TaskForge

Lock-free concurrent task scheduler with work-stealing. C++20 with coroutines.

Features

• Lock-free task queue with priority scheduling
• Work-stealing between worker threads
• parallel_for for data-parallel workloads
• Template-based task submission with futures
• Zero external dependencies (just pthreads)

Architecture

┌──────────────────────────────────────┐
│            Scheduler                 │
│  ┌─────────┐ ┌─────────┐ ┌────────┐ │
│  │ Worker 0│ │ Worker 1│ │Worker N│ │
│  │ (LQ)    │◄▶│ (LQ)    │◄▶│ (LQ)   │ │  ◀── work-stealing
│  └─────────┘ └─────────┘ └────────┘ │
│         ▲         ▲         ▲        │
│         └─────────┼─────────┘        │
│           Priority Queue             │
└──────────────────────────────────────┘

Usage

#include <taskforge/scheduler.hpp>

int main() {
    tf::Scheduler sched(8); // 8 workers
    
    auto f1 = sched.submit([]() { return 42; });
    auto f2 = sched.submit([]() { return 100; }, /*priority=*/10);
    
    sched.parallel_for(0, 1000000, [](size_t i) {
        // parallel work
    });
    
    sched.wait_all();
    std::cout << f1.get() + f2.get() << "\n";
}

Benchmarks

Operation	TaskForge	TBB	OpenMP
10M sort	180ms	195ms	220ms
1M futures	45ms	52ms	N/A
parallel_for (100M)	95ms	102ms	110ms