Skip to content

Latest commit

 

History

History
126 lines (99 loc) · 3.14 KB

README.md

File metadata and controls

126 lines (99 loc) · 3.14 KB
Raw

fwp

Go Reference

fwp (fast worker pool) is a simple, very fast bounded worker pool with an unlimited work queue.

When the worker pool is idle it consumes no memory (or goroutines).

Usage

// A worker pool with up to 1000 workers.
p := fwp.WorkerPool{Max: 1000}

p.Go(func() {
    // ...
})
p.Go(func() {
    // ...
})
// ...

If you need to wait for completion:

p := fwp.WorkerPool{Max: 1000}
var wg sync.WaitGroup

wg.Add(1)
p.Go(func() {
    // ...
    wg.Done()
})

wg.Wait()

It is possible to submit tasks from inside other tasks:

p := fwp.WorkerPool{Max: 1000}

p.Go(func() {
    p.Go(func() {
        // ...
    })
    // ...
})

If tasks depend on each other it is recommended, to prevent deadlocks that may be caused by Max tasks becoming blocked at the same time, to resubmit tasks (instead of blocking) in case a task is executed before its dependencies are ready:

p := fwp.WorkerPool{Max: 1000}

var fn func()
fn = func() {
    if some_precondition_is_not_yet_met {
        p.Go(fn)
        return
    }
    // ...
}
p.Go(fn)

Performance

fwp is pretty fast. Indeed it is faster than any other workerpool tested, and for high volumes of short tasks it can even be faster than spawning goroutines without a semaphore:

name                       time/op
FastWorkerPool-6            242ns ± 4%
GammazeroWorkerPool-6      1.42µs ± 1%
AlittoPond-6                405ns ± 8%
Panjf2000Ants-6            1.17µs ± 1%
GoroutineCond-6            1.04µs ± 3%
GoroutineCondPre-6          907ns ± 2%
GoroutineChannelSema-6      491ns ± 3%
GoroutineChannelSemaPre-6   735ns ± 3%
GoroutineXSyncSemaphore-6  1.66µs ± 5%
Goroutine-6                 266ns ±18%

The performance is due to three factors:

  • Goroutines are reused to process multiple tasks (this minimizes allocation of new goroutines as well as stack growths).
  • The number and length of critical sections is kept as low as possible (this minimizes contention on the mutex that guards the internals of the worker pool).
  • The internal behavior of the pool is adaptive to the workload, with 2 different regimes selected automatically based on the number and duration of tasks submitted (this prevents performance cliffs).

TODO

  • Use Intel TSX, ARM TME, or similar mechanisms to further minimize lock contention.
  • Investigate additional regimes:
    • When queue is empty delay worker shutdown by a few ns (or a roundtrip to the go scheduler) to wait for new tasks: as long as the delay is shorter than the time it takes for a hypotethical new goroutine to start executing a new task, it should be beneficial.
  • Reduce contention by using sharded queues or global+local queues with work stealing.
  • Use a chunked circular buffer (with chunk reuse). This should avoid copying the queue contents when the buffer needs to grow.

License

MIT