FastIPC

Zero-copy, lock-free IPC for Linux, designed as a drop-in replacement for FlexKV's ZMQ+pickle client-server control plane.

Design highlights

• POSIX shm slab pools for token_ids / slot_mapping / token_mask
• Fixed-size POD messages (104 B) — no pickle, no serialization
• Per-client SPSC rings in shm, multi-worker CAS-on-tail for MPMC consumption
• epoll_wait + eventfd edge-triggered wakeup — no sleep, no busy-loop
• C++17 core, pthread workers — no Python GIL contention on the hot path
• pybind11 bindings expose push_put(ndarray) / pull() -> dict

See ../fastipc_design.md for the full design document.

Build

pip install pybind11 numpy
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j
# Outputs:
#   build/libfastipc_core.a
#   build/fastipc_selftest             # C++ smoke test
#   build/python/fastipc/_fastipc.so   # Python extension

Quick test

# C++ smoke
./build/fastipc_selftest

# Python smoke
cd build/python && python3 -c "
import fastipc, numpy as np, threading
srv = fastipc.Server.create('demo', max_clients=1, num_workers=1,
                             pools=[(256*1024, 32), (16*1024*1024, 4)])
srv.start()

def consume():
    while True:
        req = srv.pull(timeout_ms=1000)
        if req is None: break
        srv.ack(req['dp_client_id'], req['task_id'], 0, None)

threading.Thread(target=consume, daemon=True).start()

cli = fastipc.Client.create('demo', 0, [(256*1024, 32), (16*1024*1024, 4)])
tid = cli.push_put(np.arange(1000, dtype=np.int64),
                   np.arange(1000, dtype=np.int64))
print('pushed', tid, '->', cli.pull(timeout_ms=2000))
srv.stop()
"

Benchmark vs ZMQ

python3 bench/bench_compare.py
# Writes fastipc_vs_zmq_report.md