From 3411089908f7ac667ff5587c08ac804da50e4a33 Mon Sep 17 00:00:00 2001
From: Peisong Xiao <peisong.xiao.xps@gmail.com>
Date: Sun, 7 Sep 2025 11:47:15 -0400
Subject: [PATCH] added networking components, hosts are next

---
 .dir-locals.el                          |  31 ++
 .gitignore                              |   2 +
 .gitmodules                             |   3 +
 CMakeLists.txt                          |  38 ++
 src/core/CMakeLists.txt                 |  56 ++
 src/core/error.cc                       |  34 ++
 src/core/error.h                        |  54 ++
 src/core/host.cc                        |  29 ++
 src/core/host.h                         |  45 ++
 src/core/logger.cc                      |  67 +++
 src/core/logger.h                       |  53 ++
 src/core/node.cc                        |  67 +++
 src/core/node.h                         |  39 ++
 src/core/rng.h                          | 131 +++++
 src/core/simulator.cc                   | 157 ++++++
 src/core/simulator.h                    | 120 +++++
 src/core/time.h                         | 129 +++++
 src/core/timer.h                        |  56 ++
 src/core/types.h                        |  70 +++
 src/hosts/CMakeLists.txt                |  49 ++
 src/hosts/nodes_dummy.cc                |   6 +
 src/hosts/nodes_dummy.h                 |   4 +
 src/network/CMakeLists.txt              |  58 +++
 src/network/link.cc                     | 182 +++++++
 src/network/link.h                      | 120 +++++
 src/network/network_nic.cc              | 660 ++++++++++++++++++++++++
 src/network/network_nic.h               | 154 ++++++
 src/network/network_node.cc             |   9 +
 src/network/network_node.h              |  24 +
 src/network/network_switch.cc           | 238 +++++++++
 src/network/network_switch.h            |  81 +++
 src/network/nic/CMakeLists.txt          |  50 ++
 src/network/nic/congestion_control.cc   | 100 ++++
 src/network/nic/congestion_control.h    |  47 ++
 src/network/nic/load_balance.cc         |  18 +
 src/network/nic/load_balance.h          |  36 ++
 src/network/nic/nic_telemetry.h         |  28 +
 src/network/packet.cc                   | 180 +++++++
 src/network/packet.h                    |  97 ++++
 src/network/switch/CMakeLists.txt       |  63 +++
 src/network/switch/dedicated_buffer.cc  |  65 +++
 src/network/switch/dedicated_buffer.h   |  38 ++
 src/network/switch/ecn_dedicated_red.cc |  50 ++
 src/network/switch/ecn_dedicated_red.h  |  54 ++
 src/network/switch/ecn_engine.h         |  51 ++
 src/network/switch/ecn_shared_red.cc    | 126 +++++
 src/network/switch/ecn_shared_red.h     |  71 +++
 src/network/switch/multicast_table.cc   | 208 ++++++++
 src/network/switch/multicast_table.h    |  52 ++
 src/network/switch/routing_alg.cc       |  36 ++
 src/network/switch/routing_alg.h        |  26 +
 src/network/switch/routing_tables.h     |  16 +
 src/network/switch/shared_buffer.cc     |  63 +++
 src/network/switch/shared_buffer.h      |  39 ++
 src/network/switch/switch_buffer.cc     | 244 +++++++++
 src/network/switch/switch_buffer.h      | 162 ++++++
 src/network/switch/unicast_table.cc     |  54 ++
 src/network/switch/unicast_table.h      |  36 ++
 third_party/yaml-cpp                    |   1 +
 59 files changed, 4777 insertions(+)
 create mode 100644 .dir-locals.el
 create mode 100644 .gitignore
 create mode 100644 .gitmodules
 create mode 100644 CMakeLists.txt
 create mode 100644 src/core/CMakeLists.txt
 create mode 100644 src/core/error.cc
 create mode 100644 src/core/error.h
 create mode 100644 src/core/host.cc
 create mode 100644 src/core/host.h
 create mode 100644 src/core/logger.cc
 create mode 100644 src/core/logger.h
 create mode 100644 src/core/node.cc
 create mode 100644 src/core/node.h
 create mode 100644 src/core/rng.h
 create mode 100644 src/core/simulator.cc
 create mode 100644 src/core/simulator.h
 create mode 100644 src/core/time.h
 create mode 100644 src/core/timer.h
 create mode 100644 src/core/types.h
 create mode 100644 src/hosts/CMakeLists.txt
 create mode 100644 src/hosts/nodes_dummy.cc
 create mode 100644 src/hosts/nodes_dummy.h
 create mode 100644 src/network/CMakeLists.txt
 create mode 100644 src/network/link.cc
 create mode 100644 src/network/link.h
 create mode 100644 src/network/network_nic.cc
 create mode 100644 src/network/network_nic.h
 create mode 100644 src/network/network_node.cc
 create mode 100644 src/network/network_node.h
 create mode 100644 src/network/network_switch.cc
 create mode 100644 src/network/network_switch.h
 create mode 100644 src/network/nic/CMakeLists.txt
 create mode 100644 src/network/nic/congestion_control.cc
 create mode 100644 src/network/nic/congestion_control.h
 create mode 100644 src/network/nic/load_balance.cc
 create mode 100644 src/network/nic/load_balance.h
 create mode 100644 src/network/nic/nic_telemetry.h
 create mode 100644 src/network/packet.cc
 create mode 100644 src/network/packet.h
 create mode 100644 src/network/switch/CMakeLists.txt
 create mode 100644 src/network/switch/dedicated_buffer.cc
 create mode 100644 src/network/switch/dedicated_buffer.h
 create mode 100644 src/network/switch/ecn_dedicated_red.cc
 create mode 100644 src/network/switch/ecn_dedicated_red.h
 create mode 100644 src/network/switch/ecn_engine.h
 create mode 100644 src/network/switch/ecn_shared_red.cc
 create mode 100644 src/network/switch/ecn_shared_red.h
 create mode 100644 src/network/switch/multicast_table.cc
 create mode 100644 src/network/switch/multicast_table.h
 create mode 100644 src/network/switch/routing_alg.cc
 create mode 100644 src/network/switch/routing_alg.h
 create mode 100644 src/network/switch/routing_tables.h
 create mode 100644 src/network/switch/shared_buffer.cc
 create mode 100644 src/network/switch/shared_buffer.h
 create mode 100644 src/network/switch/switch_buffer.cc
 create mode 100644 src/network/switch/switch_buffer.h
 create mode 100644 src/network/switch/unicast_table.cc
 create mode 100644 src/network/switch/unicast_table.h
 create mode 160000 third_party/yaml-cpp

diff --git a/.dir-locals.el b/.dir-locals.el
new file mode 100644
index 0000000..2f048c3
--- /dev/null
+++ b/.dir-locals.el
@@ -0,0 +1,31 @@
+((c++-mode . ((cmake-ide-project-dir . "/home/peisongxiao/Projects/dofs")
+              (cmake-ide-build-dir   . "/home/peisongxiao/Projects/dofs/build/ide")
+              (cmake-ide-cmake-args  . ("-DCMAKE_EXPORT_COMPILE_COMMANDS=ON"
+                                        "-DCMAKE_BUILD_TYPE=Debug"
+                                        "-DDOFS_GLOB_SOURCES=ON"
+                                        "-DYAML_BUILD_SHARED_LIBS=OFF"
+                                        "-DYAML_CPP_BUILD_TESTS=OFF"
+                                        "-DYAML_CPP_BUILD_TOOLS=OFF"
+                                        "-DYAML_ENABLE_PIC=ON"))
+              ;; KEY BIT: let ff-find-other-file look in build/ide/src/** for stubs (time.h -> build/ide/src/.../time.cpp)
+              (eval . (let* ((proj (locate-dominating-file default-directory "CMakeLists.txt"))
+                             (build (expand-file-name "build/ide" proj)))
+                        (setq-local ff-search-directories
+                                    (list
+                                     "."            ; current dir
+                                     "../src/**"    ; sibling source
+				     "/home/peisongxiao/Projects/dofs/build/ide/src/**"
+                                     (concat build "/src/**") ; generated stubs live here
+                                     "/usr/include" "/usr/local/include"))
+                        ;; prefer “first TU that includes this header” when available
+                        (with-eval-after-load 'cmake-ide
+                          (setq cmake-ide-header-search-first-including t
+                                cmake-ide-header-search-other-file t))
+                        ;; keep Flycheck on
+                        (flycheck-mode 1)
+			(company-mode 1))))))
+
+((nil . ((cmake-ide-build-dir . "build")
+         (cmake-ide-project-dir . ".")
+         (cmake-ide-cmake-opts . "-DCMAKE_EXPORT_COMPILE_COMMANDS=ON")
+         (cmake-ide-debug . t))))
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..bc7942d
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,2 @@
+build/
+prompt.txt
diff --git a/.gitmodules b/.gitmodules
new file mode 100644
index 0000000..7378aa9
--- /dev/null
+++ b/.gitmodules
@@ -0,0 +1,3 @@
+[submodule "third_party/yaml-cpp"]
+	path = third_party/yaml-cpp
+	url = https://git.peisongxiao.com/peisongxiao/yaml-cpp.git
diff --git a/CMakeLists.txt b/CMakeLists.txt
new file mode 100644
index 0000000..dd70782
--- /dev/null
+++ b/CMakeLists.txt
@@ -0,0 +1,38 @@
+cmake_minimum_required(VERSION 3.20)
+project(dofs LANGUAGES CXX)
+
+# Global C++ settings
+set(CMAKE_CXX_STANDARD 20)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON CACHE BOOL "Export compile_commands.json")
+
+# Dev toggle: convenient but less deterministic than listing files explicitly
+option(DOFS_GLOB_SOURCES "Dev: auto-add *.cc files via GLOB (not ideal for CI)" OFF)
+
+# Interface target to export common config (include dirs, features)
+add_library(dofs_config INTERFACE)
+target_compile_features(dofs_config INTERFACE cxx_std_20)
+
+# Let everyone include headers like: #include "core/error.h"
+target_include_directories(dofs_config INTERFACE "${PROJECT_SOURCE_DIR}/src")
+
+# yaml-cpp submodule (static)
+if(EXISTS "${PROJECT_SOURCE_DIR}/third_party/yaml-cpp/CMakeLists.txt")
+    # Use yaml-cpp's options; force them at configure-time
+    set(YAML_BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
+    set(YAML_CPP_BUILD_TESTS OFF CACHE BOOL "" FORCE)
+    set(YAML_CPP_BUILD_TOOLS OFF CACHE BOOL "" FORCE)
+    set(YAML_CPP_BUILD_CONTRIB ON  CACHE BOOL "" FORCE)
+    set(YAML_ENABLE_PIC ON          CACHE BOOL "" FORCE)
+    set(YAML_CPP_INSTALL OFF        CACHE BOOL "" FORCE)
+
+    add_subdirectory(third_party/yaml-cpp EXCLUDE_FROM_ALL)
+else()
+    message(WARNING "yaml-cpp submodule missing at third_party/yaml-cpp")
+endif()
+
+# add_subdirectory(src)
+add_subdirectory(src/core)
+add_subdirectory(src/network)
+add_subdirectory(src/hosts)
diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
new file mode 100644
index 0000000..94e4dca
--- /dev/null
+++ b/src/core/CMakeLists.txt
@@ -0,0 +1,56 @@
+add_library(dofs_core STATIC)
+
+# Propagate include dir via dofs_config to anyone linking dofs_core
+target_link_libraries(dofs_core PUBLIC dofs_config)
+
+target_include_directories(dofs_core
+  PUBLIC
+    $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/src>
+    $<INSTALL_INTERFACE:include>
+)
+
+if(TARGET yaml-cpp)
+    target_link_libraries(dofs_core PUBLIC yaml-cpp)
+endif()
+
+if(DOFS_GLOB_SOURCES)
+    file(GLOB CORE_CC CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.cc")
+    target_sources(dofs_core PRIVATE ${CORE_CC})
+else()
+    # Prefer explicit lists for reproducibility
+    target_sources(dofs_core
+    PUBLIC
+    	time.h
+	timer.h
+	types.h
+	error.h
+	rng.h
+	logger.h
+	simulator.h
+	node.h
+	host.h
+    PRIVATE
+        error.cc
+        logger.cc
+        simulator.cc
+	node.cc
+	host.cc
+    )
+endif()
+
+add_library(dofs::core ALIAS dofs_core)
+
+# --- Tooling-only unity TU for headers (core) ---
+# Generate per-header stubs with the SAME basename (time.h -> time.cpp)
+file(GLOB CORE_HEADERS CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.h")
+set(CORE_STUBS)
+foreach(h ${CORE_HEADERS})
+    get_filename_component(_base "${h}" NAME_WE)
+    set(_stub "${CMAKE_CURRENT_BINARY_DIR}/${_base}.cpp")
+    file(WRITE "${_stub}" "// tooling stub for ${_base}.h\n#include \"core/${_base}.h\"\n")
+    list(APPEND CORE_STUBS "${_stub}")
+endforeach()
+
+add_library(dofs_core_headers_tooling STATIC ${CORE_STUBS})
+target_link_libraries(dofs_core_headers_tooling PRIVATE dofs_core)
+# This target exists purely for compile_commands; no need to link it elsewhere.
diff --git a/src/core/error.cc b/src/core/error.cc
new file mode 100644
index 0000000..432d8fd
--- /dev/null
+++ b/src/core/error.cc
@@ -0,0 +1,34 @@
+#include "error.h"
+
+#include <iostream>
+#include <mutex>
+#include <string>
+
+namespace dofs {
+
+static std::mutex g_error_mutex;
+
+std::mutex &error_mutex() noexcept {
+        return g_error_mutex;
+}
+
+static inline void emit_error_line(std::string_view type,
+                                   std::optional<uint64_t> ts,
+                                   std::string_view message) noexcept {
+        std::lock_guard<std::mutex> lock(g_error_mutex);
+        std::cerr << '[' << type << ']';
+
+        if (ts.has_value()) {
+                std::cerr << '[' << *ts << ']';
+        }
+
+        std::cerr << ": " << message << '\n';
+}
+
+void log_error(std::string_view type,
+               std::string_view message,
+               std::optional<uint64_t> timestamp) noexcept {
+        emit_error_line(type, timestamp, message);
+}
+
+} // namespace dofs
diff --git a/src/core/error.h b/src/core/error.h
new file mode 100644
index 0000000..5ab2eef
--- /dev/null
+++ b/src/core/error.h
@@ -0,0 +1,54 @@
+#ifndef CORE_ERROR_H
+#define CORE_ERROR_H
+
+#include <cstdint>
+#include <optional>
+#include <string_view>
+#include <mutex>
+#include <utility>
+#include <iostream>
+#include <format>
+
+namespace dofs {
+
+void log_error(std::string_view type,
+               std::string_view message,
+               std::optional<uint64_t> timestamp = std::nullopt) noexcept;
+
+std::mutex &error_mutex() noexcept;
+
+template <class T>
+inline T&& show(std::string_view name, T&& value) noexcept {
+        std::lock_guard<std::mutex> lock(error_mutex());
+        std::cerr << name << '=' << value << '\n';
+        return std::forward<T>(value);
+}
+
+template <class T>
+inline T&& eval_and_show(std::string_view expr, T&& value) noexcept {
+        std::lock_guard<std::mutex> lock(error_mutex());
+        std::cerr << expr << '=' << value << '\n';
+        return std::forward<T>(value);
+}
+
+#define DOFS_ERROR(TYPE, MSG) \
+        ::dofs::log_error((TYPE), (MSG))
+
+#define DOFS_ERROR_T(TYPE, MSG, TS) \
+        ::dofs::log_error((TYPE), (MSG), (TS))
+
+#define DOFS_ERROR_ST(TYPE, SRC, MSG, TS) \
+        ::dofs::log_error((TYPE), (SRC), (MSG), (TS))
+
+#define DOFS_SHOW(VAR) \
+        (::dofs::show(#VAR, (VAR)))
+
+#define DOFS_EVAL(EXPR) \
+        (::dofs::eval_and_show(#EXPR, (EXPR)))
+
+#define FORMAT(VAR) \
+        (std::format("{}={}", #VAR, VAR))
+
+} // namespace dofs
+
+#endif // CORE_ERROR_H
diff --git a/src/core/host.cc b/src/core/host.cc
new file mode 100644
index 0000000..28c429d
--- /dev/null
+++ b/src/core/host.cc
@@ -0,0 +1,29 @@
+#include "core/host.h"
+#include "core/error.h"
+
+namespace dofs {
+
+Host::Host(Simulator *const sim, NodeId id) noexcept
+        : Node(sim, id, NodeType::HOST), _nic(nullptr) {
+        // Hosts start in OK unless caller overrides later.
+        Node::set_status(NodeStatus::OK);
+}
+
+void Host::attach_nic(NetworkNic* nic) noexcept {
+        if (_nic && _nic != nic) {
+                log_error("ERROR", "Host::attach_nic called while NIC already attached");
+        }
+
+        _nic = nic;
+}
+
+void Host::detach_nic(NetworkNic* nic) noexcept {
+        if (_nic && _nic != nic) {
+                log_error("ERROR", "Host::detach_nic called with non-matching NIC pointer");
+                return;
+        }
+
+        _nic = nullptr;
+}
+
+} // namespace dofs
diff --git a/src/core/host.h b/src/core/host.h
new file mode 100644
index 0000000..107ddfd
--- /dev/null
+++ b/src/core/host.h
@@ -0,0 +1,45 @@
+#ifndef CORE_HOST_H
+#define CORE_HOST_H
+
+#include <cstdint>
+
+#include "core/node.h"
+#include "core/types.h"
+#include "network/packet.h"
+
+namespace dofs {
+
+class NetworkNic;
+
+class Host : public Node {
+public:
+        Host(Simulator *const sim, NodeId id) noexcept;
+        virtual ~Host() = default;
+
+        // Access to the attached NIC (may be null if detached)
+        NetworkNic *nic() const noexcept {
+                return _nic;
+        }
+
+        // NIC lifecycle hooks (called by the NIC)
+        void attach_nic(NetworkNic* nic) noexcept;
+        void detach_nic(NetworkNic* nic) noexcept;
+
+        // Data-plane completion: a whole flow has arrived.
+        virtual void recv_flow(NodeId src,
+                               FlowPriority priority,
+                               Bytes flow_size) = 0;
+
+        // Control/telemetry interrupt: ACK/NACK/TRIM_BACK, etc.
+        virtual void recv_frame(const Packet& frame) = 0;
+
+        Host(const Host &) = delete;
+        Host &operator=(const Host &) = delete;
+
+private:
+        NetworkNic *_nic{nullptr};
+};
+
+} // namespace dofs
+
+#endif // CORE_HOST_H
diff --git a/src/core/logger.cc b/src/core/logger.cc
new file mode 100644
index 0000000..76471b7
--- /dev/null
+++ b/src/core/logger.cc
@@ -0,0 +1,67 @@
+#include "logger.h"
+
+#include <algorithm>
+#include <iostream>
+
+namespace dofs {
+
+Logger::Logger(std::string_view path, bool append) noexcept
+        : _path(path) {
+        const std::ios::openmode mode = append ? std::ios::app : std::ios::trunc;
+        (void)open(mode);
+}
+
+bool Logger::open(std::ios::openmode mode) noexcept {
+        _stream.exceptions(std::ios::goodbit);
+        _stream.open(_path, std::ios::out | mode);
+        _is_open = _stream.is_open();
+
+        if (!_is_open) {
+                DOFS_ERROR("Logger", ("failed to open file: " + _path).c_str());
+        }
+
+        return _is_open;
+}
+
+Logger::~Logger() noexcept {
+        close();
+}
+
+void Logger::flush() noexcept {
+        std::lock_guard<std::mutex> _lk(_mutex);
+
+        if (_is_open) {
+                _stream.flush();
+
+                if (!_stream) {
+                        DOFS_ERROR("Logger", ("flush failed for: " + _path).c_str());
+                }
+        }
+}
+
+void Logger::close() noexcept {
+        std::lock_guard<std::mutex> _lk(_mutex);
+
+        if (_is_open) {
+                _stream.flush();
+                _stream.close();
+                _is_open = false;
+        }
+}
+
+void Logger::write_line(std::string_view line) noexcept {
+        std::lock_guard<std::mutex> _lk(_mutex);
+
+        if (!_is_open) {
+                DOFS_ERROR("Logger", ("write on closed file: " + _path).c_str());
+                return;
+        }
+
+        _stream << line << '\n';
+
+        if (!_stream) {
+                DOFS_ERROR("Logger", ("write failed for: " + _path).c_str());
+        }
+}
+
+} // namespace dofs
diff --git a/src/core/logger.h b/src/core/logger.h
new file mode 100644
index 0000000..fefdbfc
--- /dev/null
+++ b/src/core/logger.h
@@ -0,0 +1,53 @@
+#ifndef CORE_LOGGER_H
+#define CORE_LOGGER_H
+
+#include <cstdint>
+#include <fstream>
+#include <memory>
+#include <mutex>
+#include <optional>
+#include <string>
+#include <string_view>
+#include <unordered_map>
+#include <utility>
+
+#include "core/error.h"
+#include "core/types.h"
+
+namespace dofs {
+
+class Logger final {
+public:
+        Logger(std::string_view path, bool append) noexcept;
+        ~Logger() noexcept;
+
+        Logger(const Logger &) = delete;
+        Logger &operator=(const Logger &) = delete;
+        Logger(Logger &&) = delete;
+        Logger &operator=(Logger &&) = delete;
+
+        bool is_open() const noexcept {
+                return _is_open;
+        }
+
+        void write_line(std::string_view line) noexcept;
+
+        void flush() noexcept;
+        void close() noexcept;
+
+        std::string_view path() const noexcept {
+                return _path;
+        }
+
+private:
+        bool open(std::ios::openmode mode) noexcept;
+
+        std::string _path;
+        std::ofstream _stream;
+        bool _is_open{false};
+        std::mutex _mutex;
+};
+
+} // namespace dofs
+
+#endif // CORE_LOGGER_H
diff --git a/src/core/node.cc b/src/core/node.cc
new file mode 100644
index 0000000..0758335
--- /dev/null
+++ b/src/core/node.cc
@@ -0,0 +1,67 @@
+#include "core/node.h"
+
+#include "core/simulator.h"
+#include "core/error.h"
+
+#include <optional>
+
+namespace dofs {
+
+Node::Node(Simulator *const sim, NodeId id, NodeType type) noexcept
+        : _sim(sim), _id(id), _status(NodeStatus::OK), _type(type) {}
+
+// Accessors
+NodeId Node::id() const noexcept {
+        return _id;
+}
+NodeStatus Node::status() const noexcept {
+        return _status;
+}
+NodeType Node::type() const noexcept {
+        return _type;
+}
+
+void Node::set_status(NodeStatus s) noexcept {
+        _status = s;
+}
+
+static inline bool schedule_status_ok_after(Simulator *const sim, Time delay_ns,
+                Node* self) {
+        if (!sim) {
+                log_error("ERROR", "Node: simulator instance not found for boot/reboot");
+                return false;
+        }
+
+        sim->schedule_after(delay_ns, [self]() {
+                self->set_status(NodeStatus::OK);
+        });
+        return true;
+}
+
+void Node::boot(Time boottime_ns) {
+        if (_status != NodeStatus::DOWN) {
+                log_error("ERROR", "boot() ignored: node not in DOWN state");
+                return;
+        }
+
+        _status = NodeStatus::BOOTING;
+
+        if (!schedule_status_ok_after(_sim, boottime_ns, this)) {
+                // Intentionally left BOOTING; caller can handle recovery/logs.
+        }
+}
+
+void Node::reboot(Time boottime_ns) {
+        if (_status != NodeStatus::OK && _status != NodeStatus::THROTTLING) {
+                log_error("ERROR", "reboot() ignored: node not in OK/THROTTLING state");
+                return;
+        }
+
+        _status = NodeStatus::BOOTING;
+
+        if (!schedule_status_ok_after(_sim, boottime_ns, this)) {
+                // Intentionally left BOOTING; caller can handle recovery/logs.
+        }
+}
+
+} // namespace dofs
diff --git a/src/core/node.h b/src/core/node.h
new file mode 100644
index 0000000..a89875b
--- /dev/null
+++ b/src/core/node.h
@@ -0,0 +1,39 @@
+#ifndef CORE_NODE_H
+#define CORE_NODE_H
+
+#include <cstdint>
+
+#include "core/simulator.h"
+#include "core/time.h"
+#include "core/types.h"
+
+namespace dofs {
+
+class Node {
+public:
+        Node(Simulator *const sim, NodeId id, NodeType type) noexcept;
+        virtual ~Node() = default;
+
+        NodeId id() const noexcept;
+        NodeStatus status() const noexcept;
+        NodeType type() const noexcept;
+
+        void set_status(NodeStatus s) noexcept;
+
+        void boot(Time boottime_ns);
+        void reboot(Time boottime_ns);
+
+        Node(const Node &) = delete;
+        Node &operator=(const Node &) = delete;
+
+protected:
+        Simulator *const _sim;
+
+        NodeId _id;
+        NodeStatus _status;
+        NodeType _type;
+};
+
+} // namespace dofs
+
+#endif // CORE_NODE_H
diff --git a/src/core/rng.h b/src/core/rng.h
new file mode 100644
index 0000000..59c6c36
--- /dev/null
+++ b/src/core/rng.h
@@ -0,0 +1,131 @@
+#ifndef CORE_RNG_H
+#define CORE_RNG_H
+
+#include <cstdint>
+#include <random>
+#include <vector>
+#include <utility>
+#include <initializer_list>
+#include <type_traits>
+#include <limits>
+
+namespace dofs {
+
+class Rng final {
+public:
+        using engine_type = std::mt19937_64;
+        using seed_type   = engine_type::result_type;
+
+        explicit Rng(seed_type seed = default_seed()) noexcept
+                : _eng(seed) {}
+
+        void seed(seed_type s) noexcept {
+                _eng.seed(s);
+        }
+
+        // Real in [0, 1)
+        double uniform01() {
+                return std::uniform_real_distribution<double> {}(_eng);
+        }
+
+        // Integer in [lo, hi) — lo inclusive, hi exclusive
+        template <typename Int,
+                  typename = std::enable_if_t<std::is_integral<Int>::value>>
+        Int uniform_range(Int lo_inclusive, Int hi_exclusive) {
+                return std::uniform_int_distribution<Int>(lo_inclusive, hi_exclusive - 1)(_eng);
+        }
+
+        // Integer in [0, hi) — convenience overload
+        template <typename Int,
+                  typename = std::enable_if_t<std::is_integral<Int>::value>>
+        Int uniform_range(Int hi_exclusive) {
+                return uniform_range<Int>(0, hi_exclusive);
+        }
+
+        double uniform_range(double lo_inclusive, double hi_exclusive) {
+
+                return lo_inclusive + (hi_exclusive - lo_inclusive) * uniform01();
+        }
+
+        std::uint64_t poisson(double lambda) {
+                return static_cast<std::uint64_t>(std::poisson_distribution<std::uint64_t>
+                                                  (lambda)(_eng));
+        }
+
+        template <typename T>
+        T choose_weighted(const std::vector<std::pair<double, T>>& items) {
+                return choose_weighted_impl(items.begin(), items.end());
+        }
+
+        template <typename T>
+        T choose_weighted(std::initializer_list<std::pair<double, T>> items) {
+                return choose_weighted_impl(items.begin(), items.end());
+        }
+
+private:
+        engine_type _eng;
+
+        static constexpr seed_type default_seed() noexcept {
+                return 0xC0FFEEULL ^ 0x9E3779B97F4A7C15ULL;
+        }
+
+        template <typename Iter>
+        auto choose_weighted_impl(Iter first,
+                                  Iter last) -> typename std::iterator_traits<Iter>::value_type::second_type {
+                using Pair = typename std::iterator_traits<Iter>::value_type;
+                using T    = typename Pair::second_type;
+
+                double total = 0.0;
+
+                for (auto it = first; it != last; ++it) {
+                        const double w = it->first;
+
+                        if (w > 0.0 && std::isfinite(w))
+                                total += w;
+                }
+
+                if (!(total > 0.0)) {
+                        std::size_t n = 0;
+
+                        for (auto it = first; it != last; ++it)
+                                ++n;
+
+                        if (n == 0)
+                                return T{};
+
+                        std::size_t idx = static_cast<std::size_t>(uniform_range<std::uint64_t>(0,
+                                          static_cast<std::uint64_t>(n)));
+
+                        auto it = first;
+
+                        while (idx--)
+                                ++it;
+
+                        return it->second;
+                }
+
+                double u = uniform01() * total;
+
+                for (auto it = first; it != last; ++it) {
+                        const double w = (it->first > 0.0 &&
+                                          std::isfinite(it->first)) ? it->first : 0.0;
+
+                        if (u < w)
+                                return it->second;
+
+                        u -= w;
+                }
+
+                auto it = last;
+
+                do {
+                        --it;
+                } while (it->first <= 0.0 || !std::isfinite(it->first));
+
+                return it->second;
+        }
+};
+
+} // namespace dofs
+
+#endif // CORE_RNG_H
diff --git a/src/core/simulator.cc b/src/core/simulator.cc
new file mode 100644
index 0000000..61aef69
--- /dev/null
+++ b/src/core/simulator.cc
@@ -0,0 +1,157 @@
+#include "core/simulator.h"
+
+#include <memory>
+
+#include "core/rng.h"
+#include "network/link.h"
+
+namespace dofs {
+
+bool Simulator::Cmp::operator()(const Item& a, const Item& b) const noexcept {
+        if (a.when != b.when)
+                return a.when > b.when;
+
+        return a.id > b.id;
+}
+
+Time Simulator::now() const noexcept {
+        return _now;
+}
+
+bool Simulator::cancel(EventId id) {
+        return _cancelled.insert(id).second;
+}
+
+bool Simulator::run_next() {
+        while (!_event_pq.empty()) {
+                Item it = _event_pq.top();
+                _event_pq.pop();
+
+                if (_cancelled.erase(it.id) > 0)
+                        continue;
+
+                _now = it.when;
+                it.fn();
+                return true;
+        }
+
+        return false;
+}
+
+void Simulator::run_until(Time end_time) {
+        while (!_event_pq.empty()) {
+                const Item& top = _event_pq.top();
+
+                if (end_time < top.when)
+                        break;
+
+                run_next();
+        }
+}
+
+std::vector<std::unique_ptr<Simulator>>& Simulator::registry_() {
+        static std::vector<std::unique_ptr<Simulator>> reg;
+        return reg;
+}
+
+std::pair<InstanceId, Simulator *> Simulator::create_simulator(InstanceId id) {
+        if (id == INVALID_INSTANCE_ID) {
+                return {INVALID_INSTANCE_ID, nullptr};
+        }
+
+        auto& reg = registry_();
+
+        if (static_cast<size_t>(id) >= reg.size()) {
+                reg.resize(static_cast<size_t>(id) + 1);
+        }
+
+        if (!reg[static_cast<size_t>(id)]) {
+                auto sim = std::make_unique<Simulator>();
+                Simulator* ptr = sim.get();
+                reg[static_cast<size_t>(id)] = std::move(sim);
+                return {id, ptr};
+        }
+
+        return {id, reg[static_cast<size_t>(id)].get()};
+}
+
+Simulator * Simulator::get_simulator(InstanceId id) noexcept {
+        if (id == INVALID_INSTANCE_ID)
+                return nullptr;
+
+        auto& reg = registry_();
+        const size_t idx = static_cast<size_t>(id);
+
+        if (idx >= reg.size())
+                return nullptr;
+
+        return reg[idx].get();
+}
+
+Rng * Simulator::create_rng(std::uint64_t seed) {
+        if (_rng)
+                return nullptr;
+
+        _rng = std::make_unique<Rng>(seed);
+        return _rng.get();
+}
+Rng * Simulator::get_rng() noexcept {
+        return _rng.get();
+}
+
+Rng const * Simulator::get_rng() const noexcept {
+        return _rng.get();
+}
+
+std::pair<LinkId, Link *> Simulator::create_link(
+        NetworkNode* a, PortId a_port,
+        NetworkNode* b, PortId b_port,
+        Time latency,
+        double bandwidth_gbps) {
+
+        if (!a || !b)
+                return {INVALID_LINK_ID, nullptr};
+
+        if (bandwidth_gbps <= 0.0)
+                return {INVALID_LINK_ID, nullptr};
+
+        if (latency < Time{0})
+                return {INVALID_LINK_ID, nullptr};
+
+        const LinkId id = static_cast<LinkId>(_links.size());
+        auto up =
+                std::unique_ptr<Link>(new Link(this, id,
+                                               a, a_port, b, b_port,
+                                               latency,
+                                               bandwidth_gbps));
+
+        Link* raw = up.get();
+        _links.push_back(std::move(up));
+        return {id, raw};
+}
+
+Link * Simulator::get_link(LinkId id) noexcept {
+        if (id == INVALID_LINK_ID)
+                return nullptr;
+
+        const size_t idx = static_cast<size_t>(id);
+
+        if (idx >= _links.size())
+                return nullptr;
+
+        return _links[idx].get();
+}
+
+Link const * Simulator::get_link(LinkId id) const noexcept {
+        if (id == INVALID_LINK_ID)
+                return nullptr;
+
+        const size_t idx = static_cast<size_t>(id);
+
+        if (idx >= _links.size())
+                return nullptr;
+
+        return _links[idx].get();
+}
+
+} // namespace dofs
diff --git a/src/core/simulator.h b/src/core/simulator.h
new file mode 100644
index 0000000..5fc45a8
--- /dev/null
+++ b/src/core/simulator.h
@@ -0,0 +1,120 @@
+#ifndef CORE_SIMULATOR_H
+#define CORE_SIMULATOR_H
+
+// TODO: implement concrete node methods for different nodes, store them all uniformly in vector<u_ptr<Node>> and return a get function
+
+#include <cstdint>
+#include <functional>
+#include <queue>
+#include <vector>
+#include <unordered_set>
+#include <utility>
+#include <tuple>
+#include <type_traits>
+#include <limits>
+#include <memory>
+
+#include "core/time.h"
+#include "core/types.h"
+
+namespace dofs {
+
+class Rng;
+class Logger;
+
+class Node;
+class NetworkNode;
+class Link;
+
+constexpr EventId NULL_EVENT = 0;
+
+constexpr InstanceId INVALID_INSTANCE_ID =
+        std::numeric_limits<InstanceId>::max();
+constexpr LinkId     INVALID_LINK_ID     =
+        std::numeric_limits<LinkId>::max();
+
+class Simulator final {
+private:
+        struct Item {
+                Time when;
+                EventId id;
+                std::function<void()> fn;
+        };
+
+        struct Cmp {
+                bool operator()(const Item& a, const Item& b) const noexcept;
+        };
+
+        std::priority_queue<Item, std::vector<Item>, Cmp> _event_pq;
+        std::unordered_set<EventId> _cancelled;
+        Time    _now{};
+        EventId _next_id{1};
+
+        std::unique_ptr<Rng> _rng;
+
+        std::vector<std::unique_ptr<Node>> _nodes;
+        std::vector<std::unique_ptr<Link>> _links;
+
+public:
+        Simulator() = default;
+
+        static std::pair<InstanceId, Simulator *> create_simulator(InstanceId id);
+        static Simulator *get_simulator(InstanceId id) noexcept;
+
+        Time now() const noexcept;
+
+        template <class F, class... Args>
+        EventId schedule_at(Time abs_time, F&& f, Args&&... args) {
+                if (abs_time < _now)
+                        return NULL_EVENT;
+
+                const EventId eid = _next_id++;
+                Item it{
+                        abs_time,
+                        eid,
+                        make_callable(std::forward<F>(f), std::forward<Args>(args)...)
+                };
+                _event_pq.push(std::move(it));
+                return eid;
+        }
+
+        template <class F, class... Args>
+        EventId schedule_after(Time delay, F&& f, Args&&... args) {
+                return schedule_at(_now + delay, std::forward<F>(f),
+                                   std::forward<Args>(args)...);
+        }
+
+        bool cancel(EventId id);
+        bool run_next();
+        void run_until(Time end_time);
+
+        // ---------- Object management ----------
+        Rng* create_rng(std::uint64_t seed);
+        Rng*       get_rng() noexcept;
+        Rng const* get_rng() const noexcept;
+
+        std::pair<LinkId, Link *> create_link(NetworkNode* a, PortId a_port,
+                                              NetworkNode* b, PortId b_port,
+                                              Time latency,
+                                              double bandwidth_gbps);
+
+        Link*       get_link(LinkId id) noexcept;
+        Link const* get_link(LinkId id) const noexcept;
+
+private:
+        template <class F, class... Args>
+        static auto make_callable(F&& f, Args&&... args) {
+                using Fn  = std::decay_t<F>;
+                using Tup = std::tuple<std::decay_t<Args>...>;
+                return [fn = Fn(std::forward<F>(f)),
+                tup = Tup(std::forward<Args>(args)...)]() mutable {
+                        std::apply(fn, std::move(tup));
+                };
+        }
+
+        static std::vector<std::unique_ptr<Simulator>>& registry_();
+};
+
+} // namespace dofs
+
+#endif // CORE_SIMULATOR_H
diff --git a/src/core/time.h b/src/core/time.h
new file mode 100644
index 0000000..ed58636
--- /dev/null
+++ b/src/core/time.h
@@ -0,0 +1,129 @@
+#ifndef CORE_TIME_H
+#define CORE_TIME_H
+
+#include <cstdint>
+#include <optional>
+
+namespace dofs {
+
+class Time final {
+public:
+        using rep = uint64_t;
+
+        // ----- Constructors -----
+        constexpr Time() : _nsec(0) {}
+        explicit constexpr Time(rep ns) : _nsec(ns) {}
+
+        // ----- Factories -----
+        static constexpr Time from_ns(rep ns) noexcept {
+                return Time(ns);
+        }
+        static constexpr Time from_us(rep us) noexcept {
+                return Time(us * 1000ULL);
+        }
+        static constexpr Time from_ms(rep ms) noexcept {
+                return Time(ms * 1000ULL * 1000ULL);
+        }
+        static constexpr Time from_s (rep s ) noexcept {
+                return Time(s  * 1000ULL * 1000ULL * 1000ULL);
+        }
+
+        // ----- Accessors -----
+        constexpr rep ns() const noexcept {
+                return _nsec;
+        }
+        constexpr rep count() const noexcept {
+                return _nsec;
+        }
+
+        // ----- Converters (plain uint64_t) -----
+        static constexpr rep us_to_ns(rep us) noexcept {
+                return us * 1000ULL;
+        }
+        static constexpr rep ms_to_ns(rep ms) noexcept {
+                return ms * 1000ULL * 1000ULL;
+        }
+
+        // ----- Comparisons -----
+        friend constexpr bool operator==(Time a, Time b) noexcept {
+                return a._nsec == b._nsec;
+        }
+
+        friend constexpr bool operator!=(Time a, Time b) noexcept {
+                return !(a == b);
+        }
+
+        friend constexpr bool operator< (Time a, Time b) noexcept {
+                return a._nsec <  b._nsec;
+        }
+
+        friend constexpr bool operator> (Time a, Time b) noexcept {
+                return b < a;
+        }
+
+        friend constexpr bool operator<=(Time a, Time b) noexcept {
+                return !(b < a);
+        }
+
+        friend constexpr bool operator>=(Time a, Time b) noexcept {
+                return !(a < b);
+        }
+
+        // ----- Arithmetic -----
+        constexpr Time &operator+=(Time rhs) noexcept {
+                _nsec += rhs._nsec;
+                return *this;
+        }
+        constexpr Time &operator-=(Time rhs) noexcept {
+                if (_nsec < rhs._nsec)
+                        _nsec = 0;
+                else
+                        _nsec -= rhs._nsec;
+
+                return *this;
+        }
+
+        friend constexpr Time operator+(Time a, Time b) noexcept {
+                return Time(a._nsec + b._nsec);
+        }
+
+        friend constexpr Time operator*(Time a, Time b) noexcept {
+                return Time(a._nsec * b._nsec);
+        }
+
+        friend constexpr std::optional<Time> operator-(Time a, Time b) noexcept {
+                return safe_sub(a, b);
+        }
+
+        friend constexpr std::optional<Time> safe_sub(Time a, Time b) noexcept {
+                if (a._nsec < b._nsec) {
+                        return std::nullopt;
+                }
+
+                return Time(a._nsec - b._nsec);
+        }
+
+        constexpr Time unsafe_sub(Time t) const noexcept {
+                return Time(this->_nsec - t._nsec);
+        }
+
+private:
+        rep _nsec;
+};
+
+constexpr Time operator""_ns(unsigned long long v) noexcept {
+        return Time::from_ns(static_cast<Time::rep>(v));
+}
+constexpr Time operator""_us(unsigned long long v) noexcept {
+        return Time::from_us(static_cast<Time::rep>(v));
+}
+constexpr Time operator""_ms(unsigned long long v) noexcept {
+        return Time::from_ms(static_cast<Time::rep>(v));
+}
+constexpr Time operator""_s (unsigned long long v) noexcept {
+        return Time::from_s (static_cast<Time::rep>(v));
+}
+
+} // namespace dofs
+
+#endif // CORE_TIME_H
diff --git a/src/core/timer.h b/src/core/timer.h
new file mode 100644
index 0000000..1acc51a
--- /dev/null
+++ b/src/core/timer.h
@@ -0,0 +1,56 @@
+#ifndef CORE_TIMER_H
+#define CORE_TIMER_H
+
+#include <chrono>
+#include "core/time.h"
+
+namespace dofs {
+
+class Timer final {
+public:
+        Timer() {
+                init();
+        }
+
+        void init() noexcept {
+                _start = clock::now();
+        }
+
+        Time start() const noexcept {
+                using namespace std;
+
+                auto tp = _start.time_since_epoch();
+                return Time::from_ns(
+                               static_cast<Time::rep>(
+                                       chrono::duration_cast<chrono::nanoseconds>(tp).count()
+                               ));
+        }
+
+        Time now() const noexcept {
+                using namespace std;
+
+                auto tp = clock::now().time_since_epoch();
+                return Time::from_ns(
+                               static_cast<Time::rep>(
+                                       chrono::duration_cast<chrono::nanoseconds>(tp).count()
+                               ));
+        }
+
+        Time elapsed() const noexcept {
+                using namespace std;
+
+                auto dur = clock::now() - _start;
+                return Time::from_ns(
+                               static_cast<Time::rep>(
+                                       chrono::duration_cast<chrono::nanoseconds>(dur).count()
+                               ));
+        }
+
+private:
+        using clock = std::chrono::steady_clock;
+        clock::time_point _start;
+};
+
+} // namespace dofs
+
+#endif // CORE_TIMER_H
diff --git a/src/core/types.h b/src/core/types.h
new file mode 100644
index 0000000..330278f
--- /dev/null
+++ b/src/core/types.h
@@ -0,0 +1,70 @@
+#ifndef CORE_TYPES_H
+#define CORE_TYPES_H
+
+#include <cstdint>
+
+namespace dofs {
+
+using uint128_t = __uint128_t;
+using PacketGroups = uint128_t;
+
+using EventId = uint64_t;
+using InstanceId = uint8_t;
+
+using NodeId = uint16_t;
+enum class NodeStatus : uint8_t {
+        OK,
+        THROTTLING,
+        DOWN,
+        BOOTING
+};
+
+enum class NodeType : uint8_t {
+        HOST,
+        SWITCH,
+        NIC
+};
+
+using PortId = uint16_t;
+
+using IPv4Addr = uint32_t;
+inline IPv4Addr ipv4(NodeId n, PortId p) noexcept {
+        return (static_cast<uint32_t>(n) << 16) | static_cast<uint32_t>(p);
+}
+
+using LinkId = uint32_t;
+enum class LinkStatus : uint8_t { OK, DOWN, THROTTLING };
+
+using Bytes = uint32_t;
+using PacketSeq = uint16_t;
+using FlowId = uint16_t;
+
+enum class PacketProtocol : uint8_t {
+        DATA,
+        ACK,
+        NACK,
+        HEADER_TRIM,
+        HEADER_TRIM_BACK
+};
+
+enum class SwitchBufferType : uint8_t { SHARED = 0, DEDICATED = 1 };
+
+enum class FlowPriority : uint8_t {
+        CTRL = 0,
+        MICE = 1,
+        ELEPHANT = 2,
+        AUTO
+};
+
+enum class CCType : uint8_t {
+        DCQCN,
+        NSCC
+};
+
+enum class LBType : uint8_t {
+        RANDOM_PACKET_SPRAYING
+};
+
+} // namespace dofs
+
+#endif // CORE_TYPES_H
diff --git a/src/hosts/CMakeLists.txt b/src/hosts/CMakeLists.txt
new file mode 100644
index 0000000..6faedb7
--- /dev/null
+++ b/src/hosts/CMakeLists.txt
@@ -0,0 +1,49 @@
+# --- hosts core (no switch sources here) ---
+add_library(dofs_hosts STATIC)
+
+# Make both "hosts/…" and "core/…" resolvable from anywhere
+target_include_directories(dofs_hosts
+  PUBLIC
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>
+    $<INSTALL_INTERFACE:include>
+)
+
+target_link_libraries(dofs_hosts PUBLIC dofs_core)
+
+if(DOFS_GLOB_SOURCES)
+  file(GLOB HOSTS_CC CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.cc")
+  # Exclude switch/*.cc from this glob just in case
+  list(FILTER HOSTS_CC EXCLUDE REGEX ".*/switch/.*")
+  target_sources(dofs_hosts PRIVATE ${HOSTS_CC})
+else()
+  target_sources(dofs_hosts
+    PUBLIC
+      nodes_dummy.h
+    PRIVATE
+      nodes_dummy.cc
+  )
+endif()
+
+add_library(dofs::hosts ALIAS dofs_hosts)
+
+# --- Tooling stubs for headers in this folder only (not switch) ---
+file(GLOB HOSTS_HEADERS CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.h")
+set(HOSTS_STUBS)
+foreach(h ${HOSTS_HEADERS})
+  get_filename_component(_base "${h}" NAME_WE)
+  set(_stub "${CMAKE_CURRENT_BINARY_DIR}/${_base}.cpp")
+  file(WRITE "${_stub}" "// tooling stub for ${_base}.h\n#include \"hosts/${_base}.h\"\n")
+  list(APPEND HOSTS_STUBS "${_stub}")
+endforeach()
+
+add_library(dofs_hosts_headers_tooling STATIC ${HOSTS_STUBS})
+
+# inherit include dirs/usage reqs from dofs_hosts
+target_link_libraries(dofs_hosts_headers_tooling PRIVATE dofs_hosts dofs_core)
+
+# This ensures the same -I flags even if cmake-ide builds the stubs alone
+target_include_directories(dofs_hosts_headers_tooling
+  PRIVATE
+    $<TARGET_PROPERTY:dofs_hosts,INTERFACE_INCLUDE_DIRECTORIES>
+    $<TARGET_PROPERTY:dofs_hosts,INCLUDE_DIRECTORIES>
+)
diff --git a/src/hosts/nodes_dummy.cc b/src/hosts/nodes_dummy.cc
new file mode 100644
index 0000000..e39c771
--- /dev/null
+++ b/src/hosts/nodes_dummy.cc
@@ -0,0 +1,6 @@
+#include "core/error.h"
+#include "core/logger.h"
+
+namespace dofs {
+int _nodes_tooling_dummy = 0;
+}
diff --git a/src/hosts/nodes_dummy.h b/src/hosts/nodes_dummy.h
new file mode 100644
index 0000000..0b1a0ef
--- /dev/null
+++ b/src/hosts/nodes_dummy.h
@@ -0,0 +1,4 @@
+#ifndef NODES_DUMMY
+#define NODES_DUMMY
+
+#endif
diff --git a/src/network/CMakeLists.txt b/src/network/CMakeLists.txt
new file mode 100644
index 0000000..c71cdbb
--- /dev/null
+++ b/src/network/CMakeLists.txt
@@ -0,0 +1,58 @@
+# --- Library: dofs_network ---
+add_library(dofs_network STATIC)
+
+# Expose src/ so "network/..." and "core/..." resolve from anywhere
+target_include_directories(dofs_network
+  PUBLIC
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>
+    $<INSTALL_INTERFACE:include>
+)
+
+target_link_libraries(dofs_network PUBLIC dofs_core)
+
+# Sources
+if(DOFS_GLOB_SOURCES)
+  file(GLOB NETWORK_CC CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.cc")
+  target_sources(dofs_network PRIVATE ${NETWORK_CC})
+else()
+  target_sources(dofs_network
+    PUBLIC
+      packet.h
+      link.h
+      network_node.h
+      network_switch.h
+      network_nic.h
+    PRIVATE
+      packet.cc
+      link.cc
+      network_node.cc
+      network_switch.cc
+      network_nic.cc
+  )
+endif()
+
+# Alias
+add_library(dofs::network ALIAS dofs_network)
+
+# --- Tooling: per-header stubs for "network/*.h" ---
+file(GLOB NETWORK_HEADERS CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.h")
+
+set(NETWORK_STUBS)
+foreach(h ${NETWORK_HEADERS})
+  get_filename_component(_base "${h}" NAME_WE)
+  set(_stub "${CMAKE_CURRENT_BINARY_DIR}/${_base}.cpp")
+  file(WRITE "${_stub}" "// tooling stub for ${_base}.h\n#include \"network/${_base}.h\"\n")
+  list(APPEND NETWORK_STUBS "${_stub}")
+endforeach()
+
+add_library(dofs_network_headers_tooling STATIC ${NETWORK_STUBS})
+# Make stubs see the same include roots as the real target
+target_include_directories(dofs_network_headers_tooling
+  PRIVATE
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>
+)
+target_link_libraries(dofs_network_headers_tooling PRIVATE dofs_network dofs_core)
+
+# --- Subdir: switch ---
+add_subdirectory(switch)
+add_subdirectory(nic)
diff --git a/src/network/link.cc b/src/network/link.cc
new file mode 100644
index 0000000..62f5ae8
--- /dev/null
+++ b/src/network/link.cc
@@ -0,0 +1,182 @@
+#include "network/link.h"
+
+#include <unordered_map>
+#include <memory>
+#include <cmath>
+
+namespace dofs {
+
+// If gbps <= 0, return a very large sentinel duration (treated as "can't send").
+Time Link::serialization_time(Bytes bytes, double gbps) noexcept {
+        if (gbps <= 0.0) {
+                return Time::from_ns(UINT64_C(0xFFFFFFFFFFFFFFFF));
+        }
+
+        const double ns = std::ceil(static_cast<double>(bytes) * 8.0 / gbps);
+        return Time::from_ns(static_cast<uint64_t>(ns));
+}
+
+Link::Link(Simulator *const sim,
+           LinkId id,
+           NetworkNode *a, PortId a_port,
+           NetworkNode *b, PortId b_port,
+           Time latency,
+           double bandwidth_gbps) noexcept
+        : _sim(sim),
+          _id(id),
+          _a(a), _b(b),
+          _a_id(a->id()),
+          _b_id(b->id()),
+          _a_port(a_port), _b_port(b_port),
+          _latency_init(latency),
+          _bandwidth_gbps_init(bandwidth_gbps),
+          _latency_cur(latency),
+          _bandwidth_gbps_cur(bandwidth_gbps),
+          _status(LinkStatus::OK),
+          _next_available_ab(Time(0)),
+          _next_available_ba(Time(0)) {}
+
+// Map sender to a direction; return nullopt if not an endpoint.
+std::optional<Link::Dir> Link::_dir_from(NodeId sender) const noexcept {
+        if (sender == _a_id)
+                return Dir::AtoB;
+
+        if (sender == _b_id)
+                return Dir::BtoA;
+
+        return std::nullopt;
+}
+
+Time Link::next_available(NodeId sender) const noexcept {
+        auto d = _dir_from(sender);
+
+        if (!d.has_value()) {
+                // Unknown sender; be defensive: report "now" so caller won't stall forever.
+                // The caller (switch) should treat this as a logic error.
+                return _sim ? _sim->now() : Time(0);
+        }
+
+        return (d.value() == Dir::AtoB) ? _next_available_ab : _next_available_ba;
+}
+
+std::optional<Link::Reservation> Link::reserve(Bytes bytes,
+                NodeId sender) noexcept {
+        if (_status == LinkStatus::DOWN) {
+                return std::nullopt;
+        }
+
+        if (!_sim) {
+                log_error("ERROR", "Link: simulator instance not found");
+                return std::nullopt;
+        }
+
+        auto d = _dir_from(sender);
+
+        if (!d.has_value()) {
+                log_error("ERROR", "Link::reserve called by non-endpoint node");
+                return std::nullopt;
+        }
+
+        Time& cursor = (d.value() == Dir::AtoB) ? _next_available_ab :
+                       _next_available_ba;
+
+        const Time now = _sim->now();
+        const Time start = (cursor > now) ? cursor : now;
+        const Time ser = serialization_time(bytes, _bandwidth_gbps_cur);
+        const Time finish = start + ser;
+
+        cursor = finish;
+
+        return Reservation{start, finish};
+}
+
+// --- Legacy path: send + schedule internally via the link ---
+// This keeps existing callers working while the switch migrates to reserve().
+void Link::send_pkt(Packet &pkt, NodeId caller) {
+        if (_status == LinkStatus::DOWN) {
+                // Drop silently
+                return;
+        }
+
+        if (!_sim) {
+                log_error("ERROR", "Link: simulator instance not found");
+                return;
+        }
+
+        auto d = _dir_from(caller);
+
+        if (!d.has_value()) {
+                log_error("ERROR", "Link::send_pkt called by non-endpoint node");
+                return;
+        }
+
+        auto rsv = reserve(pkt.total_size(), caller);
+
+        if (!rsv.has_value()) {
+                // DOWN or invalid sender; nothing to do.
+                return;
+        }
+
+        const Time now = _sim->now();
+        const Time to_finish = (rsv->finish > now) ?
+                               rsv->finish.unsafe_sub(now) : Time(0);
+        const Time total_delay = to_finish + _latency_cur;
+
+        NetworkNode* target_node = (d.value() == Dir::AtoB) ? _b : _a;
+        PortId target_ingress = (d.value() == Dir::AtoB) ? _b_port : _a_port;
+
+        _sim->schedule_after(total_delay,
+        [target_node, target_ingress, p = pkt]() mutable {
+                target_node->recv_pkt(p, target_ingress);
+        });
+}
+
+void Link::schedule_delivery_after(Packet &pkt, NodeId caller, Time after) {
+        if (_status == LinkStatus::DOWN || !_sim)
+                return;
+
+        auto d = _dir_from(caller);
+
+        if (!d.has_value()) {
+                log_error("ERROR", "Link::schedule_delivery_after called by non-endpoint node");
+                return;
+        }
+
+        NetworkNode* target_node = (d.value() == Dir::AtoB) ? _b : _a;
+        PortId target_ingress = (d.value() == Dir::AtoB) ? _b_port : _a_port;
+
+        _sim->schedule_after(after,
+        [target_node, target_ingress, p = pkt]() mutable {
+                target_node->recv_pkt(p, target_ingress);
+        });
+}
+
+void Link::set_status(LinkStatus s, Time new_latency,
+                      double new_bandwidth_gbps) {
+        switch (s) {
+        case LinkStatus::OK:
+                _status = LinkStatus::OK;
+                _latency_cur = _latency_init;
+                _bandwidth_gbps_cur = _bandwidth_gbps_init;
+                return;
+
+        case LinkStatus::DOWN:
+                _status = LinkStatus::DOWN;
+                return;
+
+        case LinkStatus::THROTTLING:
+                if (new_latency < _latency_init ||
+                    new_bandwidth_gbps > _bandwidth_gbps_init) {
+                        log_error("ERROR",
+                                  "Link::set_status(THROTTLING) ignored: invalid latency/bandwidth");
+                        return;
+                }
+
+                _status = LinkStatus::THROTTLING;
+                _latency_cur = new_latency;
+                _bandwidth_gbps_cur = new_bandwidth_gbps;
+                return;
+        }
+}
+
+} // namespace dofs
diff --git a/src/network/link.h b/src/network/link.h
new file mode 100644
index 0000000..3d24e73
--- /dev/null
+++ b/src/network/link.h
@@ -0,0 +1,120 @@
+#ifndef NETWORK_LINK_H
+#define NETWORK_LINK_H
+
+#include <cstdint>
+#include <optional>
+#include <functional>
+
+#include "core/time.h"
+#include "core/types.h"
+#include "core/node.h"
+#include "core/simulator.h"
+#include "core/error.h"
+#include "network/packet.h"
+#include "network/network_node.h"
+
+namespace dofs {
+
+// Link represents a full-duplex point-to-point link between two NetworkNodes.
+// It exposes a reservation-based timing API; the switch owns scheduling.
+//
+// API for switch-owned scheduling:
+//   - next_available(sender): when this direction can start sending next
+//   - serialization_time(bytes): how long to serialize 'bytes' at current bw
+//   - propagation_latency(): current one-way latency
+//   - reserve(bytes, sender): returns {start, finish} and advances cursor
+//   - send_pkt(pkt, caller): uses reserve() + schedules delivery itself.
+class Link {
+public:
+        struct Reservation {
+                Time start;   // serialization start time on the link
+                Time finish;  // serialization finish time on the link
+        };
+
+        Link(Simulator *const sim,
+             LinkId id,
+             NetworkNode *a, PortId a_port,
+             NetworkNode *b, PortId b_port,
+             Time latency,
+             double bandwidth_gbps) noexcept;
+
+        void send_pkt(Packet &pkt, NodeId caller);
+
+        void schedule_delivery_after(Packet &pkt, NodeId caller, Time after);
+
+        Time next_available(NodeId sender) const noexcept;
+
+        std::optional<Reservation> reserve(Bytes bytes, NodeId sender) noexcept;
+
+        Time serialization_time(Bytes bytes) const noexcept {
+                return serialization_time(bytes, _bandwidth_gbps_cur);
+        }
+
+        Time propagation_latency() const noexcept {
+                return _latency_cur;
+        }
+
+        LinkId id() const noexcept {
+                return _id;
+        }
+        LinkStatus status() const noexcept {
+                return _status;
+        }
+        double bandwidth_gbps() const noexcept {
+                return _bandwidth_gbps_cur;
+        }
+
+        // Endpoint accessors (A/B as constructed)
+        NodeId src_id() const noexcept {
+                return _a_id;        // 'A' end
+        }
+        NodeId dst_id() const noexcept {
+                return _b_id;        // 'B' end
+        }
+        PortId src_port() const noexcept {
+                return _a_port;
+        }
+        PortId dst_port() const noexcept {
+                return _b_port;
+        }
+
+        // Status changes:
+        // - set_status(OK): restore initial latency/bandwidth
+        // - set_status(DOWN): link becomes inactive (reservations/sends rejected)
+        // - set_status(THROTTLING, new_latency, new_bw_gbps): degrade
+        //   (reject if new_latency < initial or new_bw_gbps > initial)
+        void set_status(LinkStatus s,
+                        Time new_latency = Time(0),
+                        double new_bandwidth_gbps = 0.0);
+
+        static Time serialization_time(Bytes bytes, double gbps) noexcept;
+
+private:
+        enum class Dir : uint8_t { AtoB, BtoA };
+
+        std::optional<Dir> _dir_from(NodeId sender) const noexcept;
+
+        Simulator *const _sim;
+
+        LinkId _id;
+
+        NetworkNode *_a;
+        NetworkNode *_b;
+        NodeId _a_id;
+        NodeId _b_id;
+        PortId _a_port;
+        PortId _b_port;
+
+        Time _latency_init;
+        double _bandwidth_gbps_init;
+        Time _latency_cur;
+        double _bandwidth_gbps_cur;
+        LinkStatus _status;
+
+        Time _next_available_ab;
+        Time _next_available_ba;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_LINK_H
diff --git a/src/network/network_nic.cc b/src/network/network_nic.cc
new file mode 100644
index 0000000..f587d41
--- /dev/null
+++ b/src/network/network_nic.cc
@@ -0,0 +1,660 @@
+#include "network/network_nic.h"
+
+#include <algorithm>
+#include <utility>
+#include "core/host.h"
+#include "core/error.h"
+
+namespace dofs {
+
+NetworkNic::NetworkNic(Simulator *const sim,
+                       NodeId id,
+                       uint16_t total_ports,
+                       SwitchBuffer *const buf,
+                       Time nic_latency,
+                       Bytes mice_elephant_threshold,
+                       PacketSeq ooo_threshold,
+                       CCType cc_type,
+                       LBType lb_type,
+                       Bytes cc_init_cwnd,
+                       Bytes cc_max_cwnd,
+                       const NicSchedulingWeights& schedw) noexcept
+        : NetworkNode(sim, id, NodeType::NIC),
+          _buf(buf),
+          _port_cnt(total_ports),
+          _nic_latency_init(nic_latency),
+          _nic_latency_cur(nic_latency),
+          _mice_thresh(mice_elephant_threshold),
+          _ooo_thresh(ooo_threshold),
+          _cc_type(cc_type),
+          _lb_type(lb_type),
+          _schedw(schedw),
+          _ports(total_ports) {
+
+        Node::set_status(NodeStatus::OK);
+
+        _port_blacklisted.assign(_port_cnt, false);
+
+        if (!_buf) {
+                log_error("ERROR", "NetworkNic constructed with null SwitchBuffer");
+        }
+
+        switch (_cc_type) {
+        case CCType::DCQCN:
+                _cc = std::make_unique<DCQCN>(cc_init_cwnd, cc_max_cwnd);
+                break;
+
+        case CCType::NSCC:
+                _cc = std::make_unique<NSCC>(cc_init_cwnd, cc_max_cwnd);
+                break;
+
+        default:
+                _cc = std::make_unique<DCQCN>(cc_init_cwnd, cc_max_cwnd);
+                break;
+        }
+
+        switch (_lb_type) {
+        case LBType::RANDOM_PACKET_SPRAYING:
+                _lb = std::make_unique<LBRandomPacketSpraying>(&_rng);
+                break;
+
+        default:
+                _lb = std::make_unique<LBRandomPacketSpraying>(&_rng);
+                break;
+        }
+
+        for (auto& pq : _ports) {
+                pq.b_control  = _schedw.control;
+                pq.b_retrans  = _schedw.retrans;
+                pq.b_mice     = _schedw.mice;
+                pq.b_elephant = _schedw.elephant;
+        }
+}
+
+void NetworkNic::attach_host(Host* host) noexcept {
+        _host = host;
+
+        if (_host)
+                _host->attach_nic(this);
+}
+
+void NetworkNic::detach_host(Host* host) noexcept {
+        if (_host && _host != host) {
+                log_error("ERROR", "NetworkNic::detach_host pointer mismatch");
+                return;
+        }
+
+        if (_host)
+                _host->detach_nic(this);
+
+        _host = nullptr;
+}
+
+void NetworkNic::set_status(NodeStatus s, Time new_latency) noexcept {
+        switch (s) {
+        case NodeStatus::OK:
+                Node::set_status(NodeStatus::OK);
+                _nic_latency_cur = _nic_latency_init;
+                return;
+
+        case NodeStatus::THROTTLING:
+                if (new_latency < _nic_latency_init) {
+                        log_error("ERROR", "NetworkNic::set_status THROTTLING new latency < initial");
+                        return;
+                }
+
+                Node::set_status(NodeStatus::THROTTLING);
+                _nic_latency_cur = new_latency;
+                return;
+
+        case NodeStatus::DOWN:
+        case NodeStatus::BOOTING:
+                log_error("ERROR", "NetworkNic::set_status ignored for DOWN/BOOTING");
+                return;
+        }
+}
+
+bool NetworkNic::is_elephant(Bytes sz) const noexcept {
+        return sz >= _mice_thresh;
+}
+
+std::uint64_t NetworkNic::txkey(FlowId f, PacketSeq s) noexcept {
+        return (static_cast<std::uint64_t>(static_cast<std::uint32_t>(f)) << 16) |
+               static_cast<std::uint64_t>(static_cast<std::uint16_t>(s));
+}
+
+void NetworkNic::record_tx_timestamp(const Packet& pkt) noexcept {
+        if (!_sim)
+                return;
+
+        const auto key = txkey(pkt.flow_id(), pkt.seq());
+        _tx_sent_at[key] = _sim->now();
+}
+
+bool NetworkNic::lookup_rtt_and_erase(const Packet& ack_like, Time now,
+                                      Time& out_rtt) noexcept {
+        const auto key = txkey(ack_like.flow_id(), ack_like.seq());
+        auto it = _tx_sent_at.find(key);
+
+        if (it == _tx_sent_at.end())
+                return false;
+
+        out_rtt = now.unsafe_sub(it->second);
+        _tx_sent_at.erase(it);
+        return true;
+}
+
+void NetworkNic::recv_pkt(Packet &pkt_in, PortId ingress) {
+        if (status() == NodeStatus::DOWN || status() == NodeStatus::BOOTING) {
+                return;
+        }
+
+        Packet pkt = pkt_in;
+        const Time now = _sim ? _sim->now() : Time(0);
+
+        switch (pkt.protocol()) {
+        case PacketProtocol::DATA: {
+                _telemetry.rx_data++;
+
+                // Auto-ACK after NIC latency
+                schedule_ack(pkt, ingress);
+
+                // For OOO detection & NACK: basic placeholder based on seq gap.
+                // (A real impl would maintain RxFlowState with expected_seq.)
+                // Here we keep it minimal and do not NACK unless explicitly enabled later.
+
+                // Flow completion: if EOF, notify host with total size from payload (accumulation would live in RxFlowState).
+                if (pkt.is_eof() && _host) {
+                        _host->recv_flow(pkt.src_node(), pkt.priority(), pkt.payload_size());
+                }
+
+                break;
+        }
+
+        case PacketProtocol::ACK: {
+                _telemetry.rx_acks++;
+
+                // RTT measurement and CC update
+                Time rtt;
+
+                if (lookup_rtt_and_erase(pkt, now, rtt)) {
+                        if (_cc) {
+                                _cc->update(pkt, rtt);
+                                _telemetry.cc_updates++;
+                        }
+                } else {
+                        if (_cc) {
+                                _cc->update(pkt, Time(0));
+                                _telemetry.cc_updates++;
+                        }
+                }
+
+                if (_lb) {
+                        _lb->update(pkt);
+                        _telemetry.lb_updates++;
+                }
+
+                if (_host)
+                        _host->recv_frame(pkt);
+
+                break;
+        }
+
+        case PacketProtocol::NACK: {
+                _telemetry.rx_nacks++;
+
+                if (_cc) {
+                        _cc->update(pkt, Time(0));
+                        _telemetry.cc_updates++;
+                }
+
+                if (_lb) {
+                        _lb->update(pkt);
+                        _telemetry.lb_updates++;
+                }
+
+                if (_host)
+                        _host->recv_frame(pkt);
+
+                // Retransmission path would re-enqueue missing seq as CONTROL/RETRANS.
+                break;
+        }
+
+        case PacketProtocol::HEADER_TRIM: {
+                _telemetry.rx_trims++;
+                schedule_trim_back_response(pkt, ingress);
+
+                if (_cc) {
+                        _cc->update(pkt, Time(0));
+                        _telemetry.cc_updates++;
+                }
+
+                if (_lb) {
+                        _lb->update(pkt);
+                        _telemetry.lb_updates++;
+                }
+
+                if (_host)
+                        _host->recv_frame(pkt);
+
+                break;
+        }
+
+        case PacketProtocol::HEADER_TRIM_BACK: {
+                _telemetry.rx_trims++;
+
+                if (_cc) {
+                        _cc->update(pkt, Time(0));
+                        _telemetry.cc_updates++;
+                }
+
+                if (_lb) {
+                        _lb->update(pkt);
+                        _telemetry.lb_updates++;
+                }
+
+                if (_host)
+                        _host->recv_frame(pkt);
+
+                break;
+        }
+
+        default:
+                // Unknown proto: ignore
+                break;
+        }
+}
+
+void NetworkNic::send_flow(NodeId dst, Bytes size, FlowPriority desired) {
+        const bool ele = is_elephant(size);
+        const FlowPriority prio = (desired == FlowPriority::AUTO)
+                                  ? (ele ? FlowPriority::ELEPHANT : FlowPriority::MICE)
+                                  : desired;
+        start_tx_unicast(dst, size, prio);
+}
+
+void NetworkNic::send_flow(PacketGroups gmask, Bytes size,
+                           FlowPriority desired) {
+        const bool ele = is_elephant(size);
+        const FlowPriority prio = (desired == FlowPriority::AUTO)
+                                  ? (ele ? FlowPriority::ELEPHANT : FlowPriority::MICE)
+                                  : desired;
+        start_tx_multicast(gmask, size, prio);
+}
+
+void NetworkNic::set_port_blacklisted(PortId port, bool blacklisted) noexcept {
+        if (port >= _port_cnt) {
+                log_error("ERROR", "NetworkNic::set_port_blacklisted invalid port");
+                return;
+        }
+
+        if (port >= _port_blacklisted.size())
+                _port_blacklisted.resize(_port_cnt, false);
+
+        _port_blacklisted[port] = blacklisted;
+}
+
+bool NetworkNic::is_port_blacklisted(PortId port) const noexcept {
+        if (port >= _port_cnt || port >= _port_blacklisted.size())
+                return false;
+
+        return _port_blacklisted[port];
+}
+
+PortId NetworkNic::pick_src_port_for_flow(NodeId dst) noexcept {
+        (void)dst;
+
+        if (_port_cnt == 0)
+                return static_cast<PortId>(0);
+
+        // Simple RR scheme
+        const uint16_t start = _rr_cursor % _port_cnt;
+
+        for (uint16_t i = 0; i < _port_cnt; ++i) {
+                const PortId cand = static_cast<PortId>((start + i) % _port_cnt);
+
+                if (cand < _port_blacklisted.size() && !_port_blacklisted[cand]) {
+                        _rr_cursor = static_cast<PortId>((cand + 1) % _port_cnt);
+                        return cand;
+                }
+        }
+
+        log_error("WARN",
+                  "NetworkNic: all ports blacklisted; choosing RR port despite blacklist");
+        const PortId fallback = static_cast<PortId>(start);
+        _rr_cursor = static_cast<PortId>((fallback + 1) % _port_cnt);
+        return fallback;
+}
+
+Packet NetworkNic::make_data_packet(NodeId dst, PortId out_port,
+                                    FlowPriority prio,
+                                    FlowId fid, PacketSeq seq, Bytes payload) noexcept {
+        Packet p(this->id(), out_port,
+                 dst, static_cast<PortId>(0),
+                 PacketProtocol::DATA,
+                 prio, seq, fid, 0, 0, payload);
+
+        if (_lb) {
+                const uint16_t e = _lb->get_entropy(p);
+                p.set_entropy(e);
+        }
+
+        return p;
+}
+
+Packet NetworkNic::make_ack_packet(const Packet& rx_data,
+                                   PortId ingress) noexcept {
+        Packet ack(rx_data.dst_node(), rx_data.dst_port(),
+                   rx_data.src_node(), rx_data.src_port(),
+                   PacketProtocol::ACK,
+                   FlowPriority::CTRL,
+                   rx_data.seq(), rx_data.flow_id(),
+                   rx_data.entropy(), 0, Bytes(0));
+        (void)ingress;
+        return ack;
+}
+
+Packet NetworkNic::make_nack_packet(NodeId peer, FlowId flow, PacketSeq miss,
+                                    FlowPriority prio,
+                                    PortId ingress) noexcept {
+        Packet nack(this->id(), ingress,
+                    peer, static_cast<PortId>(0),
+                    PacketProtocol::NACK,
+                    prio, miss, flow, 0, 0, Bytes(0));
+        return nack;
+}
+
+Packet NetworkNic::make_trim_back_response(const Packet& trim,
+                PortId ingress) noexcept {
+        Packet tb(trim.dst_node(), trim.dst_port(),
+                  trim.src_node(), trim.src_port(),
+                  PacketProtocol::HEADER_TRIM_BACK,
+                  FlowPriority::CTRL,
+                  trim.seq(), trim.flow_id(),
+                  trim.entropy(), 0, Bytes(0));
+        (void)ingress;
+        return tb;
+}
+
+void NetworkNic::schedule_ack(const Packet& rx_data, PortId ingress) noexcept {
+        if (!_sim)
+                return;
+
+        const Packet ack = make_ack_packet(rx_data, ingress);
+        const PortId egress = static_cast<PortId> (ingress);
+
+        _sim->schedule_after(_nic_latency_cur,
+        [this, ack, egress]() mutable {
+                enqueue_packet(egress, QClass::CONTROL, ack);
+        });
+        _telemetry.tx_acks++;
+}
+
+void NetworkNic::schedule_nack(NodeId peer, FlowId flow, PacketSeq missing_seq,
+                               FlowPriority prio, PortId ingress) noexcept {
+        if (!_sim)
+                return;
+
+        const Packet nack = make_nack_packet(peer, flow, missing_seq, prio, ingress);
+        const PortId egress = ingress;
+
+        _sim->schedule_after(_nic_latency_cur,
+        [this, nack, egress]() mutable {
+                enqueue_packet(egress, QClass::CONTROL, nack);
+        });
+        _telemetry.tx_nacks++;
+}
+
+void NetworkNic::schedule_trim_back_response(const Packet& trim,
+                PortId ingress) noexcept {
+        if (!_sim)
+                return;
+
+        const Packet tb = make_trim_back_response(trim, ingress);
+        const PortId egress = ingress;
+
+        _sim->schedule_after(_nic_latency_cur,
+        [this, tb, egress]() mutable {
+                enqueue_packet(egress, QClass::CONTROL, tb);
+        });
+        _telemetry.tx_trims++;
+}
+
+void NetworkNic::start_tx_unicast(NodeId dst, Bytes size, FlowPriority prio) {
+        // Minimal segmentation loop: 1 flow_id per call; simple seq starting at 0
+        static FlowId next_flow = 1;
+        const FlowId fid = next_flow++;
+
+        PortId port = pick_src_port_for_flow(dst);
+
+        Bytes remaining = size;
+        PacketSeq seq = 0;
+
+        while (remaining > Bytes(0)) {
+                const Bytes payload = std::min(remaining, DEFAULT_MSS_BYTES);
+                Packet pkt = make_data_packet(dst, port, prio, fid, seq, payload);
+
+                if (remaining <= DEFAULT_MSS_BYTES) {
+                        pkt.set_eof(true);
+                }
+
+                enqueue_packet(port, (prio == FlowPriority::ELEPHANT) ? QClass::ELEPHANT :
+                               QClass::MICE, pkt);
+
+                _telemetry.tx_data++;
+                remaining -= payload;
+                seq = static_cast<PacketSeq>(static_cast<std::uint16_t>(seq + 1));
+        }
+}
+
+void NetworkNic::start_tx_multicast(PacketGroups gmask, Bytes size,
+                                    FlowPriority prio) {
+        static FlowId next_flow = 1;
+        const FlowId fid = next_flow++;
+
+        PortId port = static_cast<PortId>(0);
+
+        Bytes remaining = size;
+        PacketSeq seq = 0;
+
+        while (remaining > Bytes(0)) {
+                const Bytes payload = std::min(remaining, DEFAULT_MSS_BYTES);
+
+                Packet pkt(this->id(), port, NodeId(0), PortId(0),
+                           PacketProtocol::DATA, prio, seq, fid, 0, 0, payload);
+                pkt.set_groups(gmask);
+
+                if (_lb) {
+                        const uint16_t e = _lb->get_entropy(pkt);
+                        pkt.set_entropy(e);
+                }
+
+                if (remaining <= DEFAULT_MSS_BYTES) {
+                        pkt.set_eof(true);
+                }
+
+                enqueue_packet(port, (prio == FlowPriority::ELEPHANT) ? QClass::ELEPHANT :
+                               QClass::MICE, pkt);
+
+                _telemetry.tx_data++;
+                remaining -= payload;
+                seq = static_cast<PacketSeq>(static_cast<std::uint16_t>(seq + 1));
+        }
+}
+
+void NetworkNic::enqueue_packet(PortId port, QClass cls, Packet pkt) noexcept {
+        if (port >= _port_cnt) {
+                log_error("ERROR", "NetworkNic::enqueue_packet invalid port");
+                return;
+        }
+
+        auto& pq = _ports[port];
+
+        switch (cls) {
+        case QClass::CONTROL:
+                pq.control.push_back(std::move(pkt));
+                break;
+
+        case QClass::RETRANS:
+                pq.retrans.push_back(std::move(pkt));
+                break;
+
+        case QClass::MICE:
+                pq.mice.push_back(std::move(pkt));
+                break;
+
+        case QClass::ELEPHANT:
+                pq.elephant.push_back(std::move(pkt));
+                break;
+        }
+
+        schedule_port_if_needed(port);
+}
+
+void NetworkNic::schedule_port_if_needed(PortId port) noexcept {
+        auto& pq = _ports[port];
+
+        if (pq.scheduled || !_sim)
+                return;
+
+        pq.scheduled = true;
+        _sim->schedule_after(Time(0),
+                             std::mem_fn(&NetworkNic::port_drain_task),
+                             this, port);
+}
+
+bool NetworkNic::pick_next_qclass(const PortQueues& pq,
+                                  QClass& out_cls) const noexcept {
+        if (!pq.control.empty()  && pq.b_control  > Bytes(0))  {
+                out_cls = QClass::CONTROL;
+                return true;
+        }
+
+        if (!pq.retrans.empty()  && pq.b_retrans  > Bytes(0))  {
+                out_cls = QClass::RETRANS;
+                return true;
+        }
+
+        if (!pq.mice.empty()     && pq.b_mice     > Bytes(0))  {
+                out_cls = QClass::MICE;
+                return true;
+        }
+
+        if (!pq.elephant.empty() && pq.b_elephant > Bytes(0))  {
+                out_cls = QClass::ELEPHANT;
+                return true;
+        }
+
+        return false;
+}
+
+bool NetworkNic::try_send_one(PortId port, QClass cls) noexcept {
+        auto& pq = _ports[port];
+        std::deque<Packet> *q = nullptr;
+        Bytes* budget = nullptr;
+
+        switch (cls) {
+        case QClass::CONTROL:
+                q = &pq.control;
+                budget = &pq.b_control;
+                break;
+
+        case QClass::RETRANS:
+                q = &pq.retrans;
+                budget = &pq.b_retrans;
+                break;
+
+        case QClass::MICE:
+                q = &pq.mice;
+                budget = &pq.b_mice;
+                break;
+
+        case QClass::ELEPHANT:
+                q = &pq.elephant;
+                budget = &pq.b_elephant;
+                break;
+        }
+
+        if (!q || q->empty() || *budget <= Bytes(0))
+                return false;
+
+        Packet pkt = std::move(q->front());
+        q->pop_front();
+
+        const bool is_ctrl = (cls == QClass::CONTROL || cls == QClass::RETRANS);
+
+        if (!is_ctrl && _cc) {
+                const Bytes next_bytes = pkt.total_size();
+                Bytes& out = _tx_outstanding[pkt.flow_id()];
+
+                if (!_cc->is_allowed_to_send(out, next_bytes)) {
+                        q->push_front(std::move(pkt));
+                        return false;
+                }
+
+                out += next_bytes;
+        }
+
+        if (_sim) {
+                _sim->schedule_after(_nic_latency_cur,
+                [this, pkt, port]() mutable {
+                        if (pkt.protocol() == PacketProtocol::DATA) {
+                                record_tx_timestamp(pkt);
+                        }
+                        (void)_buf->enqueue_packet(pkt, port, pkt.priority());
+                });
+        }
+
+        const Bytes cost = pkt.total_size();
+
+        if (cost >= *budget)
+                *budget = Bytes(0);
+        else *budget = *budget - cost;
+
+        return true;
+}
+
+void NetworkNic::port_drain_task(PortId port) noexcept {
+        auto& pq = _ports[port];
+
+        // Attempt to send while there is any budget and any queue has packets.
+        // Refill budgets when all depleted or all queues empty.
+        for (;;) {
+                QClass cls;
+
+                if (pick_next_qclass(pq, cls)) {
+                        if (!try_send_one(port, cls)) {
+                                // Could not send, stop and retry later.
+                                break;
+                        }
+
+                        continue;
+                }
+
+                const bool any_nonempty = !(pq.control.empty() && pq.retrans.empty() &&
+                                            pq.mice.empty() && pq.elephant.empty());
+                const bool any_budget = (pq.b_control  > Bytes(0) ||
+                                         pq.b_retrans  > Bytes(0) ||
+                                         pq.b_mice     > Bytes(0) ||
+                                         pq.b_elephant > Bytes(0));
+
+                if (!any_budget && any_nonempty) {
+                        pq.b_control  = _schedw.control;
+                        pq.b_retrans  = _schedw.retrans;
+                        pq.b_mice     = _schedw.mice;
+                        pq.b_elephant = _schedw.elephant;
+                        continue;
+                }
+
+                break;
+        }
+
+        pq.scheduled = false;
+
+        if (!(pq.control.empty() && pq.retrans.empty() &&
+              pq.mice.empty() && pq.elephant.empty())) {
+                schedule_port_if_needed(port);
+        }
+}
+
+} // namespace dofs
diff --git a/src/network/network_nic.h b/src/network/network_nic.h
new file mode 100644
index 0000000..91423b0
--- /dev/null
+++ b/src/network/network_nic.h
@@ -0,0 +1,154 @@
+#ifndef NETWORK_NETWORK_NIC_H
+#define NETWORK_NETWORK_NIC_H
+
+#include <cstdint>
+#include <deque>
+#include <unordered_map>
+#include <vector>
+
+#include "core/time.h"
+#include "core/types.h"
+#include "core/rng.h"
+#include "network/network_node.h"
+#include "network/packet.h"
+#include "network/switch/switch_buffer.h"
+#include "network/nic/nic_telemetry.h"
+#include "network/nic/congestion_control.h"
+#include "network/nic/load_balance.h"
+
+namespace dofs {
+
+class Host;
+
+struct NicSchedulingWeights {
+        // Bytes of budget per round (refilled when all budgets deplete).
+        Bytes control  { 64 * 1024 };
+        Bytes retrans  { 256 * 1024 };
+        Bytes mice     { 1024 * 1024 };
+        Bytes elephant { 1024 * 1024 };
+};
+
+class NetworkNic final : public NetworkNode {
+public:
+        NetworkNic(Simulator *const sim,
+                   NodeId id,
+                   uint16_t total_ports,
+                   SwitchBuffer *const buf,
+                   Time nic_latency,
+                   Bytes mice_elephant_threshold,
+                   PacketSeq ooo_threshold,
+                   CCType cc_type,
+                   LBType lb_type,
+                   Bytes cc_init_cwnd,
+                   Bytes cc_max_cwnd,
+                   const NicSchedulingWeights& schedw) noexcept;
+
+        virtual void recv_pkt(Packet &pkt, PortId ingress) override;
+
+        void attach_host(Host* host) noexcept;
+        void detach_host(Host* host) noexcept;
+
+        // Host API (TX):
+        void send_flow(NodeId dst, Bytes size,
+                       FlowPriority desired = FlowPriority::AUTO);
+        void send_flow(PacketGroups group_mask, Bytes size,
+                       FlowPriority desired = FlowPriority::AUTO);
+
+        void set_status(NodeStatus s, Time new_latency = Time(0)) noexcept;
+
+        const NicTelemetry &telemetry() const noexcept {
+                return _telemetry;
+        }
+
+        void set_port_blacklisted(PortId port, bool blacklisted) noexcept;
+        bool is_port_blacklisted(PortId port) const noexcept;
+
+private:
+        enum class QClass : uint8_t { CONTROL = 0, RETRANS = 1, MICE = 2, ELEPHANT = 3 };
+
+        struct PortQueues {
+                std::deque<Packet> control;
+                std::deque<Packet> retrans;
+                std::deque<Packet> mice;
+                std::deque<Packet> elephant;
+
+                Bytes b_control  {0};
+                Bytes b_retrans  {0};
+                Bytes b_mice     {0};
+                Bytes b_elephant {0};
+
+                bool scheduled {false};
+        };
+
+        void schedule_port_if_needed(PortId port) noexcept;
+        void port_drain_task(PortId port) noexcept;
+
+        bool pick_next_qclass(const PortQueues& pq, QClass& out_cls) const noexcept;
+
+        bool try_send_one(PortId port, QClass cls) noexcept;
+
+        void enqueue_packet(PortId port, QClass cls, Packet pkt) noexcept;
+
+        void schedule_ack(const Packet& rx_data, PortId ingress) noexcept;
+        void schedule_nack(NodeId peer, FlowId flow, PacketSeq missing_seq,
+                           FlowPriority prio, PortId ingress) noexcept;
+        void schedule_trim_back_response(const Packet& trim, PortId ingress) noexcept;
+
+        bool is_elephant(Bytes sz) const noexcept;
+
+        static inline std::uint64_t txkey(FlowId f, PacketSeq s) noexcept;
+
+        void record_tx_timestamp(const Packet& pkt) noexcept;
+        bool lookup_rtt_and_erase(const Packet& ack_like, Time now,
+                                  Time& out_rtt) noexcept;
+
+        PortId pick_src_port_for_flow(NodeId dst) noexcept;
+
+        void start_tx_unicast(NodeId dst, Bytes size, FlowPriority prio);
+        void start_tx_multicast(PacketGroups gmask, Bytes size, FlowPriority prio);
+
+        Packet make_data_packet(NodeId dst, PortId out_port, FlowPriority prio,
+                                FlowId fid, PacketSeq seq, Bytes payload) noexcept;
+        Packet make_ack_packet(const Packet& rx_data, PortId ingress) noexcept;
+        Packet make_nack_packet(NodeId peer, FlowId flow, PacketSeq miss,
+                                FlowPriority prio,
+                                PortId ingress) noexcept;
+        Packet make_trim_back_response(const Packet& trim, PortId ingress) noexcept;
+
+private:
+        Host                  *_host {nullptr};
+        SwitchBuffer          *_buf  {nullptr};
+
+        const uint16_t         _port_cnt;
+        Time                   _nic_latency_init;
+        Time                   _nic_latency_cur;
+
+        const Bytes            _mice_thresh;
+        const PacketSeq        _ooo_thresh;
+
+        CCType                 _cc_type;
+        LBType                 _lb_type;
+
+        NicSchedulingWeights   _schedw;
+
+        NicTelemetry           _telemetry;
+        Rng                    _rng;
+
+        std::vector<PortQueues> _ports;
+
+        std::vector<bool>       _port_blacklisted;
+        PortId                  _rr_cursor {0};
+
+        std::unique_ptr<CongestionControl> _cc;
+        std::unique_ptr<LoadBalance>       _lb;
+
+        // Outstanding bytes per flow (TX) for cwnd gating
+        std::unordered_map<FlowId, Bytes> _tx_outstanding;
+
+        // TX timestamps for RTT: (flow, seq) -> send_time
+        std::unordered_map<std::uint64_t, Time> _tx_sent_at;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_NETWORK_NIC_H
diff --git a/src/network/network_node.cc b/src/network/network_node.cc
new file mode 100644
index 0000000..7f94a6e
--- /dev/null
+++ b/src/network/network_node.cc
@@ -0,0 +1,9 @@
+#include "network/network_node.h"
+
+namespace dofs {
+
+NetworkNode::NetworkNode(Simulator *const sim, NodeId id,
+                         NodeType type) noexcept
+        : Node(sim, id, type) {}
+
+} // namespace dofs
diff --git a/src/network/network_node.h b/src/network/network_node.h
new file mode 100644
index 0000000..9307ad0
--- /dev/null
+++ b/src/network/network_node.h
@@ -0,0 +1,24 @@
+#ifndef NETWORK_NETWORK_NODE_H
+#define NETWORK_NETWORK_NODE_H
+
+#include <cstdint>
+
+#include "core/node.h"
+#include "core/types.h"
+#include "network/packet.h"
+
+namespace dofs {
+
+class NetworkNode : public Node {
+public:
+
+        explicit NetworkNode(Simulator *const sim, NodeId id, NodeType type) noexcept;
+
+        virtual ~NetworkNode() = default;
+
+        virtual void recv_pkt(Packet &pkt, PortId ingress) = 0;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_NETWORK_NODE_H
diff --git a/src/network/network_switch.cc b/src/network/network_switch.cc
new file mode 100644
index 0000000..7bc3179
--- /dev/null
+++ b/src/network/network_switch.cc
@@ -0,0 +1,238 @@
+#include "network/network_switch.h"
+
+#include <algorithm>
+#include <iterator>
+#include <set>
+
+#include "network/switch/multicast_table.h"
+#include "network/switch/unicast_table.h"
+
+namespace dofs {
+
+NetworkSwitch::NetworkSwitch(Simulator *const sim,
+                             NodeId id,
+                             uint16_t total_ports,
+                             ECNEngine *const ecn,
+                             SwitchBuffer *const buf,
+                             const RoutingTables *const rt,
+                             Time forwarding_latency,
+                             Time multicast_dup_delay) noexcept
+        : NetworkNode(sim, id, NodeType::SWITCH),
+          _ecn(ecn),
+          _buf(buf),
+          _rt(rt),
+          _port_cnt(total_ports),
+          _fwd_latency_init(forwarding_latency),
+          _fwd_latency_cur(forwarding_latency),
+          _dup_delay(multicast_dup_delay) {
+
+        Node::set_status(NodeStatus::OK);
+
+        if (!_ecn) {
+                log_error("ERROR", "NetworkSwitch constructed with null ECN engine");
+        }
+
+        if (!_buf) {
+                log_error("ERROR", "NetworkSwitch constructed with null SwitchBuffer");
+        }
+
+        if (!_rt) {
+                log_error("ERROR", "NetworkSwitch constructed with null RoutingTables");
+        }
+}
+
+void NetworkSwitch::set_status(NodeStatus s,
+                               Time new_forward_latency) noexcept {
+        switch (s) {
+        case NodeStatus::OK:
+                Node::set_status(NodeStatus::OK);
+                _fwd_latency_cur = _fwd_latency_init;
+                return;
+
+        case NodeStatus::THROTTLING:
+                if (new_forward_latency < _fwd_latency_init) {
+                        log_error("ERROR",
+                                  "NetworkSwitch::set_status(THROTTLING) new latency < initial");
+                        return;
+                }
+
+                Node::set_status(NodeStatus::THROTTLING);
+                _fwd_latency_cur = new_forward_latency;
+                return;
+
+        case NodeStatus::DOWN:
+        case NodeStatus::BOOTING:
+                log_error("ERROR", "NetworkSwitch::set_status ignored for DOWN/BOOTING");
+                return;
+        }
+}
+
+void NetworkSwitch::recv_pkt(Packet &pkt_in, PortId ingress) {
+        // If switch is DOWN or BOOTING, drop silently
+        if (status() == NodeStatus::DOWN || status() == NodeStatus::BOOTING) {
+                return;
+        }
+
+        Packet pkt = pkt_in;
+
+        if (_ecn && _buf) {
+                pkt = _ecn->process_packet(pkt, _buf);
+        }
+
+        if (_sim) {
+                _sim->schedule_after(_fwd_latency_cur,
+                                     std::mem_fn(&NetworkSwitch::forward_after_delay),
+                                     this, pkt, ingress);
+        }
+}
+
+void NetworkSwitch::forward_after_delay(Packet pkt, PortId ingress) noexcept {
+        if (!_buf || !_rt)
+                return;
+
+        // Decide multicast vs unicast based on groups mask
+        std::vector<PortId> fanout;
+        fanout.reserve(8); // small typical fanout
+
+        const bool is_mc = (pkt.groups() != static_cast<PacketGroups>(0));
+
+        if (is_mc) {
+                build_mc_fanout(pkt, ingress, fanout);
+        } else {
+                build_uc_egress(pkt, fanout);
+        }
+
+        // Filter invalid/self ports and keep determinism
+        fanout.erase(std::remove_if(fanout.begin(), fanout.end(),
+        [&](PortId p) {
+                return (p >= _port_cnt) || (p == ingress);
+        }),
+        fanout.end());
+
+        if (fanout.empty())
+                return;
+
+        const FlowPriority prio = pkt.priority();
+
+        for (std::size_t i = 0; i < fanout.size(); ++i) {
+                const PortId egress = fanout[i];
+                const Time offset = _dup_delay * Time(i);
+
+                if (_sim && offset > Time(0)) {
+                        _sim->schedule_after(offset,
+                                             std::mem_fn(&NetworkSwitch::enqueue_one),
+                                             this, pkt, egress, prio);
+                } else {
+                        enqueue_one(pkt, egress, prio);
+                }
+        }
+}
+
+void NetworkSwitch::enqueue_one(Packet pkt, PortId egress,
+                                FlowPriority prio) noexcept {
+        (void)_buf->enqueue_packet(pkt, egress, prio);
+}
+
+void NetworkSwitch::build_uc_egress(const Packet& pkt,
+                                    std::vector<PortId> &out) const {
+        // Unicast: consult table; if multiple candidates, ECMP-pick one by entropy
+        std::vector<PortId> candidates =
+                _rt->unicast.get_port_list(pkt.dst_node(), pkt.dst_port());
+
+        if (candidates.empty()) {
+                // Fallback to dst_port if sane
+                if (pkt.dst_port() < _port_cnt) {
+                        out.clear();
+                        out.push_back(pkt.dst_port());
+                }
+
+                return;
+        }
+
+        if (candidates.size() == 1) {
+                out = candidates;
+                return;
+        }
+
+        // Multiple candidates: pick one deterministically via hash_ecmp
+        const uint32_t differentiator =
+                (static_cast<uint32_t>(pkt.entropy()) << 16) ^
+                static_cast<uint32_t>(pkt.dst_port());
+
+        const uint16_t idx = hash_ecmp(pkt.src_node(), pkt.src_port(),
+                                       pkt.dst_node(), pkt.dst_port(),
+                                       differentiator,
+                                       static_cast<uint16_t>(candidates.size()));
+
+        out.clear();
+        out.push_back(candidates[idx % candidates.size()]);
+}
+
+void NetworkSwitch::build_mc_fanout(const Packet& pkt, PortId ingress,
+                                    std::vector<PortId> &fanout) const {
+        if (!_rt)
+                return;
+
+        const PacketGroups mask = pkt.groups();
+        const std::size_t gcount = _rt->multicast.group_count();
+
+        for (std::size_t gid = 0; gid < gcount; ++gid) {
+                if (!group_bit_set(mask, gid))
+                        continue;
+
+                const auto* trees = _rt->multicast.trees_of(gid);
+
+                if (!trees || trees->empty())
+                        continue;
+
+                const uint16_t k = static_cast<uint16_t>(trees->size());
+                uint16_t chosen_idx = 0;
+
+                if (k > 1) {
+                        // Entropy-based multipath spraying:
+                        // differentiator mixes 16b entropy with 16b group index
+                        const uint32_t differentiator =
+                                (static_cast<uint32_t>(pkt.entropy()) << 16) ^
+                                static_cast<uint32_t>(gid & 0xFFFF);
+
+                        chosen_idx = hash_ecmp(pkt.src_node(), pkt.src_port(),
+                                               pkt.dst_node(), pkt.dst_port(),
+                                               differentiator, k);
+                        chosen_idx %= k;
+                }
+
+                const McTree& t = (*trees)[chosen_idx];
+
+                if (t.parent_port.has_value() && ingress != t.parent_port.value()) {
+                        // Skip this group's replication from the wrong direction
+                        continue;
+                }
+
+                merge_sorted_unique(fanout, t.child_ports);
+        }
+}
+
+void NetworkSwitch::merge_sorted_unique(std::vector<PortId> &base,
+                                        const std::vector<PortId> &add) {
+        if (add.empty())
+                return;
+
+        if (base.empty()) {
+                base = add;
+                return;
+        }
+
+        std::vector<PortId> merged;
+        merged.reserve(base.size() + add.size());
+        std::set_union(base.begin(), base.end(),
+                       add.begin(), add.end(),
+                       std::back_inserter(merged));
+        base.swap(merged);
+}
+
+inline bool NetworkSwitch::group_bit_set(PacketGroups mask,
+                std::size_t gid) noexcept {
+        return (mask >> gid) & static_cast<PacketGroups>(1);
+}
+
+} // namespace dofs
diff --git a/src/network/network_switch.h b/src/network/network_switch.h
new file mode 100644
index 0000000..4c46d5d
--- /dev/null
+++ b/src/network/network_switch.h
@@ -0,0 +1,81 @@
+#ifndef NETWORK_NETWORK_SWITCH_H
+#define NETWORK_NETWORK_SWITCH_H
+
+#include <cstdint>
+#include <vector>
+
+#include "core/time.h"
+#include "core/types.h"
+#include "core/error.h"
+#include "network/network_node.h"
+#include "network/packet.h"
+#include "network/switch/ecn_engine.h"
+#include "network/switch/switch_buffer.h"
+#include "network/switch/routing_tables.h"
+#include "network/switch/routing_alg.h"
+
+namespace dofs {
+
+class NetworkSwitch final : public NetworkNode {
+public:
+        NetworkSwitch(Simulator *const sim,
+                      NodeId id,
+                      uint16_t total_ports,
+                      ECNEngine *const ecn,
+                      SwitchBuffer *const buf,
+                      const RoutingTables *const rt,
+                      Time forwarding_latency,
+                      Time multicast_dup_delay) noexcept;
+
+        virtual void recv_pkt(Packet &pkt, PortId ingress) override;
+
+        // Status control:
+        // - OK: restore forwarding latency to initial
+        // - THROTTLING(new_latency): set _fwd_latency_cur = new_latency (>= initial)
+        // - DOWN/BOOTING: ignored with error log
+        void set_status(NodeStatus s, Time new_forward_latency = Time(0)) noexcept;
+
+        NodeStatus get_status() const noexcept {
+                return status();
+        }
+
+        uint16_t port_cnt() const noexcept {
+                return _port_cnt;
+        }
+
+private:
+        void forward_after_delay(Packet pkt, PortId ingress) noexcept;
+
+        // Schedules an enqueue of a single copy after delay offset
+        void enqueue_one(Packet pkt, PortId egress, FlowPriority prio) noexcept;
+
+        // Build multicast fanout: pick exactly one tree per active group bit,
+        // optionally enforcing RPF via parent_port, and union child ports.
+        void build_mc_fanout(const Packet& pkt, PortId ingress,
+                             std::vector<PortId> &fanout) const;
+
+        // Build unicast egress list: if multiple candidates, ECMP-pick one via hash_ecmp.
+        void build_uc_egress(const Packet& pkt, std::vector<PortId> &out) const;
+
+        static void merge_sorted_unique(std::vector<PortId> &base,
+                                        const std::vector<PortId> &add);
+
+        // Test if group bit gid is set in PacketGroups
+        static inline bool group_bit_set(PacketGroups mask, std::size_t gid) noexcept;
+
+private:
+        ECNEngine               *const _ecn;
+        SwitchBuffer            *const _buf;
+        const RoutingTables     *const _rt;
+
+        const uint16_t _port_cnt;
+
+        const Time _fwd_latency_init;
+        Time       _fwd_latency_cur;
+
+        const Time _dup_delay;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_NETWORK_SWITCH_H
diff --git a/src/network/nic/CMakeLists.txt b/src/network/nic/CMakeLists.txt
new file mode 100644
index 0000000..b220e31
--- /dev/null
+++ b/src/network/nic/CMakeLists.txt
@@ -0,0 +1,50 @@
+# --- Library: dofs_nic ---
+add_library(dofs_nic STATIC)
+
+target_include_directories(dofs_nic
+  PUBLIC
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>          # "network/...", "core/..."
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src/network>  # ** enables "nic/..." **
+    $<INSTALL_INTERFACE:include>
+)
+
+# Sources
+if(DOFS_GLOB_SOURCES)
+  file(GLOB NIC_CC CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.cc")
+  target_sources(dofs_nic PRIVATE ${NIC_CC})
+else()
+  target_sources(dofs_nic
+    PUBLIC
+	nic_telemetry.h
+	congestion_control.h
+      	load_balance.h
+    PRIVATE
+	congestion_control.cc
+      	load_balance.cc
+  )
+endif()
+
+target_link_libraries(dofs_nic PUBLIC dofs_network dofs_core)
+add_library(dofs::nic ALIAS dofs_nic)
+
+# --- Tooling stubs for nic headers ---
+file(GLOB NIC_HEADERS CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.h")
+
+set(NIC_STUBS)
+foreach(h ${NIC_HEADERS})
+  get_filename_component(_base "${h}" NAME_WE)
+  set(_stub "${CMAKE_CURRENT_BINARY_DIR}/nic_${_base}.cpp")
+  file(WRITE "${_stub}" "// stub nic/${_base}.h\n#include \"nic/${_base}.h\"\n")
+  list(APPEND NIC_STUBS "${_stub}")
+endforeach()
+
+add_library(dofs_nic_headers_tooling STATIC ${NIC_STUBS})
+
+# Make sure stubs see the SAME include roots
+target_include_directories(dofs_nic_headers_tooling
+  PRIVATE
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src/network>  # ** not src/network/nic **
+)
+
+target_link_libraries(dofs_nic_headers_tooling PRIVATE dofs_nic dofs_network dofs_core)
diff --git a/src/network/nic/congestion_control.cc b/src/network/nic/congestion_control.cc
new file mode 100644
index 0000000..321180d
--- /dev/null
+++ b/src/network/nic/congestion_control.cc
@@ -0,0 +1,100 @@
+#include "network/nic/congestion_control.h"
+
+namespace dofs {
+
+CongestionControl::CongestionControl(Bytes init_cwnd, Bytes max_cwnd) noexcept
+        : _cwnd(init_cwnd), _cwnd_max(max_cwnd) {
+        if (_cwnd > _cwnd_max)
+                _cwnd = _cwnd_max;
+
+        if (_cwnd == Bytes(0))
+                _cwnd = Bytes(1);
+}
+
+bool CongestionControl::is_allowed_to_send(Bytes bytes_outstanding,
+                Bytes next_bytes) const noexcept {
+        const Bytes needed = bytes_outstanding + next_bytes;
+        return needed <= _cwnd;
+}
+
+// IMPORTANT: placeholders
+DCQCN::DCQCN(Bytes init_cwnd, Bytes max_cwnd) noexcept
+        : CongestionControl(init_cwnd, max_cwnd) {}
+
+void DCQCN::update(const Packet& pkt, Time rtt) noexcept {
+        // Extremely simple placeholder logic:
+        // - If ECN marked or TRIM_BACK/NACK, cut cwnd multiplicatively.
+        // - If ACK of data (no ECN), increase cwnd additively (slowly).
+        // This gives you a cwnd-based shape; replace with real DCQCN later.
+
+        const bool is_ctrl = (pkt.protocol() == PacketProtocol::ACK ||
+                              pkt.protocol() == PacketProtocol::NACK ||
+                              pkt.protocol() == PacketProtocol::HEADER_TRIM_BACK);
+
+        if (pkt.protocol() == PacketProtocol::HEADER_TRIM_BACK ||
+            pkt.protocol() == PacketProtocol::NACK ||
+            pkt.is_ecn()) {
+                // MD: halve cwnd, floor at 1 MSS
+                Bytes new_cwnd = _cwnd / 2;
+
+                if (new_cwnd < DEFAULT_MSS_BYTES)
+                        new_cwnd = DEFAULT_MSS_BYTES;
+
+                _cwnd = new_cwnd;
+
+                if (_cwnd > _cwnd_max)
+                        _cwnd = _cwnd_max;
+
+                (void)rtt;
+                (void)is_ctrl;
+                return;
+        }
+
+        // Additive increase on clean ACKs; bound by cwnd_max.
+        if (pkt.protocol() == PacketProtocol::ACK) {
+                Bytes inc = DEFAULT_MSS_BYTES; // 1 MSS per RTT-ish sample
+                Bytes new_cwnd = _cwnd + inc;
+
+                if (new_cwnd > _cwnd_max)
+                        new_cwnd = _cwnd_max;
+
+                _cwnd = new_cwnd;
+                (void)rtt;
+                (void)is_ctrl;
+        }
+}
+
+NSCC::NSCC(Bytes init_cwnd, Bytes max_cwnd) noexcept
+        : CongestionControl(init_cwnd, max_cwnd) {}
+
+void NSCC::update(const Packet& pkt, Time rtt) noexcept {
+        if (pkt.protocol() == PacketProtocol::HEADER_TRIM_BACK ||
+            pkt.protocol() == PacketProtocol::NACK ||
+            pkt.is_ecn()) {
+                Bytes new_cwnd = (_cwnd * 3) / 4; // 25% cut
+
+                if (new_cwnd < DEFAULT_MSS_BYTES)
+                        new_cwnd = DEFAULT_MSS_BYTES;
+
+                _cwnd = new_cwnd;
+
+                if (_cwnd > _cwnd_max)
+                        _cwnd = _cwnd_max;
+
+                (void)rtt;
+                return;
+        }
+
+        if (pkt.protocol() == PacketProtocol::ACK) {
+                Bytes inc = DEFAULT_MSS_BYTES;
+                Bytes new_cwnd = _cwnd + inc;
+
+                if (new_cwnd > _cwnd_max)
+                        new_cwnd = _cwnd_max;
+
+                _cwnd = new_cwnd;
+                (void)rtt;
+        }
+}
+
+} // namespace dofs
diff --git a/src/network/nic/congestion_control.h b/src/network/nic/congestion_control.h
new file mode 100644
index 0000000..1fe6465
--- /dev/null
+++ b/src/network/nic/congestion_control.h
@@ -0,0 +1,47 @@
+#ifndef NETWORK_NIC_CONGESTION_CONTROL_H
+#define NETWORK_NIC_CONGESTION_CONTROL_H
+
+#include <cstdint>
+#include "core/time.h"
+#include "core/types.h"
+#include "network/packet.h"
+
+namespace dofs {
+
+class CongestionControl {
+public:
+        explicit CongestionControl(Bytes init_cwnd, Bytes max_cwnd) noexcept;
+        virtual ~CongestionControl() = default;
+
+        virtual void update(const Packet& pkt, Time rtt) noexcept = 0;
+
+        virtual bool is_allowed_to_send(Bytes bytes_outstanding,
+                                        Bytes next_bytes) const noexcept;
+
+        Bytes cwnd() const noexcept {
+                return _cwnd;
+        }
+        Bytes cwnd_max() const noexcept {
+                return _cwnd_max;
+        }
+
+protected:
+        Bytes _cwnd;
+        Bytes _cwnd_max;
+};
+
+class DCQCN final : public CongestionControl {
+public:
+        explicit DCQCN(Bytes init_cwnd, Bytes max_cwnd) noexcept;
+        virtual void update(const Packet& pkt, Time rtt) noexcept override;
+};
+
+class NSCC final : public CongestionControl {
+public:
+        explicit NSCC(Bytes init_cwnd, Bytes max_cwnd) noexcept;
+        virtual void update(const Packet& pkt, Time rtt) noexcept override;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_NIC_CONGESTION_CONTROL_H
diff --git a/src/network/nic/load_balance.cc b/src/network/nic/load_balance.cc
new file mode 100644
index 0000000..62af324
--- /dev/null
+++ b/src/network/nic/load_balance.cc
@@ -0,0 +1,18 @@
+#include "network/nic/load_balance.h"
+
+namespace dofs {
+
+void LBRandomPacketSpraying::update(const Packet& pkt) noexcept {
+        (void)pkt;
+}
+
+uint16_t LBRandomPacketSpraying::get_entropy(const Packet& context) noexcept {
+        (void)context;
+
+        if (!_rng)
+                return 0;
+
+        return static_cast<uint16_t>(_rng->uniform_range<std::uint32_t>(0, 0x10000u));
+}
+
+} // namespace dofs
diff --git a/src/network/nic/load_balance.h b/src/network/nic/load_balance.h
new file mode 100644
index 0000000..7dc9c50
--- /dev/null
+++ b/src/network/nic/load_balance.h
@@ -0,0 +1,36 @@
+#ifndef NETWORK_NIC_LOAD_BALANCE_H
+#define NETWORK_NIC_LOAD_BALANCE_H
+
+#include <cstdint>
+#include "network/packet.h"
+#include "core/rng.h"
+
+namespace dofs {
+
+// Base class for NIC load balancing policy.
+// Only exposes update() and get_entropy().
+class LoadBalance {
+public:
+        explicit LoadBalance(Rng *const rng) noexcept : _rng(rng) {}
+        virtual ~LoadBalance() = default;
+
+        virtual void update(const Packet& pkt) noexcept = 0;
+
+        // Produce a 16-bit entropy value for the next packet.
+        virtual uint16_t get_entropy(const Packet& context) noexcept = 0;
+
+protected:
+        Rng *const _rng;
+};
+
+// Random per-packet spraying (testing/baseline).
+class LBRandomPacketSpraying final : public LoadBalance {
+public:
+        explicit LBRandomPacketSpraying(Rng *const rng) noexcept : LoadBalance(rng) {}
+        virtual void update(const Packet& pkt) noexcept override;
+        virtual uint16_t get_entropy(const Packet& context) noexcept override;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_NIC_LOAD_BALANCE_H
diff --git a/src/network/nic/nic_telemetry.h b/src/network/nic/nic_telemetry.h
new file mode 100644
index 0000000..a535c30
--- /dev/null
+++ b/src/network/nic/nic_telemetry.h
@@ -0,0 +1,28 @@
+#ifndef NETWORK_NIC_NIC_TELEMETRY_H
+#define NETWORK_NIC_NIC_TELEMETRY_H
+
+#include <cstdint>
+
+namespace dofs {
+
+struct NicTelemetry {
+        std::uint64_t rx_acks        {0};
+        std::uint64_t rx_nacks       {0};
+        std::uint64_t rx_trims       {0};
+        std::uint64_t rx_data        {0};
+
+        std::uint64_t tx_acks        {0};
+        std::uint64_t tx_nacks       {0};
+        std::uint64_t tx_trims       {0};
+        std::uint64_t tx_data        {0};
+
+        std::uint64_t retrans        {0};
+        std::uint64_t timeouts       {0};
+
+        std::uint64_t cc_updates     {0};
+        std::uint64_t lb_updates     {0};
+};
+
+} // namespace dofs
+
+#endif // NETWORK_NIC_NIC_TELEMETRY_H
diff --git a/src/network/packet.cc b/src/network/packet.cc
new file mode 100644
index 0000000..242f856
--- /dev/null
+++ b/src/network/packet.cc
@@ -0,0 +1,180 @@
+#include "network/packet.h"
+
+namespace dofs {
+
+Packet::Packet(NodeId src_node,
+               PortId src_port,
+               NodeId dst_node,
+               PortId dst_port,
+               PacketProtocol proto,
+               FlowPriority prio,
+               PacketSeq seq,
+               FlowId flow,
+               uint16_t entropy,
+               uint8_t notifications,
+               Bytes payload_bytes) noexcept
+        : _src_node(src_node),
+          _src_port(src_port),
+          _dst_node(dst_node),
+          _dst_port(dst_port),
+          _protocol(proto),
+          _seq(seq),
+          _flow(flow),
+          _entropy(entropy),
+          _payload_size((proto == PacketProtocol::DATA) ?
+                        payload_bytes : Bytes{0}),
+          _groups(0),
+          _notifications(static_cast<uint8_t>(notifications & 0x07)) {
+        const uint8_t p = static_cast<uint8_t>(prio) & 0x03;
+        _notifications = static_cast<uint8_t>(
+                                 (_notifications & ~PRIORITY_MASK) |
+                                 (p << PRIORITY_SHIFT));
+}
+
+NodeId Packet::src_node() const noexcept {
+        return _src_node;
+}
+PortId Packet::src_port() const noexcept {
+        return _src_port;
+}
+NodeId Packet::dst_node() const noexcept {
+        return _dst_node;
+}
+PortId Packet::dst_port() const noexcept {
+        return _dst_port;
+}
+PacketProtocol Packet::protocol() const noexcept {
+        return _protocol;
+}
+PacketSeq Packet::seq() const noexcept {
+        return _seq;
+}
+FlowId Packet::flow_id() const noexcept {
+        return _flow;
+}
+uint32_t Packet::entropy() const noexcept {
+        return _entropy;
+}
+
+void Packet::set_src_node(NodeId n) noexcept {
+        _src_node = n;
+}
+void Packet::set_src_port(PortId p) noexcept {
+        _src_port = p;
+}
+void Packet::set_dst_node(NodeId n) noexcept {
+        _dst_node = n;
+}
+void Packet::set_dst_port(PortId p) noexcept {
+        _dst_port = p;
+}
+void Packet::set_seq(PacketSeq s) noexcept {
+        _seq = s;
+}
+void Packet::set_flow_id(FlowId f) noexcept {
+        _flow = f;
+}
+void Packet::set_entropy(uint32_t e) noexcept {
+        _entropy = e;
+}
+void Packet::set_protocol(PacketProtocol p) noexcept {
+        _protocol = p;
+}
+
+void Packet::set_payload_size(Bytes size) noexcept {
+        _payload_size = size;
+}
+
+void Packet::set_ecn_enabled(bool v) noexcept {
+        if (v)
+                _notifications |= ECN_ENABLED_MASK;
+        else
+                _notifications &= ~ECN_ENABLED_MASK;
+}
+void Packet::set_ecn_marked(bool v) noexcept {
+        if (v)
+                _notifications |= ECN_MARK_MASK;
+        else
+                _notifications &= ~ECN_MARK_MASK;
+}
+void Packet::set_eof(bool v) noexcept {
+        if (v)
+                _notifications |= EOF_MASK;
+        else
+                _notifications &= ~EOF_MASK;
+}
+
+bool Packet::is_ecn_enabled() const noexcept {
+        return (_notifications & ECN_ENABLED_MASK) != 0;
+}
+
+bool Packet::is_ecn() const noexcept {
+        return _notifications & ECN_MARK_MASK;
+}
+
+bool Packet::is_eof() const noexcept {
+        return (_notifications & EOF_MASK) != 0;
+}
+
+FlowPriority Packet::priority() const noexcept {
+        const uint8_t pr = static_cast<uint8_t>(
+                                   (_notifications & PRIORITY_MASK) >>
+                                   PRIORITY_SHIFT);
+
+        return static_cast<FlowPriority>(pr);
+}
+
+uint8_t Packet::priority_raw() const noexcept {
+        return static_cast<uint8_t>((_notifications & PRIORITY_MASK) >>
+                                    PRIORITY_SHIFT);
+}
+
+Bytes Packet::header_size() const noexcept {
+        return HEADER_SIZE_BYTES;
+}
+
+Bytes Packet::payload_size() const noexcept {
+        switch (_protocol) {
+        case PacketProtocol::DATA:
+                return _payload_size;
+
+        case PacketProtocol::ACK:
+        case PacketProtocol::NACK:
+                return 0;
+
+        case PacketProtocol::HEADER_TRIM:
+        case PacketProtocol::HEADER_TRIM_BACK:
+                return 0;
+        }
+
+        return 0;
+}
+
+Bytes Packet::total_size() const noexcept {
+        switch (_protocol) {
+        case PacketProtocol::DATA:
+                return HEADER_SIZE_BYTES + _payload_size;
+
+        case PacketProtocol::ACK:
+        case PacketProtocol::NACK:
+                return ACK_SIZE_BYTES;
+
+        case PacketProtocol::HEADER_TRIM:
+        case PacketProtocol::HEADER_TRIM_BACK:
+                return HEADER_SIZE_BYTES;
+        }
+
+        return HEADER_SIZE_BYTES;
+}
+
+PacketGroups Packet::groups() const noexcept {
+        return _groups;
+}
+void Packet::set_groups(PacketGroups g) noexcept {
+        _groups = g;
+}
+void Packet::add_groups(PacketGroups gmask) noexcept {
+        _groups |= gmask;
+}
+
+} // namespace dofs
diff --git a/src/network/packet.h b/src/network/packet.h
new file mode 100644
index 0000000..330ba2c
--- /dev/null
+++ b/src/network/packet.h
@@ -0,0 +1,97 @@
+#ifndef NETWORK_PACKET_H
+#define NETWORK_PACKET_H
+
+#include <cstdint>
+#include <memory>
+#include "core/types.h"
+#include "core/node.h"
+
+namespace dofs {
+
+inline const std::size_t MAX_GROUPS = 128;
+
+inline Bytes HEADER_SIZE_BYTES   = 64;
+inline Bytes ACK_SIZE_BYTES      = 64;
+inline Bytes DEFAULT_MTU_BYTES   = 4160;
+inline Bytes DEFAULT_MSS_BYTES   = 4096;
+
+class Packet {
+public:
+        Packet(NodeId src_node,
+               PortId src_port,
+               NodeId dst_node,
+               PortId dst_port,
+               PacketProtocol proto,
+               FlowPriority prio,
+               PacketSeq seq = 0,
+               FlowId flow = 0,
+               uint16_t entropy = 0,
+               uint8_t notifications = 0,
+               Bytes payload_bytes = DEFAULT_MSS_BYTES) noexcept;
+
+        NodeId         src_node() const noexcept;
+        PortId         src_port() const noexcept;
+        NodeId         dst_node() const noexcept;
+        PortId         dst_port() const noexcept;
+        PacketProtocol protocol() const noexcept;
+        PacketSeq      seq() const noexcept;
+        FlowId         flow_id() const noexcept;
+        uint32_t       entropy() const noexcept;
+
+        void set_src_node(NodeId n) noexcept;
+        void set_src_port(PortId p) noexcept;
+        void set_dst_node(NodeId n) noexcept;
+        void set_dst_port(PortId p) noexcept;
+        void set_seq(PacketSeq s) noexcept;
+        void set_flow_id(FlowId f) noexcept;
+        void set_entropy(uint32_t e) noexcept;
+        void set_protocol(PacketProtocol p) noexcept;
+
+        void set_payload_size(Bytes size) noexcept;
+
+        void set_ecn_enabled(bool v) noexcept;
+        void set_ecn_marked(bool v) noexcept;
+        // end of flow
+        void set_eof(bool v) noexcept;
+
+        bool is_ecn_enabled() const noexcept;
+        bool is_ecn() const noexcept;
+        bool is_eof() const noexcept;
+
+        FlowPriority priority() const noexcept;
+        uint8_t        priority_raw() const noexcept;
+
+        Bytes header_size() const noexcept;
+        Bytes payload_size() const noexcept;
+        Bytes total_size() const noexcept;
+
+        PacketGroups groups() const noexcept;
+        void set_groups(PacketGroups g) noexcept;
+        void add_groups(PacketGroups gmask) noexcept;
+
+private:
+        NodeId   _src_node;
+        PortId   _src_port;
+        NodeId   _dst_node;
+        PortId   _dst_port;
+
+        PacketProtocol _protocol;
+        PacketSeq      _seq;
+        FlowId         _flow;
+        uint16_t       _entropy;
+
+        Bytes        _payload_size;
+        PacketGroups _groups;
+
+        uint8_t  _notifications;
+
+        static constexpr uint8_t ECN_ENABLED_MASK = 0x01;
+        static constexpr uint8_t ECN_MARK_MASK    = 0x02;
+        static constexpr uint8_t EOF_MASK         = 0x04;
+        static constexpr uint8_t PRIORITY_MASK    = 0x18;
+        static constexpr uint8_t PRIORITY_SHIFT   = 3;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_PACKET_H
diff --git a/src/network/switch/CMakeLists.txt b/src/network/switch/CMakeLists.txt
new file mode 100644
index 0000000..b5e8e4d
--- /dev/null
+++ b/src/network/switch/CMakeLists.txt
@@ -0,0 +1,63 @@
+# --- Library: dofs_switch ---
+add_library(dofs_switch STATIC)
+
+target_include_directories(dofs_switch
+  PUBLIC
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>          # "network/...", "core/..."
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src/network>  # ** enables "switch/..." **
+    $<INSTALL_INTERFACE:include>
+)
+
+# Sources
+if(DOFS_GLOB_SOURCES)
+  file(GLOB SWITCH_CC CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.cc")
+  target_sources(dofs_switch PRIVATE ${SWITCH_CC})
+else()
+  target_sources(dofs_switch
+    PUBLIC
+      routing_alg.h
+      routing_tables.h
+      unicast_table.h
+      multicast_table.h
+      switch_buffer.h
+      shared_buffer.h
+      dedicated_buffer.h
+      ecn_engine.h
+      ecn_shared_red.h
+      ecn_dedicated_red.h
+    PRIVATE
+      routing_alg.cc
+      unicast_table.cc
+      multicast_table.cc
+      switch_buffer.cc
+      shared_buffer.cc
+      dedicated_buffer.cc
+      ecn_shared_red.cc
+      ecn_dedicated_red.cc
+  )
+endif()
+
+target_link_libraries(dofs_switch PUBLIC dofs_network dofs_core)
+add_library(dofs::switch ALIAS dofs_switch)
+
+# --- Tooling stubs for switch headers ---
+file(GLOB SWITCH_HEADERS CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.h")
+
+set(SWITCH_STUBS)
+foreach(h ${SWITCH_HEADERS})
+  get_filename_component(_base "${h}" NAME_WE)
+  set(_stub "${CMAKE_CURRENT_BINARY_DIR}/switch_${_base}.cpp")
+  file(WRITE "${_stub}" "// stub switch/${_base}.h\n#include \"switch/${_base}.h\"\n")
+  list(APPEND SWITCH_STUBS "${_stub}")
+endforeach()
+
+add_library(dofs_switch_headers_tooling STATIC ${SWITCH_STUBS})
+
+# Make sure stubs see the SAME include roots
+target_include_directories(dofs_switch_headers_tooling
+  PRIVATE
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src>
+    $<BUILD_INTERFACE:${CMAKE_SOURCE_DIR}/src/network>  # ** not src/network/switch **
+)
+
+target_link_libraries(dofs_switch_headers_tooling PRIVATE dofs_switch dofs_network dofs_core)
diff --git a/src/network/switch/dedicated_buffer.cc b/src/network/switch/dedicated_buffer.cc
new file mode 100644
index 0000000..332b1d3
--- /dev/null
+++ b/src/network/switch/dedicated_buffer.cc
@@ -0,0 +1,65 @@
+#include "network/switch/dedicated_buffer.h"
+
+namespace dofs {
+
+DedicatedBuffer::DedicatedBuffer(Simulator* const sim,
+                                 NetworkSwitch* const owner,
+                                 Bytes total_bytes,
+                                 uint16_t ports)
+        : SwitchBuffer(sim, owner, SwitchBufferType::DEDICATED, total_bytes, ports),
+          _per_port_cap(ports ? total_bytes / ports : Bytes(0)),
+          _per_port_queues(ports)
+{}
+
+std::array<std::deque<SwitchBuffer::Queued>, SwitchBuffer::PRI_COUNT>&DedicatedBuffer::queues_for(
+        PortId p) {
+        return _per_port_queues[p];
+}
+
+const std::array<std::deque<SwitchBuffer::Queued>, SwitchBuffer::PRI_COUNT>&DedicatedBuffer::queues_for(
+        PortId p) const {
+        return _per_port_queues[p];
+}
+
+bool DedicatedBuffer::on_enqueue_cap_check(PortId port, Bytes sz) {
+        if (port >= _port_cnt)
+                return false;
+
+        // Guard per-port cap
+        if (_per_port_bytes[port] + sz > _per_port_cap)
+                return false;
+
+        return true;
+}
+void DedicatedBuffer::on_enqueue_commit(PortId port, Bytes sz) {
+        _per_port_bytes[port] += sz;
+}
+void DedicatedBuffer::on_dequeue_commit(PortId port, Bytes sz) {
+        _per_port_bytes[port] -= sz;
+}
+
+bool DedicatedBuffer::enqueue_packet(const Packet& pkt, PortId egress,
+                                     FlowPriority prio) {
+        if (egress >= _port_cnt)
+                return false;
+
+        const Bytes sz = pkt.total_size();
+
+        if (!on_enqueue_cap_check(egress, sz))
+                return false;
+
+        Queued q{pkt, egress, prio, _sim->now(), sz};
+        auto& qset = queues_for(egress);
+        qset[to_idx(prio)].push_back(std::move(q));
+        on_enqueue_commit(egress, sz);
+
+        schedule_drain_if_needed(egress);
+
+        return true;
+}
+
+bool DedicatedBuffer::drain_one(PortId port) {
+        return drain_one_common(port);
+}
+
+} // namespace dofs
diff --git a/src/network/switch/dedicated_buffer.h b/src/network/switch/dedicated_buffer.h
new file mode 100644
index 0000000..a403aa5
--- /dev/null
+++ b/src/network/switch/dedicated_buffer.h
@@ -0,0 +1,38 @@
+#ifndef NETWORK_SWITCH_DEDICATED_BUFFER_H
+#define NETWORK_SWITCH_DEDICATED_BUFFER_H
+
+#include "network/switch/switch_buffer.h"
+
+namespace dofs {
+
+// DedicatedBuffer: buffer_size is split across ports (equal split by default).
+// Each port has a hard cap; no borrowing.
+class DedicatedBuffer : public SwitchBuffer {
+public:
+        DedicatedBuffer(Simulator* const sim,
+                        NetworkSwitch* const owner,
+                        Bytes total_bytes,
+                        uint16_t ports);
+
+        bool enqueue_packet(const Packet& pkt, PortId egress,
+                            FlowPriority prio) override;
+        bool drain_one(PortId port) override;
+
+protected:
+        std::array<std::deque<Queued>, PRI_COUNT> &queues_for(PortId p) override;
+        const std::array<std::deque<Queued>, PRI_COUNT> &queues_for(
+                PortId p) const override;
+
+        bool on_enqueue_cap_check(PortId port, Bytes sz) override;
+        void on_enqueue_commit(PortId port, Bytes sz) override;
+        void on_dequeue_commit(PortId port, Bytes sz) override;
+
+private:
+        // Hard cap per port (equal share of total; final remainder goes to low ports).
+        Bytes _per_port_cap;
+        std::vector<std::array<std::deque<Queued>, PRI_COUNT >> _per_port_queues;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_DEDICATED_BUFFER_H
diff --git a/src/network/switch/ecn_dedicated_red.cc b/src/network/switch/ecn_dedicated_red.cc
new file mode 100644
index 0000000..6d22d06
--- /dev/null
+++ b/src/network/switch/ecn_dedicated_red.cc
@@ -0,0 +1,50 @@
+#include "network/switch/ecn_dedicated_red.h"
+
+#include <algorithm>
+#include <cmath>
+
+namespace dofs {
+
+Packet &DedicatedREDEngine::process_packet(Packet &pkt,
+                SwitchBuffer* buf) noexcept {
+        if (!buf)
+                return pkt;
+
+        const PortId  port = pkt.dst_port();
+        const uint16_t pc  = buf->port_cnt();
+
+        if (port >= pc)
+                return pkt;
+
+        ensure_size(pc);
+
+        // instantaneous queue (bytes) for the egress port
+        const double q_inst = static_cast<double>(buf->port_buffered(port));
+        double &q_avg       = _avg_bytes[port];
+        q_avg               = ewma(q_inst, q_avg);
+
+        const double min_th = static_cast<double>(_min_th);
+        const double max_th = static_cast<double>(_max_th);
+
+        if (q_avg <= min_th) {
+                // Below min_th: no marking
+                return pkt;
+        }
+
+        if (q_avg >= max_th) {
+                // Above max_th: trim (replace drop)
+                header_trim(pkt, _back_to_sender);
+                return pkt;
+        }
+
+        // Between thresholds: mark ECN with probability p(q)
+        const double p = _p_max * (q_avg - min_th) / std::max(1.0, (max_th - min_th));
+
+        if (_rng && _rng->uniform01() < p) {
+                maybe_mark_ecn(pkt);
+        }
+
+        return pkt;
+}
+
+} // namespace dofs
diff --git a/src/network/switch/ecn_dedicated_red.h b/src/network/switch/ecn_dedicated_red.h
new file mode 100644
index 0000000..0cfb50b
--- /dev/null
+++ b/src/network/switch/ecn_dedicated_red.h
@@ -0,0 +1,54 @@
+#ifndef NETWORK_SWITCH_ECN_DEDICATED_RED_H
+#define NETWORK_SWITCH_ECN_DEDICATED_RED_H
+
+#include <vector>
+#include "network/switch/ecn_engine.h"
+#include "network/switch/switch_buffer.h"
+#include "core/rng.h"
+
+namespace dofs {
+
+// Classic per-port RED over DedicatedBuffer caps.
+// Probabilistic ECN in [min_th, max_th]; trim at > max_th.
+class DedicatedREDEngine final : public ECNEngine {
+public:
+        DedicatedREDEngine(Bytes min_th,
+                           Bytes max_th,
+                           double p_max,
+                           bool back_to_sender,
+                           Rng *const rng) noexcept
+                : _min_th(min_th),
+                  _max_th(max_th),
+                  _p_max(p_max),
+                  _back_to_sender(back_to_sender),
+                  _rng(rng) {}
+
+        virtual Packet &process_packet(Packet &pkt,
+                                       SwitchBuffer *buf) noexcept override;
+
+private:
+        // EWMA queue estimator per port (bytes)
+        void ensure_size(uint16_t port_cnt) {
+                if (_avg_bytes.size() != port_cnt)
+                        _avg_bytes.assign(port_cnt, 0.0);
+        }
+
+        double ewma(double cur, double avg) const noexcept {
+                // w_q ~ 1/8 for responsiveness without flapping
+                constexpr double W = 0.125;
+                return (1.0 - W) * avg + W * cur;
+        }
+
+private:
+        Bytes      _min_th;
+        Bytes      _max_th;
+        double     _p_max;          // in [0,1]
+        bool       _back_to_sender; // whether to trim back to sender
+        Rng *const _rng;            // not owned
+
+        std::vector<double> _avg_bytes;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_ECN_DEDICATED_RED_H
diff --git a/src/network/switch/ecn_engine.h b/src/network/switch/ecn_engine.h
new file mode 100644
index 0000000..aa0ab50
--- /dev/null
+++ b/src/network/switch/ecn_engine.h
@@ -0,0 +1,51 @@
+#ifndef NETWORK_SWITCH_ECN_ENGINE_H
+#define NETWORK_SWITCH_ECN_ENGINE_H
+
+#include <cstdint>
+#include "core/types.h"
+#include "network/packet.h"
+
+namespace dofs {
+
+class SwitchBuffer;
+
+class ECNEngine {
+public:
+        virtual ~ECNEngine() = default;
+
+        // Decide ECN mark / header-trim for a packet about to enter 'buf'.
+        // Returns the same packet reference, possibly modified in-place.
+        virtual Packet &process_packet(Packet &pkt, SwitchBuffer* buf) noexcept = 0;
+
+protected:
+        // Helpers for derived engines
+        // Now defined directly in the header
+        static inline void maybe_mark_ecn(Packet &pkt) noexcept {
+                if (pkt.is_ecn_enabled())
+                        pkt.set_ecn_marked(true);
+        }
+
+        // Trim payload to zero and set protocol accordingly.
+        // If back_to_sender, flip src/dst like an RX NIC reflection.
+        static inline void header_trim(Packet &pkt, bool back_to_sender) noexcept {
+                pkt.set_payload_size(Bytes(0));
+
+                if (back_to_sender) {
+                        const NodeId s_n = pkt.src_node();
+                        const PortId s_p = pkt.src_port();
+                        const NodeId d_n = pkt.dst_node();
+                        const PortId d_p = pkt.dst_port();
+                        pkt.set_src_node(d_n);
+                        pkt.set_src_port(d_p);
+                        pkt.set_dst_node(s_n);
+                        pkt.set_dst_port(s_p);
+                        pkt.set_protocol(PacketProtocol::HEADER_TRIM_BACK);
+                } else {
+                        pkt.set_protocol(PacketProtocol::HEADER_TRIM);
+                }
+        }
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_ECN_ENGINE_
diff --git a/src/network/switch/ecn_shared_red.cc b/src/network/switch/ecn_shared_red.cc
new file mode 100644
index 0000000..0208ba2
--- /dev/null
+++ b/src/network/switch/ecn_shared_red.cc
@@ -0,0 +1,126 @@
+#include "network/switch/ecn_shared_red.h"
+
+#include <algorithm>
+#include <numeric>
+#include <cmath>
+
+namespace dofs {
+
+SharedREDEngine::Thresholds SharedREDEngine::dynamic_thresholds(
+        const Packet &pkt,
+        const SwitchBuffer *buf,
+        double total_bytes_avg) const {
+        const Bytes cap_b = buf->buffer_size();
+        const double cap  = static_cast<double>(cap_b);
+
+        if (cap <= 0.0) {
+                return {0.0, 0.0, 0.0, false};
+        }
+
+        // Active ports heuristic
+        const auto &per = buf->ports_buffered();
+        const std::size_t active = std::max<std::size_t>(1,
+        std::count_if(per.begin(), per.end(), [](Bytes b) {
+                return b > 0;
+        }));
+
+        const double fair_share = cap / static_cast<double>(active);
+
+        // Class factors: be generous to CTRL, moderate for MICE, strict for ELE.
+        double alpha = 0.20, beta = 0.85, pmax = 0.10; // mice defaults
+        bool allow_trim = true;
+
+        if (is_ctrl(pkt)) {
+                alpha = 0.80;
+                beta = 0.98;
+                pmax = 0.02;
+                allow_trim = false; // almost no marking, never trim
+        } else if (is_ele(pkt)) {
+                alpha = 0.10;
+                beta = 0.60;
+                pmax = 0.25;
+                allow_trim = true;  // aggressive on elephants
+        }
+
+        // Dynamic bias: if total avg usage is high, tighten thresholds a bit
+        // shrink factor in [0.8, 1.0] as pool fills
+        const double util = std::clamp(total_bytes_avg / cap, 0.0, 1.0);
+        const double tighten = 1.0 - 0.2 * util;
+
+        const double min_th = alpha * fair_share * tighten;
+        const double max_th = std::max(min_th + 1.0, beta * fair_share * tighten);
+
+        return {min_th, max_th, pmax, allow_trim};
+}
+
+Packet &SharedREDEngine::process_packet(Packet &pkt,
+                                        SwitchBuffer *buf) noexcept {
+        if (!buf)
+                return pkt;
+
+        const uint16_t pc = buf->port_cnt();
+        ensure_size(pc);
+
+        // Update EWMA of total bytes in the shared pool.
+        const auto &per = buf->ports_buffered();
+        const double total_now = std::accumulate(per.begin(), per.end(), 0.0,
+        [](double acc, Bytes b) {
+                return acc + static_cast<double>(b);
+        });
+        _avg_total_bytes = ewma(total_now, _avg_total_bytes);
+
+        // Also keep per-port EWMA to stabilize active-port detection (optional).
+        for (uint16_t p = 0; p < pc; ++p) {
+                _avg_port_bytes[p] = ewma(static_cast<double>(per[p]),
+                                          _avg_port_bytes[p]);
+        }
+
+        // Use per-port instantaneous occupancy for the target port.
+        const PortId port = pkt.dst_port();
+
+        if (port >= pc)
+                return pkt;
+
+        // Derive class-aware dynamic thresholds
+        const auto thr = dynamic_thresholds(pkt, buf, _avg_total_bytes);
+
+        // Choose which queue measure to feed RED: per-port EWMA
+        const double q_avg = _avg_port_bytes[port];
+
+        if (q_avg <= thr.min_th) {
+                // Well below min_th: no mark
+                return pkt;
+        }
+
+        if (q_avg >= thr.max_th) {
+                // Above max_th: trim (unless forbidden for class)
+                if (thr.allow_trim) {
+                        header_trim(pkt, false);
+                } else {
+                        // CTRL: at worst, mark ECN if allowed
+                        maybe_mark_ecn(pkt);
+                }
+
+                return pkt;
+        }
+
+        // Between thresholds: probabilistic ECN
+        const double p = thr.p_max * (q_avg - thr.min_th) / std::max(1.0,
+                         (thr.max_th - thr.min_th));
+
+        if (_rng) {
+                if (_rng->uniform01() < p) {
+                        maybe_mark_ecn(pkt);
+                }
+        } else {
+                // Deterministic fallback: mark if over mid-point
+                const double mid = 0.5 * (thr.min_th + thr.max_th);
+
+                if (q_avg > mid)
+                        maybe_mark_ecn(pkt);
+        }
+
+        return pkt;
+}
+
+} // namespace dofs
diff --git a/src/network/switch/ecn_shared_red.h b/src/network/switch/ecn_shared_red.h
new file mode 100644
index 0000000..5b1f129
--- /dev/null
+++ b/src/network/switch/ecn_shared_red.h
@@ -0,0 +1,71 @@
+#ifndef NETWORK_SWITCH_ECN_SHARED_RED_H
+#define NETWORK_SWITCH_ECN_SHARED_RED_H
+
+#include <vector>
+#include "network/switch/ecn_engine.h"
+#include "network/switch/switch_buffer.h"
+#include "core/rng.h"
+
+namespace dofs {
+
+// Dynamic-threshold RED for a shared pool.
+// Intuition:
+//   - Let active ports = count of ports with non-zero backlog.
+//   - Per-port fair share ~ buffer_size / max(1, active).
+//   - Class tweak: CTRL lenient, ELE stricter.
+//   - min_th = alpha * share; max_th = beta * share; (alpha < beta < 1)
+//   - ECN prob grows linearly; above max_th → header trim (not drop).
+class SharedREDEngine final : public ECNEngine {
+public:
+        explicit SharedREDEngine(Rng *const rng = nullptr) noexcept
+                : _rng(rng),
+                  _avg_total_bytes(0.0),
+                  _avg_port_bytes() {}
+
+        virtual Packet &process_packet(Packet &pkt,
+                                       SwitchBuffer *buf) noexcept override;
+
+private:
+        void ensure_size(uint16_t port_cnt) {
+                if (_avg_port_bytes.size() != port_cnt)
+                        _avg_port_bytes.assign(port_cnt, 0.0);
+        }
+
+        static inline bool is_ctrl(const Packet &p) noexcept {
+                return p.priority() == FlowPriority::CTRL;
+        }
+        static inline bool is_mice(const Packet &p) noexcept {
+                return p.priority() == FlowPriority::MICE;
+        }
+        static inline bool is_ele (const Packet &p) noexcept {
+                return p.priority() == FlowPriority::ELEPHANT;
+        }
+
+        double ewma(double cur, double avg) const noexcept {
+                constexpr double W = 0.125; // same 1/8 smoothing
+                return (1.0 - W) * avg + W * cur;
+        }
+
+        struct Thresholds {
+                double min_th;
+                double max_th;
+                double p_max; // cap for ECN probability
+                bool   allow_trim; // whether we ever trim this class
+        };
+
+        // Compute dynamic thresholds from shared state and class.
+        Thresholds dynamic_thresholds(const Packet &pkt,
+                                      const SwitchBuffer *buf,
+                                      double total_bytes_avg) const;
+
+private:
+        Rng *const _rng;
+
+        // EWMA of total bytes (shared pool) and per-port bytes
+        double _avg_total_bytes;
+        std::vector<double> _avg_port_bytes;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_ECN_SHARED_RED_H
diff --git a/src/network/switch/multicast_table.cc b/src/network/switch/multicast_table.cc
new file mode 100644
index 0000000..60c8ce7
--- /dev/null
+++ b/src/network/switch/multicast_table.cc
@@ -0,0 +1,208 @@
+#include "network/switch/multicast_table.h"
+
+#include <set>
+#include <iterator>
+
+namespace dofs {
+
+MulticastTable::MulticastTable()
+        : _groups(MAX_GROUPS) {}
+
+std::size_t MulticastTable::group_count() const noexcept {
+        return _groups.size();
+}
+
+bool MulticastTable::add_tree(std::size_t group_id, uint16_t tree_id) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it != trees.end())
+                return false;
+
+        McTree t;
+        t.tree_id = tree_id;
+        trees.insert(std::lower_bound(trees.begin(), trees.end(), tree_id,
+        [](const McTree & a, uint16_t id) {
+                return a.tree_id < id;
+        }),
+        t);
+        return true;
+}
+
+bool MulticastTable::delete_tree(std::size_t group_id, uint16_t tree_id) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it == trees.end())
+                return false;
+
+        trees.erase(it);
+        return true;
+}
+
+bool MulticastTable::add_child_port(std::size_t group_id, uint16_t tree_id,
+                                    PortId out_port) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it == trees.end())
+                return false;
+
+        return insert_sorted_unique(it->child_ports, out_port);
+}
+
+bool MulticastTable::delete_child_port(std::size_t group_id, uint16_t tree_id,
+                                       PortId out_port) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it == trees.end())
+                return false;
+
+        return erase_sorted(it->child_ports, out_port);
+}
+
+bool MulticastTable::set_parent(std::size_t group_id, uint16_t tree_id,
+                                PortId parent) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it == trees.end())
+                return false;
+
+        it->parent_port = parent;
+        return true;
+}
+
+bool MulticastTable::set_weight(std::size_t group_id, uint16_t tree_id,
+                                uint8_t w) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it == trees.end())
+                return false;
+
+        it->weight = w;
+        return true;
+}
+
+bool MulticastTable::set_epoch(std::size_t group_id, uint16_t tree_id,
+                               uint8_t epoch) {
+        if (group_id >= _groups.size())
+                return false;
+
+        auto& trees = _groups[group_id];
+        auto it = std::find_if(trees.begin(), trees.end(),
+        [&](const McTree & t) {
+                return t.tree_id == tree_id;
+        });
+
+        if (it == trees.end())
+                return false;
+
+        it->epoch = epoch;
+        return true;
+}
+
+const std::vector<McTree> *MulticastTable::trees_of(std::size_t group_id)
+const {
+        if (group_id >= _groups.size())
+                return nullptr;
+
+        return &_groups[group_id];
+}
+
+// Legacy: union of tree_id=0 ports for all set groups
+std::vector<PortId> MulticastTable::get_port_list(PacketGroups groups) const {
+        std::vector<PortId> result;
+
+        for (std::size_t gid = 0; gid < _groups.size(); ++gid) {
+                const bool bit_set = (groups >> gid) & static_cast<PacketGroups>(1);
+
+                if (!bit_set)
+                        continue;
+
+                const auto& trees = _groups[gid];
+                auto it = std::find_if(trees.begin(), trees.end(),
+                [](const McTree & t) {
+                        return t.tree_id == 0;
+                });
+
+                if (it == trees.end())
+                        continue;
+
+                if (it->child_ports.empty())
+                        continue;
+
+                if (result.empty()) {
+                        result = it->child_ports;
+                } else {
+                        std::vector<PortId> merged;
+                        merged.reserve(result.size() + it->child_ports.size());
+                        std::set_union(result.begin(), result.end(),
+                                       it->child_ports.begin(), it->child_ports.end(),
+                                       std::back_inserter(merged));
+                        result.swap(merged);
+                }
+        }
+
+        return result;
+}
+
+bool MulticastTable::insert_sorted_unique(std::vector<PortId> &vec, PortId v) {
+        auto it = std::lower_bound(vec.begin(), vec.end(), v);
+
+        if (it != vec.end() && *it == v)
+                return false;
+
+        vec.insert(it, v);
+        return true;
+}
+
+bool MulticastTable::erase_sorted(std::vector<PortId> &vec, PortId v) {
+        auto it = std::lower_bound(vec.begin(), vec.end(), v);
+
+        if (it == vec.end() || *it != v)
+                return false;
+
+        vec.erase(it);
+        return true;
+}
+
+} // namespace dofs
diff --git a/src/network/switch/multicast_table.h b/src/network/switch/multicast_table.h
new file mode 100644
index 0000000..4729942
--- /dev/null
+++ b/src/network/switch/multicast_table.h
@@ -0,0 +1,52 @@
+#ifndef NETWORK_SWITCH_MULTICAST_TABLE_H
+#define NETWORK_SWITCH_MULTICAST_TABLE_H
+
+#include <cstdint>
+#include <vector>
+#include <optional>
+#include <algorithm>
+
+#include "core/types.h"
+#include "network/packet.h"
+
+namespace dofs {
+
+struct McTree {
+        std::vector<PortId> child_ports;
+        std::optional<PortId> parent_port;
+        uint8_t weight{1};
+        uint8_t tier{0};
+        uint16_t tree_id{0};
+        uint8_t epoch{0};
+};
+
+class MulticastTable {
+public:
+        MulticastTable();
+
+        bool add_tree(std::size_t group_id, uint16_t tree_id);
+        bool delete_tree(std::size_t group_id, uint16_t tree_id);
+
+        bool add_child_port(std::size_t group_id, uint16_t tree_id, PortId out_port);
+        bool delete_child_port(std::size_t group_id, uint16_t tree_id, PortId out_port);
+
+        bool set_parent(std::size_t group_id, uint16_t tree_id, PortId parent);
+        bool set_weight(std::size_t group_id, uint16_t tree_id, uint8_t w);
+        bool set_epoch(std::size_t group_id, uint16_t tree_id, uint8_t epoch);
+
+        const std::vector<McTree> *trees_of(std::size_t group_id) const;
+        std::size_t group_count() const noexcept;
+
+        // Legacy: return ports from tree_id=0 for each group bit
+        std::vector<PortId> get_port_list(PacketGroups groups) const;
+
+private:
+        static bool insert_sorted_unique(std::vector<PortId> &vec, PortId v);
+        static bool erase_sorted(std::vector<PortId> &vec, PortId v);
+
+        std::vector<std::vector<McTree>> _groups;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_MULTICAST_TABLE_H
diff --git a/src/network/switch/routing_alg.cc b/src/network/switch/routing_alg.cc
new file mode 100644
index 0000000..4207bc4
--- /dev/null
+++ b/src/network/switch/routing_alg.cc
@@ -0,0 +1,36 @@
+#include "network/switch/routing_alg.h"
+#include <cstdint>
+
+namespace dofs {
+
+static inline uint32_t mix32(uint32_t x) noexcept {
+        x ^= x >> 16;
+        x *= 0x7feb352dU;
+        x ^= x >> 15;
+        x *= 0x846ca68bU;
+        x ^= x >> 16;
+        return x;
+}
+
+PortId hash_ecmp(NodeId src_node, PortId src_port,
+                 NodeId dst_node, PortId dst_port,
+                 uint32_t differentiator,
+                 uint16_t port_count) noexcept {
+        if (port_count == 0) {
+                log_error("ERROR", "hash_ecmp ignored: port_count == 0");
+                return static_cast<PortId>(0);
+        }
+
+        const uint32_t sip = ipv4(src_node, src_port);
+        const uint32_t dip = ipv4(dst_node, dst_port);
+
+        uint32_t key = sip;
+        key ^= mix32(dip + 0x9e3779b9U);
+        key ^= mix32(differentiator + 0x85ebca6bU);
+
+        const uint32_t h = mix32(key);
+        const uint16_t idx = static_cast<uint16_t>(h % port_count);
+        return static_cast<PortId>(idx);
+}
+
+} // namespace dofs
diff --git a/src/network/switch/routing_alg.h b/src/network/switch/routing_alg.h
new file mode 100644
index 0000000..c757edc
--- /dev/null
+++ b/src/network/switch/routing_alg.h
@@ -0,0 +1,26 @@
+#ifndef NETWORK_SWITCH_ROUTING_ALG_H
+#define NETWORK_SWITCH_ROUTING_ALG_H
+
+#include <cstdint>
+#include "core/types.h"
+#include "core/error.h"
+
+namespace dofs {
+
+PortId hash_ecmp(NodeId src_node, PortId src_port,
+                 NodeId dst_node, PortId dst_port,
+                 uint32_t differentiator,
+                 uint16_t port_count) noexcept;
+
+inline PortId hash_ecmp(NodeId src_node,
+                        NodeId dst_node,
+                        uint32_t differentiator,
+                        uint16_t port_count) noexcept {
+        return hash_ecmp(src_node, static_cast<PortId>(0),
+                         dst_node, static_cast<PortId>(0),
+                         differentiator, port_count);
+}
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_ROUTING_ALG_H
diff --git a/src/network/switch/routing_tables.h b/src/network/switch/routing_tables.h
new file mode 100644
index 0000000..522d4b0
--- /dev/null
+++ b/src/network/switch/routing_tables.h
@@ -0,0 +1,16 @@
+#ifndef NETWORK_SWITCH_ROUTING_TABLES_H
+#define NETWORK_SWITCH_ROUTING_TABLES_H
+
+#include "network/switch/unicast_table.h"
+#include "network/switch/multicast_table.h"
+
+namespace dofs {
+
+struct RoutingTables {
+        UnicastTable   unicast;
+        MulticastTable multicast;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_ROUTING_TABLES_H
diff --git a/src/network/switch/shared_buffer.cc b/src/network/switch/shared_buffer.cc
new file mode 100644
index 0000000..fa1153d
--- /dev/null
+++ b/src/network/switch/shared_buffer.cc
@@ -0,0 +1,63 @@
+#include "network/switch/shared_buffer.h"
+
+namespace dofs {
+
+SharedBuffer::SharedBuffer(Simulator* const sim,
+                           NetworkSwitch* const owner,
+                           Bytes total_bytes,
+                           uint16_t ports)
+        : SwitchBuffer(sim, owner, SwitchBufferType::SHARED, total_bytes, ports),
+          _per_port_queues(ports),
+          _used_total(0)
+{}
+
+std::array<std::deque<SwitchBuffer::Queued>, SwitchBuffer::PRI_COUNT>&SharedBuffer::queues_for(
+        PortId p) {
+        return _per_port_queues[p];
+}
+
+const std::array<std::deque<SwitchBuffer::Queued>, SwitchBuffer::PRI_COUNT>&SharedBuffer::queues_for(
+        PortId p) const {
+        return _per_port_queues[p];
+}
+
+bool SharedBuffer::on_enqueue_cap_check(PortId, Bytes sz) {
+        if (_used_total + sz > _buffer_bytes)
+                return false;
+
+        return true;
+}
+void SharedBuffer::on_enqueue_commit(PortId port, Bytes sz) {
+        _used_total += sz;
+        _per_port_bytes[port] += sz;
+}
+void SharedBuffer::on_dequeue_commit(PortId port, Bytes sz) {
+        _used_total -= sz;
+        _per_port_bytes[port] -= sz;
+}
+
+bool SharedBuffer::enqueue_packet(const Packet& pkt, PortId egress,
+                                  FlowPriority prio) {
+        if (egress >= _port_cnt)
+                return false;
+
+        const Bytes sz = pkt.total_size();
+
+        if (!on_enqueue_cap_check(egress, sz))
+                return false;
+
+        Queued q{pkt, egress, prio, _sim->now(), sz};
+        auto& qset = queues_for(egress);
+        qset[to_idx(prio)].push_back(std::move(q));
+        on_enqueue_commit(egress, sz);
+
+        schedule_drain_if_needed(egress);
+
+        return true;
+}
+
+bool SharedBuffer::drain_one(PortId port) {
+        return drain_one_common(port);
+}
+
+} // namespace dofs
diff --git a/src/network/switch/shared_buffer.h b/src/network/switch/shared_buffer.h
new file mode 100644
index 0000000..c69899d
--- /dev/null
+++ b/src/network/switch/shared_buffer.h
@@ -0,0 +1,39 @@
+#ifndef NETWORK_SWITCH_SHARED_BUFFER_H
+#define NETWORK_SWITCH_SHARED_BUFFER_H
+
+#include "network/switch/switch_buffer.h"
+
+namespace dofs {
+
+// SharedBuffer: a single global memory pool shared by all ports.
+// Only guard is the total pool (_buffer_bytes). Per-port usage tracked for stats.
+class SharedBuffer : public SwitchBuffer {
+public:
+        SharedBuffer(Simulator* const sim,
+                     NetworkSwitch* const owner,
+                     Bytes total_bytes,
+                     uint16_t ports);
+
+        virtual bool enqueue_packet(const Packet& pkt, PortId egress,
+                                    FlowPriority prio) override;
+        virtual bool drain_one(PortId port) override;
+
+protected:
+        virtual std::array<std::deque<Queued>, PRI_COUNT> &queues_for(
+                PortId p) override;
+        virtual const std::array<std::deque<Queued>, PRI_COUNT> &queues_for(
+                PortId p) const override;
+
+        virtual bool on_enqueue_cap_check(PortId port, Bytes sz) override;
+        virtual void on_enqueue_commit(PortId port, Bytes sz) override;
+        virtual void on_dequeue_commit(PortId port, Bytes sz) override;
+
+private:
+        // Per-port, per-priority queues
+        std::vector<std::array<std::deque<Queued>, PRI_COUNT >> _per_port_queues;
+        Bytes _used_total;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_SHARED_BUFFER_H
diff --git a/src/network/switch/switch_buffer.cc b/src/network/switch/switch_buffer.cc
new file mode 100644
index 0000000..5124287
--- /dev/null
+++ b/src/network/switch/switch_buffer.cc
@@ -0,0 +1,244 @@
+#include "network/switch/switch_buffer.h"
+
+#include <cmath>
+#include <algorithm>
+#include <numeric>
+#include <functional>
+
+#include "network/network_switch.h"
+
+namespace dofs {
+
+SwitchBuffer::SwitchBuffer(Simulator* const sim,
+                           NetworkSwitch* const owner,
+                           SwitchBufferType t,
+                           Bytes total_bytes,
+                           uint16_t ports)
+        : _sim(sim),
+          _owner(owner),
+          _type(t),
+          _buffer_bytes(total_bytes),
+          _port_cnt(ports),
+          _egress_links(ports, nullptr),
+          _drain_scheduled(ports, false),
+          _per_port_bytes(ports, Bytes(0)),
+          _sched(ports),
+          _share_ctrl(30), _share_mice(50), _share_ele(20) // sane defaults
+{}
+
+Bytes SwitchBuffer::buffer_size() const noexcept {
+        return _buffer_bytes;
+}
+SwitchBufferType SwitchBuffer::type() const noexcept {
+        return _type;
+}
+uint16_t SwitchBuffer::port_cnt() const noexcept {
+        return _port_cnt;
+}
+Bytes SwitchBuffer::port_buffered(PortId p) const noexcept {
+        if (p >= _per_port_bytes.size())
+                return Bytes(0);
+
+        return _per_port_bytes[p];
+}
+const std::vector<Bytes> &SwitchBuffer::ports_buffered() const noexcept {
+        return _per_port_bytes;
+}
+uint8_t SwitchBuffer::share_ctrl() const noexcept {
+        return _share_ctrl;
+}
+uint8_t SwitchBuffer::share_mice() const noexcept {
+        return _share_mice;
+}
+uint8_t SwitchBuffer::share_elephant() const noexcept {
+        return _share_ele;
+}
+
+const Simulator *SwitchBuffer::simulator() const noexcept {
+        return _sim;
+}
+const NetworkSwitch *SwitchBuffer::owner() const noexcept {
+        return _owner;
+}
+
+void SwitchBuffer::set_egress_links(const std::vector<Link*> &links) {
+        if (links.size() != _port_cnt) {
+                log_error("ERROR", "SwitchBuffer::set_egress_links size mismatch");
+                return;
+        }
+
+        _egress_links = links;
+}
+
+static void normalize_triple(uint8_t &a, uint8_t &b, uint8_t &c) {
+        int sum = int(a) + int(b) + int(c);
+
+        if (sum == 100 || sum == 0) {
+                if (sum == 0) {
+                        a = 34;
+                        b = 33;
+                        c = 33;
+                }
+
+                return;
+        }
+
+        // proportional normalization
+        double k = 100.0 / double(sum);
+        int na = int(std::round(a * k));
+        int nb = int(std::round(b * k));
+        int nc = 100 - na - nb;
+
+        if (nc < 0)
+                nc = 0;
+
+        a = uint8_t(std::clamp(na, 0, 100));
+        b = uint8_t(std::clamp(nb, 0, 100));
+        c = uint8_t(std::clamp(nc, 0, 100));
+}
+
+void SwitchBuffer::set_share_ctrl(uint8_t pct) noexcept {
+        _share_ctrl = pct;
+        normalize_triple(_share_ctrl, _share_mice, _share_ele);
+}
+void SwitchBuffer::set_share_mice(uint8_t pct) noexcept {
+        _share_mice = pct;
+        normalize_triple(_share_ctrl, _share_mice, _share_ele);
+}
+void SwitchBuffer::set_share_elephant(uint8_t pct) noexcept {
+        _share_ele = pct;
+        normalize_triple(_share_ctrl, _share_mice, _share_ele);
+}
+
+Bytes SwitchBuffer::queued_bytes_total() const noexcept {
+        Bytes sum(0);
+
+        for (auto b : _per_port_bytes)
+                sum += b;
+
+        return sum;
+}
+Bytes SwitchBuffer::queued_bytes_port(PortId p) const noexcept {
+        return port_buffered(p);
+}
+
+std::optional<Time> SwitchBuffer::try_reserve_and_send(PortId port, Queued& q) {
+        if (port >= _egress_links.size() || _egress_links[port] == nullptr) {
+                log_error("ERROR", "SwitchBuffer: egress link missing on port");
+                return std::nullopt;
+        }
+
+        Link* link = _egress_links[port];
+
+        auto r = link->reserve(q.size_bytes, _owner->id());
+
+        if (!r.has_value()) {
+                return std::nullopt; // link down or invalid sender
+        }
+
+        const Time now = _sim->now();
+        const Time to_finish = (r->finish > now) ? r->finish.unsafe_sub(now) : Time(0);
+        const Time deliver_after = to_finish + link->propagation_latency();
+
+        link->schedule_delivery_after(q.pkt, _owner->id(), deliver_after);
+
+        return r->finish;
+}
+
+void SwitchBuffer::schedule_drain_if_needed(PortId port) {
+        if (port >= _port_cnt)
+                return;
+
+        if (_drain_scheduled[port])
+                return;
+
+        if (queued_bytes_port(port) == Bytes(0))
+                return;
+
+        Link* link = _egress_links[port];
+
+        if (!link)
+                return;
+
+        const Time now = _sim->now();
+        const Time na = link->next_available(_owner->id());
+        const Time delay = (na > now) ? na.unsafe_sub(now) : Time(0);
+
+        _drain_scheduled[port] = true;
+        _sim->schedule_after(delay,
+                             std::mem_fn(&SwitchBuffer::drain_once), this, port);
+}
+
+void SwitchBuffer::drain_once(PortId port) {
+        _drain_scheduled[port] = false;
+
+        // Try to send one packet. If we made progress, drain_one_common()
+        // already scheduled the next drain at the right time.
+        if (drain_one_common(port)) {
+                return;
+        }
+
+        // No progress: either queues empty or link not reservable now.
+        // If queues are still non-empty, retry later at next-available.
+        auto& qs = queues_for(port);
+        bool any = !qs[PRI_CTRL].empty() || !qs[PRI_MICE].empty() ||
+                   !qs[PRI_ELE].empty();
+
+        if (any) {
+                schedule_drain_if_needed(port);
+        }
+}
+
+bool SwitchBuffer::drain_one_common(PortId port) {
+        if (port >= _port_cnt)
+                return false;
+
+        auto& queues = queues_for(port);
+        auto& sched  = _sched[port];
+
+        auto head_size = [&](int pri) -> Bytes {
+                if (queues[pri].empty())
+                        return Bytes(0);
+
+                return queues[pri].front().size_bytes;
+        };
+
+        sched.deficit_bytes[PRI_CTRL] += int64_t(_share_ctrl) * QUANTUM_UNIT;
+        sched.deficit_bytes[PRI_MICE] += int64_t(_share_mice) * QUANTUM_UNIT;
+        sched.deficit_bytes[PRI_ELE]  += int64_t(_share_ele)  * QUANTUM_UNIT;
+
+        for (int k = 0; k < PRI_COUNT; ++k) {
+                int pri = (sched.next_pick + k) % PRI_COUNT;
+                Bytes sz = head_size(pri);
+
+                if (sz > Bytes(0) && sched.deficit_bytes[pri] >= int64_t(sz)) {
+                        auto& q = queues[pri].front();
+                        auto finish_opt = try_reserve_and_send(port, q);
+
+                        if (finish_opt.has_value()) {
+                                queues[pri].pop_front();
+                                sched.deficit_bytes[pri] -= int64_t(sz);
+                                on_dequeue_commit(port, sz);
+                                sched.next_pick = (pri + 1) % PRI_COUNT;
+
+                                Link* link = _egress_links[port];
+                                const Time na = link->next_available(_owner->id());
+                                const Time now = _sim->now();
+                                const Time delay = (na > now) ? na.unsafe_sub(now) : Time(0);
+                                _drain_scheduled[port] = true;
+                                _sim->schedule_after(delay,
+                                                     std::mem_fn(&SwitchBuffer::drain_once), this, port);
+
+                                return true;
+                        } else {
+                                // Could not reserve now (link down). Stop early
+                                return false;
+                        }
+                }
+        }
+
+        sched.next_pick = (sched.next_pick + 1) % PRI_COUNT;
+        return false;
+}
+
+} // namespace dofs
diff --git a/src/network/switch/switch_buffer.h b/src/network/switch/switch_buffer.h
new file mode 100644
index 0000000..0659342
--- /dev/null
+++ b/src/network/switch/switch_buffer.h
@@ -0,0 +1,162 @@
+#ifndef NETWORK_SWITCH_SWITCH_BUFFER_H
+#define NETWORK_SWITCH_SWITCH_BUFFER_H
+
+#include <cstdint>
+#include <vector>
+#include <deque>
+#include <optional>
+
+#include "core/time.h"
+#include "core/types.h"
+#include "core/simulator.h"
+#include "core/error.h"
+#include "network/packet.h"
+#include "network/link.h"
+
+namespace dofs {
+
+class NetworkSwitch;
+
+// Base class for switch egress buffering with weighted priorities.
+// Implementations: SharedBuffer, DedicatedBuffer.
+class SwitchBuffer {
+public:
+        struct Queued {
+                Packet pkt;
+                PortId egress;
+                FlowPriority prio;
+                Time enq_time;
+                Bytes size_bytes;
+        };
+
+        virtual ~SwitchBuffer() = default;
+
+        // --- API: enqueue & drain ---
+        // NOTE: ECN/drop decisions are expected to be done *before* enqueue;
+        // we still guard and return false on capacity violations.
+        virtual bool enqueue_packet(const Packet& pkt,
+                                    PortId egress,
+                                    FlowPriority prio) = 0;
+
+        // Attempt to send exactly one packet on 'port' using weighted priorities.
+        // Returns true if a packet was reserved and scheduled.
+        virtual bool drain_one(PortId port) = 0;
+
+        // --- Accessors (implemented in .cc) ---
+        Bytes buffer_size() const noexcept;
+        SwitchBufferType type() const noexcept;
+        uint16_t port_cnt() const noexcept;
+        Bytes port_buffered(PortId p) const noexcept;
+        const std::vector<Bytes> &ports_buffered() const noexcept;
+
+        // Priority shares: percentages [0..100]. Sum must be 100.
+        uint8_t share_ctrl() const noexcept;
+        uint8_t share_mice() const noexcept;
+        uint8_t share_elephant() const noexcept;
+
+        // --- Mutators for priority shares (implemented in .cc) ---
+        // If the 3 shares don't sum to 100, we normalize proportionally.
+        void set_share_ctrl(uint8_t pct) noexcept;
+        void set_share_mice(uint8_t pct) noexcept;
+        void set_share_elephant(uint8_t pct) noexcept;
+
+        // Owner & env
+        const Simulator *simulator() const noexcept;
+        const NetworkSwitch *owner() const noexcept;
+
+        // Link wiring (egress side). Size must equal port_cnt().
+        void set_egress_links(const std::vector<Link*> &links);
+
+protected:
+        SwitchBuffer(Simulator* const sim,
+                     NetworkSwitch* const owner,
+                     SwitchBufferType t,
+                     Bytes total_bytes,
+                     uint16_t ports);
+
+        // Per-priority index and helpers
+        static constexpr int PRI_COUNT = 3;
+        static constexpr int PRI_CTRL = 0;
+        static constexpr int PRI_MICE = 1;
+        static constexpr int PRI_ELE  = 2;
+
+        static int to_idx(FlowPriority p) noexcept {
+                switch (p) {
+                case FlowPriority::CTRL:
+                        return PRI_CTRL;
+
+                case FlowPriority::MICE:
+                        return PRI_MICE;
+
+                case FlowPriority::ELEPHANT:
+                        return PRI_ELE;
+                }
+
+                return PRI_ELE; // defensive
+        }
+
+        // Weighted Deficit Round Robin across priorities (per port):
+        // Maintain byte-deficits per prio; each drain attempt adds quantum
+        // proportional to the configured share. Pick first prio whose head fits.
+        struct PerPortSched {
+                std::array<int64_t, PRI_COUNT> deficit_bytes {0, 0, 0};
+                int next_pick = 0; // rotating pointer among priorities
+        };
+
+        // Quantum baseline (bytes) per "share unit". Calibrated to MSS-ish.
+        // Effective quantum = share_pct * QUANTUM_UNIT.
+        static constexpr int64_t QUANTUM_UNIT = 128; // bytes per % share
+
+        // For derived classes
+        bool drain_one_common(PortId port);
+
+        // Per-port self-scheduling
+        void schedule_drain_if_needed(PortId port);
+        void drain_once(PortId port); // event-loop callback
+
+        // Derived classes must expose queues and capacity guards:
+
+        // Access the 3 per-priority queues for a port.
+        virtual std::array<std::deque<Queued>, PRI_COUNT> &queues_for(PortId p) = 0;
+        virtual const std::array<std::deque<Queued>, PRI_COUNT> &queues_for(
+                PortId p) const = 0;
+
+        // Capacity checks and accounting; return false if not enough space.
+        virtual bool on_enqueue_cap_check(PortId port, Bytes sz) = 0;
+        virtual void on_enqueue_commit(PortId port, Bytes sz) = 0;
+        virtual void on_dequeue_commit(PortId port, Bytes sz) = 0;
+
+        // Attempt reservation+schedule on the egress link for (port, packet).
+        // Returns the reservation finish time if successful (for scheduling the next drain).
+        std::optional<Time> try_reserve_and_send(PortId port, Queued &q);
+
+        // Helpers
+        Bytes queued_bytes_total() const noexcept;
+        Bytes queued_bytes_port(PortId p) const noexcept;
+
+protected:
+        Simulator *const _sim;
+        NetworkSwitch *const _owner;
+
+        SwitchBufferType _type;
+        Bytes _buffer_bytes;
+        uint16_t _port_cnt;
+
+        // Egress link per port (owned by switch; not by buffer)
+        std::vector<Link *> _egress_links;
+
+        std::vector<bool> _drain_scheduled;
+
+        // Per-port accounting and sched state
+        std::vector<Bytes> _per_port_bytes;
+        std::vector<PerPortSched> _sched; // per-port deficits
+
+        // Priority shares (percentages). Maintained to sum to 100.
+        uint8_t _share_ctrl;
+        uint8_t _share_mice;
+        uint8_t _share_ele;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_SWITCH_BUFFER_H
diff --git a/src/network/switch/unicast_table.cc b/src/network/switch/unicast_table.cc
new file mode 100644
index 0000000..21cebca
--- /dev/null
+++ b/src/network/switch/unicast_table.cc
@@ -0,0 +1,54 @@
+#include "network/switch/unicast_table.h"
+
+namespace dofs {
+
+std::vector<PortId> UnicastTable::get_port_list(NodeId dst_node,
+                PortId dst_port) const {
+        const IPv4Addr k = ipv4(dst_node, dst_port);
+        auto it = _table.find(k);
+
+        if (it == _table.end())
+                return {};
+
+        return it->second;
+}
+
+bool UnicastTable::add_entry(NodeId dst_node, PortId dst_port,
+                             PortId out_port) {
+        const IPv4Addr k = ipv4(dst_node, dst_port);
+        auto& vec = _table[k];
+        return insert_sorted_unique(vec, out_port);
+}
+
+bool UnicastTable::delete_entry(NodeId dst_node, PortId dst_port,
+                                PortId out_port) {
+        const IPv4Addr k = ipv4(dst_node, dst_port);
+        auto it = _table.find(k);
+
+        if (it == _table.end())
+                return false;
+
+        return erase_sorted(it->second, out_port);
+}
+
+bool UnicastTable::insert_sorted_unique(std::vector<PortId> &vec, PortId v) {
+        auto it = std::lower_bound(vec.begin(), vec.end(), v);
+
+        if (it != vec.end() && *it == v)
+                return false;
+
+        vec.insert(it, v);
+        return true;
+}
+
+bool UnicastTable::erase_sorted(std::vector<PortId> &vec, PortId v) {
+        auto it = std::lower_bound(vec.begin(), vec.end(), v);
+
+        if (it == vec.end() || *it != v)
+                return false;
+
+        vec.erase(it);
+        return true;
+}
+
+} // namespace dofs
diff --git a/src/network/switch/unicast_table.h b/src/network/switch/unicast_table.h
new file mode 100644
index 0000000..e139d5f
--- /dev/null
+++ b/src/network/switch/unicast_table.h
@@ -0,0 +1,36 @@
+#ifndef NETWORK_SWITCH_UNICAST_TABLE_H
+#define NETWORK_SWITCH_UNICAST_TABLE_H
+
+#include <cstdint>
+#include <unordered_map>
+#include <vector>
+#include <algorithm>
+
+#include "core/types.h"
+
+namespace dofs {
+
+class UnicastTable {
+public:
+        UnicastTable() = default;
+
+        std::vector<PortId> get_port_list(NodeId dst_node, PortId dst_port) const;
+
+        bool add_entry(NodeId dst_node, PortId dst_port, PortId out_port);
+
+        bool delete_entry(NodeId dst_node, PortId dst_port, PortId out_port);
+
+private:
+        static constexpr uint32_t kPortShiftBits = 8;
+
+        static bool insert_sorted_unique(std::vector<PortId> &vec, PortId v);
+
+        static bool erase_sorted(std::vector<PortId> &vec, PortId v);
+
+private:
+        std::unordered_map<IPv4Addr, std::vector<PortId>> _table;
+};
+
+} // namespace dofs
+
+#endif // NETWORK_SWITCH_UNICAST_TABLE_H
diff --git a/third_party/yaml-cpp b/third_party/yaml-cpp
new file mode 160000
index 0000000..2f86d13
--- /dev/null
+++ b/third_party/yaml-cpp
@@ -0,0 +1 @@
+Subproject commit 2f86d13775d119edbb69af52e5f566fd65c6953b