Merge pull request #237 from raulk/feat/adaptive-dial-queue

Adaptive queue for staging dials

dial_queue.go

@@ -0,0 +1,293 @@
+package dht
+
+import (
+	"context"
+	"math"
+	"time"
+
+	peer "github.com/libp2p/go-libp2p-peer"
+	queue "github.com/libp2p/go-libp2p-peerstore/queue"
+)
+
+var (
+	// DialQueueMinParallelism is the minimum number of worker dial goroutines that will be alive at any time.
+	DialQueueMinParallelism = 6
+	// DialQueueMaxParallelism is the maximum number of worker dial goroutines that can be alive at any time.
+	DialQueueMaxParallelism = 20
+	// DialQueueMaxIdle is the period that a worker dial goroutine waits before signalling a worker pool downscaling.
+	DialQueueMaxIdle = 5 * time.Second
+	// DialQueueScalingMutePeriod is the amount of time to ignore further worker pool scaling events, after one is
+	// processed. Its role is to reduce jitter.
+	DialQueueScalingMutePeriod = 1 * time.Second
+)
+
+type dialQueue struct {
+	ctx    context.Context
+	dialFn func(context.Context, peer.ID) error
+
+	nWorkers      int
+	scalingFactor float64
+
+	in  *queue.ChanQueue
+	out *queue.ChanQueue
+
+	waitingCh chan waitingCh
+	dieCh     chan struct{}
+	growCh    chan struct{}
+	shrinkCh  chan struct{}
+}
+
+type waitingCh struct {
+	ch chan<- peer.ID
+	ts time.Time
+}
+
+// newDialQueue returns an adaptive dial queue that spawns a dynamically sized set of goroutines to preemptively
+// stage dials for later handoff to the DHT protocol for RPC. It identifies backpressure on both ends (dial consumers
+// and dial producers), and takes compensating action by adjusting the worker pool.
+//
+// Why? Dialing is expensive. It's orders of magnitude slower than running an RPC on an already-established
+// connection, as it requires establishing a TCP connection, multistream handshake, crypto handshake, mux handshake,
+// and protocol negotiation.
+//
+// We start with DialQueueMinParallelism number of workers, and scale up and down based on demand and supply of
+// dialled peers.
+//
+// The following events trigger scaling:
+// - we scale up when we can't immediately return a successful dial to a new consumer.
+// - we scale down when we've been idle for a while waiting for new dial attempts.
+// - we scale down when we complete a dial and realise nobody was waiting for it.
+//
+// Dialler throttling (e.g. FD limit exceeded) is a concern, as we can easily spin up more workers to compensate, and
+// end up adding fuel to the fire. Since we have no deterministic way to detect this for now, we hard-limit concurrency
+// to DialQueueMaxParallelism.
+func newDialQueue(ctx context.Context, target string, in *queue.ChanQueue, dialFn func(context.Context, peer.ID) error) *dialQueue {
+	sq := &dialQueue{
+		ctx:           ctx,
+		dialFn:        dialFn,
+		nWorkers:      DialQueueMinParallelism,
+		scalingFactor: 1.5,
+
+		in:  in,
+		out: queue.NewChanQueue(ctx, queue.NewXORDistancePQ(target)),
+
+		growCh:    make(chan struct{}, 1),
+		shrinkCh:  make(chan struct{}, 1),
+		waitingCh: make(chan waitingCh),
+		dieCh:     make(chan struct{}, DialQueueMaxParallelism),
+	}
+	for i := 0; i < DialQueueMinParallelism; i++ {
+		go sq.worker()
+	}
+	go sq.control()
+	return sq
+}
+
+func (dq *dialQueue) control() {
+	var (
+		dialled        <-chan peer.ID
+		waiting        []waitingCh
+		lastScalingEvt = time.Now()
+	)
+
+	defer func() {
+		for _, w := range waiting {
+			close(w.ch)
+		}
+		waiting = nil
+	}()
+
+	for {
+		// First process any backlog of dial jobs and waiters -- making progress is the priority.
+		// This block is copied below; couldn't find a more concise way of doing this.
+		select {
+		case <-dq.ctx.Done():
+			return
+		case w := <-dq.waitingCh:
+			waiting = append(waiting, w)
+			dialled = dq.out.DeqChan
+			continue // onto the top.
+		case p, ok := <-dialled:
+			if !ok {
+				return // we're done if the ChanQueue is closed, which happens when the context is closed.
+			}
+			w := waiting[0]
+			log.Debugf("delivering dialled peer to DHT; took %dms.", time.Since(w.ts)/time.Millisecond)
+			w.ch <- p
+			close(w.ch)
+			waiting = waiting[1:]
+			if len(waiting) == 0 {
+				// no more waiters, so stop consuming dialled jobs.
+				dialled = nil
+			}
+			continue // onto the top.
+		default:
+			// there's nothing to process, so proceed onto the main select block.
+		}
+
+		select {
+		case <-dq.ctx.Done():
+			return
+		case w := <-dq.waitingCh:
+			waiting = append(waiting, w)
+			dialled = dq.out.DeqChan
+		case p, ok := <-dialled:
+			if !ok {
+				return // we're done if the ChanQueue is closed, which happens when the context is closed.
+			}
+			w := waiting[0]
+			log.Debugf("delivering dialled peer to DHT; took %dms.", time.Since(w.ts)/time.Millisecond)
+			w.ch <- p
+			close(w.ch)
+			waiting = waiting[1:]
+			if len(waiting) == 0 {
+				// no more waiters, so stop consuming dialled jobs.
+				dialled = nil
+			}
+		case <-dq.growCh:
+			if time.Since(lastScalingEvt) < DialQueueScalingMutePeriod {
+				continue
+			}
+			dq.grow()
+			lastScalingEvt = time.Now()
+		case <-dq.shrinkCh:
+			if time.Since(lastScalingEvt) < DialQueueScalingMutePeriod {
+				continue
+			}
+			dq.shrink()
+			lastScalingEvt = time.Now()
+		}
+	}
+}
+
+func (dq *dialQueue) Consume() <-chan peer.ID {
+	ch := make(chan peer.ID, 1)
+
+	select {
+	case p := <-dq.out.DeqChan:
+		// short circuit and return a dialled peer if it's immediately available.
+		ch <- p
+		close(ch)
+		return ch
+	case <-dq.ctx.Done():
+		// return a closed channel with no value if we're done.
+		close(ch)
+		return ch
+	default:
+	}
+
+	// we have no finished dials to return, trigger a scale up.
+	select {
+	case dq.growCh <- struct{}{}:
+	default:
+	}
+
+	// park the channel until a dialled peer becomes available.
+	select {
+	case dq.waitingCh <- waitingCh{ch, time.Now()}:
+		// all good
+	case <-dq.ctx.Done():
+		// return a closed channel with no value if we're done.
+		close(ch)
+	}
+	return ch
+}
+
+func (dq *dialQueue) grow() {
+	// no mutex needed as this is only called from the (single-threaded) control loop.
+	defer func(prev int) {
+		if prev == dq.nWorkers {
+			return
+		}
+		log.Debugf("grew dial worker pool: %d => %d", prev, dq.nWorkers)
+	}(dq.nWorkers)
+
+	if dq.nWorkers == DialQueueMaxParallelism {
+		return
+	}
+	target := int(math.Floor(float64(dq.nWorkers) * dq.scalingFactor))
+	if target > DialQueueMaxParallelism {
+		target = DialQueueMinParallelism
+	}
+	for ; dq.nWorkers < target; dq.nWorkers++ {
+		go dq.worker()
+	}
+}
+
+func (dq *dialQueue) shrink() {
+	// no mutex needed as this is only called from the (single-threaded) control loop.
+	defer func(prev int) {
+		if prev == dq.nWorkers {
+			return
+		}
+		log.Debugf("shrunk dial worker pool: %d => %d", prev, dq.nWorkers)
+	}(dq.nWorkers)
+
+	if dq.nWorkers == DialQueueMinParallelism {
+		return
+	}
+	target := int(math.Floor(float64(dq.nWorkers) / dq.scalingFactor))
+	if target < DialQueueMinParallelism {
+		target = DialQueueMinParallelism
+	}
+	// send as many die signals as workers we have to prune.
+	for ; dq.nWorkers > target; dq.nWorkers-- {
+		select {
+		case dq.dieCh <- struct{}{}:
+		default:
+			log.Debugf("too many die signals queued up.")
+		}
+	}
+}
+
+func (dq *dialQueue) worker() {
+	// This idle timer tracks if the environment is slow. If we're waiting to long to acquire a peer to dial,
+	// it means that the DHT query is progressing slow and we should shrink the worker pool.
+	idleTimer := time.NewTimer(24 * time.Hour) // placeholder init value which will be overridden immediately.
+	for {
+		// trap exit signals first.
+		select {
+		case <-dq.ctx.Done():
+			return
+		case <-dq.dieCh:
+			return
+		default:
+		}
+
+		idleTimer.Stop()
+		select {
+		case <-idleTimer.C:
+		default:
+		}
+		idleTimer.Reset(DialQueueMaxIdle)
+
+		select {
+		case <-dq.dieCh:
+			return
+		case <-dq.ctx.Done():
+			return
+		case <-idleTimer.C:
+			// no new dial requests during our idle period; time to scale down.
+		case p := <-dq.in.DeqChan:
+			t := time.Now()
+			if err := dq.dialFn(dq.ctx, p); err != nil {
+				log.Debugf("discarding dialled peer because of error: %v", err)
+				continue
+			}
+			log.Debugf("dialling %v took %dms (as observed by the dht subsystem).", p, time.Since(t)/time.Millisecond)
+			waiting := len(dq.waitingCh)
+			dq.out.EnqChan <- p
+			if waiting > 0 {
+				// we have somebody to deliver this value to, so no need to shrink.
+				continue
+			}
+		}
+
+		// scaling down; control only arrives here if the idle timer fires, or if there are no goroutines
+		// waiting for the value we just produced.
+		select {
+		case dq.shrinkCh <- struct{}{}:
+		default:
+		}
+	}
+}