4 releases
Uses new Rust 2024
| 0.2.1 | Dec 10, 2025 |
|---|---|
| 0.2.0 | Dec 3, 2025 |
| 0.1.1 | Dec 3, 2025 |
| 0.1.0 | Dec 3, 2025 |
#301 in Concurrency
66KB
873 lines
SPMC broadcast channel for HFT and real-time systems
A lightweight, ultra-low-latency single-producer / multi-consumer broadcast channel designed for use in high-frequency trading (HFT) and other real-time systems.
Provides predictable performance, minimal contention, and carefully controlled memory access patterns. Works for thread-to-thread broadcasting via local memory, or inter-process communication using shared memory.
Features
- Lock-free single-producer / multi-consumer broadcast channel
- Seq_no protocol ensures overwrite detection
- Spin-wait receivers for ultra-low latency
- Cache-friendly layout (CachePadded)
- Works between threads, or between processes via shared memory
- Zero allocations after initialization
no_std-friendly design
Spin-Wait Behavior
Receivers use busy-waiting (spin loops) to wait for the producer to complete publishing the next message.
Implications:
- Lowest possible latency (no OS scheduling)
- A receiver consumes one logical CPU core while waiting
- Best when producer and receivers are on the same NUMA node
- Not ideal for scenarios where power efficiency matters
- No OS blocking (by design)
This design is intended for HFT, trading engines, matching engines, real-time telemetry, and other performance-critical workloads.
Installation
[dependencies]
hft-channel = "0.1"
Quick Example
use hft-channel::spmc_broadcast::channel;
let (tx, rx) = channel::<u64>("/test", 512);
let (_, payload) = tx.reserve();
*payload = 42;
tx.commit();
let (_, payload) = rx.recv();
Design Overview
The channel tracks state as:
[ dirty (1 bit, MSB) | seq_no (63 bits) ]
Guarantees:
- At least one slot is always dirty
- Dirty slot = being written, reserved for writing
- Clean slot = readable and stable
- If receiver sees seq_no changed, the slot was overwritten
Pattern:
- Sender marks next reserved slot dirty
- Sender writes payload into current reserved slot
- Sender clears dirty flag of current slot (= commit)
- Receiver waits for
dirty == false, then copies or peeks
Benchmarks
RUSTFLAGS="-C target-cpu=native" cargo build --release --example bench
./target/release/examples/bench --help
./target/release/examples/bench broadcast -p 1000 -m 2
Testing
cargo test
License
Copyright © 2005–2025 IKH Software, Inc.
Licensed under LGPL-3.0-or-later.
See LICENSE or https://www.gnu.org/licenses/lgpl-3.0.html.
Contributing
Contributions are welcome! Please open issues or pull requests on GitHub.
By submitting a contribution, you agree that it will be licensed under the project’s LGPL-3.0-or-later license.
Dependencies
~1MB
~17K SLoC