Low-Latency Order Matching Engine 2025

C++, Multithreading, WebSockets, CMake

Asynchronous limit order book (LOB) matching engine with lock-free ingress/publish queues, lock-free memory pool allocation for order nodes, and a WebSocket API for real-time Level 2 market data.

Highlights

• Single-writer matching core for deterministic FIFO price-time priority.
• Lock-free MPSC command queue for concurrent producer threads.
• Lock-free object pool for order-node allocation/reuse to reduce heap churn and latency spikes.
• Dedicated matcher + publisher threads for asynchronous market data fan-out.
• Native WebSocket server with RFC6455 handshake, text frames, ping/pong, and broadcast.
• Bench harness to measure p50/p95/p99 execution latencies.

Project Layout

include/lom/
  lockfree_mpsc_queue.hpp
  matching_engine.hpp
  order_book.hpp
  order_pool.hpp
  types.hpp
  websocket_server.hpp
src/
  core/order_book.cpp
  engine/matching_engine.cpp
  net/websocket_server.cpp
  util/base64.cpp
  util/sha1.cpp
  main.cpp
  benchmark.cpp
tests/
  order_book_tests.cpp
examples/
  ws_client.html

Build

cmake -S . -B build
cmake --build build --config Release

Windows (MSVC) quickstart

cmake -S . -B build -G "Visual Studio 17 2022" -A x64
cmake --build build --config Release
ctest --test-dir build -C Release --output-on-failure

Linux / WSL quickstart

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j
ctest --test-dir build --output-on-failure

Benchmarks

Performance on 300,000 order commands:

Active resting orders: 64,546
Latency p50:  0.00 µs (sub-microsecond)
Latency p95:  0.00 µs (sub-microsecond)
Latency p99:  0.00 µs (sub-microsecond)
Latency max:  3,054 µs (~3 ms)

Run the benchmark:

./build/lom_benchmark

Run Engine

./build/lom_engine

WebSocket endpoint:

ws://127.0.0.1:9002

API (WebSocket JSON)

New order

{
  "type": "new_order",
  "request_id": "client-req-001",
  "order_id": 1001,
  "user_id": 42,
  "side": "buy",
  "price_ticks": 10050,
  "quantity": 25,
  "client_ts_ns": 0
}

Cancel order

{
  "type": "cancel_order",
  "request_id": "client-req-002",
  "order_id": 1001,
  "user_id": 42,
  "client_ts_ns": 0
}

Outbound events

• ack for accepted/rejected command ingestion. Includes:
  • schema_version
  • request_id (if provided by client)
  • reason_code + reason when rejected
• l2 payload after accepted book updates:
  • schema_version
  • sequence
  • ts_ns
  • bids: [price_ticks, total_qty, order_count]
  • asks: [price_ticks, total_qty, order_count]
  • trades: array of matched fills

Example ack event

{
  "schema_version": 1,
  "type": "ack",
  "event": "new_order",
  "status": "accepted",
  "request_id": "client-req-001",
  "ts_ns": 1720000000000
}

Example l2 event

{
  "schema_version": 1,
  "type": "l2",
  "sequence": 1024,
  "ts_ns": 1720000000000,
  "bids": [[10050, 45, 3]],
  "asks": [[10060, 20, 1]],
  "trades": [
    {
      "taker_order_id": 2002,
      "maker_order_id": 1999,
      "buy_user_id": 77,
      "sell_user_id": 51,
      "price_ticks": 10060,
      "quantity": 5,
      "match_ts_ns": 1720000000000
    }
  ]
}

Risk Guardrails

Gateway validation rejects orders before enqueue if:

• order_id == 0
• user_id == 0
• quantity == 0 or quantity > 250000
• price_ticks outside [1, 2000000]

Threading Model

• Producer threads enqueue commands through a lock-free MPSC queue.
• Matcher thread is the single writer to the limit order book for deterministic matching.
• Publisher thread fans out acknowledgements and l2 snapshots to websocket clients.
• Memory pool recycles order nodes to minimize heap-driven latency spikes.

Runtime Counters

At shutdown, the engine prints:

• submit accepted/rejected totals
• commands processed/accepted/rejected
• total trades and matched quantity
• total market data events emitted

Project Planning

• Roadmap: docs/ROADMAP.md
• Issue backlog: docs/ISSUES_BACKLOG.md

Test

ctest --test-dir build --output-on-failure

Benchmark

./build/lom_benchmark

Browser Smoke Test

Open examples/ws_client.html, connect to ws://127.0.0.1:9002, and submit orders.