This C++14 library provides highly optimized, lock-free, multiple-producer-multiple-consumer (MPMC) queues based on circular buffers. It targets developers building ultra-low-latency systems, aiming to minimize the time between pushing and popping elements by employing a minimalist design with minimal atomic operations and explicit cache-friendliness.

How It Works

The queues utilize fixed-size circular buffers and std::atomic operations. Key design principles include minimizing atomic instructions, avoiding false sharing, using a linear buffer array, and avoiding heap allocations post-construction. This minimalist approach, including value semantics (copy/move on push/pop), aims for synergy to maximize CPU performance by reducing cache misses and pipeline stalls. Power-of-2 buffer sizes enable efficient index mapping and cache line contention reduction.

Quick Start & Requirements

• Install: Clone the repository and add atomic_queue/include to your build system's include paths. Alternatively, use vcpkg install atomic-queue or conan install atomic-queue.
• Prerequisites: C++14 compliant compiler. Benchmarks require Intel TBB.
• Links: GitHub, Benchmarks

Highlighted Details

• Offers AtomicQueue (for atomic types) and AtomicQueue2 (for non-atomic types), with Optimist variants for busy-waiting.
• Supports single-producer-single-consumer (SPSC) modes for further optimization.
• Provides optional totally ordered mode with zero cost on Intel x86.
• Benchmarked against std::mutex, boost::lockfree, moodycamel::ConcurrentQueue, and others, claiming superior latency and throughput.

Maintenance & Community

• The project is maintained by Maxim Egorushkin.
• Contributions are welcome.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Generally compatible with C++14 platforms supporting std::atomic. Reported compatible with Windows, but CI is Linux-only.

Limitations & Caveats

Queue size must be fixed at compile or construction time. The library relies on specific OS behaviors and thread scheduling (e.g., SCHED_FIFO) for optimal performance, and achieving the lowest latency requires careful system configuration. Debug builds include asserts for the NIL value, which must not be pushed.