Performance

FastQueue is designed for high-throughput, low-latency task execution.

Benchmark Results

Average throughput across multiple runs:

Configuration	Jobs	Throughput
1 thread, 10K jobs	10,000	~1.4M jobs/s
2 threads, 10K jobs	10,000	~1.1M jobs/s
4 threads, 10K jobs	10,000	~1.3M jobs/s
4 threads, 100K jobs	100,000	~1.0M jobs/s
4 threads, 1M jobs	1,000,000	~1.0M jobs/s

Running Benchmarks

cmake -B build -DFQ_BUILD_BENCHMARKS=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
./build/benchmarks/Release/fq_benchmark

Optimization Tips

1. Right-size the Thread Count

// Use hardware concurrency
cfg.thread_count = 0; // Auto-detect

// Or set manually
cfg.thread_count = 8; // Match your CPU cores

2. Enable Work Stealing

cfg.enable_work_stealing = FQ_TRUE;

3. Use Appropriate Queue Capacity

// For many small tasks
cfg.queue_capacity = 8192;

// For fewer large tasks
cfg.queue_capacity = 256;

4. Minimize Lock Contention

Submit tasks from multiple threads to distribute across worker queues:

// Good: Multiple threads submitting
#pragma omp parallel for
for (int i = 0; i < 1000; ++i) {
    fq_thread_pool_submit_fn(pool, task, &data[i]);
}

// Bad: Single thread submitting all
for (int i = 0; i < 1000; ++i) {
    fq_thread_pool_submit_fn(pool, task, &data[i]);
}

5. Avoid Unnecessary Waiting

// Use futures instead of wait_idle
fq_future_t *future = NULL;
fq_scheduler_submit_with_future(scheduler, task, &future);
// Do other work...
fq_future_wait(future); // Wait only when needed

Profiling

Use your platform's profiler to identify bottlenecks:

# Linux perf
perf record ./build/benchmarks/Release/fq_benchmark
perf report

# Valgrind
valgrind --tool=callgrind ./build/benchmarks/Release/fq_benchmark

# macOS Instruments
instruments -t "Time Profiler" ./build/benchmarks/Release/fq_benchmark