HFT Matrix Challenge

A small low-latency systems project built around a simple idea: a server streams matrix-multiplication challenges, and the client must parse, compute, and answer as fast as possible.

The challenge message format is:

challenge_id
N
A (N*N integers)
B (N*N integers)

The client replies with:

challenge_id checksum

where checksum = sum(A * B) mod 997.

What is in the repo

• hftserver2026/: reference server for local testing
• tools/blast_server.cpp: stress server with configurable rate, matrix size, window, and burst modes
• hftclient2026/main.cpp: current optimized client
• hftclient2026/main_v1.cpp, main_v2.cpp, main_v3.cpp: earlier client iterations kept for comparison

Current client

The current client focuses on low latency under load:

• streaming parser over a circular buffer
• flat matrix storage
• parallel compute workers
• dynamic tile scheduling
• cache-friendly tiled multiplication
• deferred modulo to remove % 997 from the inner hot loop
• macOS compute-thread QoS
• socket tuning with TCP_NODELAY and larger receive buffers
• built-in benchmark mode

In local testing, the biggest performance win came from deferred modulo, followed by better worker scheduling and tile sizing.

Version progression

• v1: first correct client, fixed TCP stream parsing and basic checksum computation
• v2: introduced a receive/parser/compute pipeline and worker threads
• v3: improved compute with tiling, cache-friendly loop order, and benchmark instrumentation
• v4 (main.cpp): benchmark-driven tuning of worker count, tile size, ring buffer size, deferred modulo, prefetching, and macOS scheduling hints

Build and run

Build with CMake:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j8

Start the reference server:

./build/bin/hftserver2026

Run the optimized client:

./build/bin/hftclient2026 127.0.0.1 12345 TeamA

Run with benchmark summary:

./build/bin/hftclient2026 127.0.0.1 12345 TeamA --benchmark

Stress testing

Example blast-server run:

./build/bin/blast_server --rate 200 --window 10 --size 128 --mode 1

The current client was tested against:

• high message rates
• burst / overlap traffic
• 64x64, 128x128, 256x256, and 512x512 matrices

and remained correct and stable in those local runs.