Merge pull request #191 from ipfs/feat/tag-ewma

engine: tag peers based on usefulness

Author: Stebalien

decision/engine.go

@@ -57,11 +57,35 @@ const (
 	outboxChanBuffer = 0
 	// maxMessageSize is the maximum size of the batched payload
 	maxMessageSize = 512 * 1024
-	// tagPrefix is the tag given to peers associated an engine
-	tagPrefix = "bs-engine-%s"
+	// tagFormat is the tag given to peers associated an engine
+	tagFormat = "bs-engine-%s-%s"
 
-	// tagWeight is the default weight for peers associated with an engine
-	tagWeight = 5
+	// queuedTagWeight is the default weight for peers that have work queued
+	// on their behalf.
+	queuedTagWeight = 10
+
+	// the alpha for the EWMA used to track short term usefulness
+	shortTermAlpha = 0.5
+
+	// the alpha for the EWMA used to track long term usefulness
+	longTermAlpha = 0.05
+
+	// long term ratio defines what "long term" means in terms of the
+	// shortTerm duration. Peers that interact once every longTermRatio are
+	// considered useful over the long term.
+	longTermRatio = 10
+
+	// long/short term scores for tagging peers
+	longTermScore  = 10 // this is a high tag but it grows _very_ slowly.
+	shortTermScore = 10 // this is a high tag but it'll go away quickly if we aren't using the peer.
+)
+
+var (
+	// how frequently the engine should sample usefulness. Peers that
+	// interact every shortTerm time period are considered "active".
+	//
+	// this is only a variable to make testing easier.
+	shortTerm = 10 * time.Second
 )
 
 // Envelope contains a message for a Peer.
@@ -105,7 +129,8 @@ type Engine struct {
 
 	peerTagger PeerTagger
 
-	tag  string
+	tagQueued, tagUseful string
+
 	lock sync.Mutex // protects the fields immediatly below
 	// ledgerMap lists Ledgers by their Partner key.
 	ledgerMap map[peer.ID]*ledger
@@ -123,18 +148,118 @@ func NewEngine(ctx context.Context, bs bstore.Blockstore, peerTagger PeerTagger)
 		workSignal: make(chan struct{}, 1),
 		ticker:     time.NewTicker(time.Millisecond * 100),
 	}
-	e.tag = fmt.Sprintf(tagPrefix, uuid.New().String())
+	e.tagQueued = fmt.Sprintf(tagFormat, "queued", uuid.New().String())
+	e.tagUseful = fmt.Sprintf(tagFormat, "useful", uuid.New().String())
 	e.peerRequestQueue = peertaskqueue.New(peertaskqueue.OnPeerAddedHook(e.onPeerAdded), peertaskqueue.OnPeerRemovedHook(e.onPeerRemoved))
 	go e.taskWorker(ctx)
+	go e.scoreWorker(ctx)
 	return e
 }
 
+// scoreWorker keeps track of how "useful" our peers are, updating scores in the
+// connection manager.
+//
+// It does this by tracking two scores: short-term usefulness and long-term
+// usefulness. Short-term usefulness is sampled frequently and highly weights
+// new observations. Long-term usefulness is sampled less frequently and highly
+// weights on long-term trends.
+//
+// In practice, we do this by keeping two EWMAs. If we see an interaction
+// within the sampling period, we record the score, otherwise, we record a 0.
+// The short-term one has a high alpha and is sampled every shortTerm period.
+// The long-term one has a low alpha and is sampled every
+// longTermRatio*shortTerm period.
+//
+// To calculate the final score, we sum the short-term and long-term scores then
+// adjust it ±25% based on our debt ratio. Peers that have historically been
+// more useful to us than we are to them get the highest score.
+func (e *Engine) scoreWorker(ctx context.Context) {
+	ticker := time.NewTicker(shortTerm)
+	defer ticker.Stop()
+
+	type update struct {
+		peer  peer.ID
+		score int
+	}
+	var (
+		lastShortUpdate, lastLongUpdate time.Time
+		updates                         []update
+	)
+
+	for i := 0; ; i = (i + 1) % longTermRatio {
+		var now time.Time
+		select {
+		case now = <-ticker.C:
+		case <-ctx.Done():
+			return
+		}
+
+		// The long term update ticks every `longTermRatio` short
+		// intervals.
+		updateLong := i == 0
+
+		e.lock.Lock()
+		for _, ledger := range e.ledgerMap {
+			ledger.lk.Lock()
+
+			// Update the short-term score.
+			if ledger.lastExchange.After(lastShortUpdate) {
+				ledger.shortScore = ewma(ledger.shortScore, shortTermScore, shortTermAlpha)
+			} else {
+				ledger.shortScore = ewma(ledger.shortScore, 0, shortTermAlpha)
+			}
+
+			// Update the long-term score.
+			if updateLong {
+				if ledger.lastExchange.After(lastLongUpdate) {
+					ledger.longScore = ewma(ledger.longScore, longTermScore, longTermAlpha)
+				} else {
+					ledger.longScore = ewma(ledger.longScore, 0, longTermAlpha)
+				}
+			}
+
+			// Calculate the new score.
+			//
+			// The accounting score adjustment prefers peers _we_
+			// need over peers that need us. This doesn't help with
+			// leeching.
+			score := int((ledger.shortScore + ledger.longScore) * ((ledger.Accounting.Score())*.5 + .75))
+
+			// Avoid updating the connection manager unless there's a change. This can be expensive.
+			if ledger.score != score {
+				// put these in a list so we can perform the updates outside _global_ the lock.
+				updates = append(updates, update{ledger.Partner, score})
+				ledger.score = score
+			}
+			ledger.lk.Unlock()
+		}
+		e.lock.Unlock()
+
+		// record the times.
+		lastShortUpdate = now
+		if updateLong {
+			lastLongUpdate = now
+		}
+
+		// apply the updates
+		for _, update := range updates {
+			if update.score == 0 {
+				e.peerTagger.UntagPeer(update.peer, e.tagUseful)
+			} else {
+				e.peerTagger.TagPeer(update.peer, e.tagUseful, update.score)
+			}
+		}
+		// Keep the memory. It's not much and it saves us from having to allocate.
+		updates = updates[:0]
+	}
+}
+
 func (e *Engine) onPeerAdded(p peer.ID) {
-	e.peerTagger.TagPeer(p, e.tag, tagWeight)
+	e.peerTagger.TagPeer(p, e.tagQueued, queuedTagWeight)
 }
 
 func (e *Engine) onPeerRemoved(p peer.ID) {
-	e.peerTagger.UntagPeer(p, e.tag)
+	e.peerTagger.UntagPeer(p, e.tagQueued)
 }
 
 // WantlistForPeer returns the currently understood want list for a given peer

decision/engine_test.go

@@ -19,38 +19,63 @@ import (
 	testutil "github.com/libp2p/go-libp2p-core/test"
 )
 
+type peerTag struct {
+	done  chan struct{}
+	peers map[peer.ID]int
+}
+
 type fakePeerTagger struct {
-	lk          sync.Mutex
-	wait        sync.WaitGroup
-	taggedPeers []peer.ID
+	lk   sync.Mutex
+	tags map[string]*peerTag
 }
 
 func (fpt *fakePeerTagger) TagPeer(p peer.ID, tag string, n int) {
-	fpt.wait.Add(1)
-
 	fpt.lk.Lock()
 	defer fpt.lk.Unlock()
-	fpt.taggedPeers = append(fpt.taggedPeers, p)
+	if fpt.tags == nil {
+		fpt.tags = make(map[string]*peerTag, 1)
+	}
+	pt, ok := fpt.tags[tag]
+	if !ok {
+		pt = &peerTag{peers: make(map[peer.ID]int, 1), done: make(chan struct{})}
+		fpt.tags[tag] = pt
+	}
+	pt.peers[p] = n
 }
 
 func (fpt *fakePeerTagger) UntagPeer(p peer.ID, tag string) {
-	defer fpt.wait.Done()
-
 	fpt.lk.Lock()
 	defer fpt.lk.Unlock()
-	for i := 0; i < len(fpt.taggedPeers); i++ {
-		if fpt.taggedPeers[i] == p {
-			fpt.taggedPeers[i] = fpt.taggedPeers[len(fpt.taggedPeers)-1]
-			fpt.taggedPeers = fpt.taggedPeers[:len(fpt.taggedPeers)-1]
-			return
-		}
+	pt := fpt.tags[tag]
+	if pt == nil {
+		return
+	}
+	delete(pt.peers, p)
+	if len(pt.peers) == 0 {
+		close(pt.done)
+		delete(fpt.tags, tag)
 	}
 }
 
-func (fpt *fakePeerTagger) count() int {
+func (fpt *fakePeerTagger) count(tag string) int {
 	fpt.lk.Lock()
 	defer fpt.lk.Unlock()
-	return len(fpt.taggedPeers)
+	if pt, ok := fpt.tags[tag]; ok {
+		return len(pt.peers)
+	}
+	return 0
+}
+
+func (fpt *fakePeerTagger) wait(tag string) {
+	fpt.lk.Lock()
+	pt := fpt.tags[tag]
+	if pt == nil {
+		fpt.lk.Unlock()
+		return
+	}
+	doneCh := pt.done
+	fpt.lk.Unlock()
+	<-doneCh
 }
 
 type engineSet struct {
@@ -241,16 +266,56 @@ func TestTaggingPeers(t *testing.T) {
 	next := <-sanfrancisco.Engine.Outbox()
 	envelope := <-next
 
-	if sanfrancisco.PeerTagger.count() != 1 {
+	if sanfrancisco.PeerTagger.count(sanfrancisco.Engine.tagQueued) != 1 {
 		t.Fatal("Incorrect number of peers tagged")
 	}
 	envelope.Sent()
 	<-sanfrancisco.Engine.Outbox()
-	sanfrancisco.PeerTagger.wait.Wait()
-	if sanfrancisco.PeerTagger.count() != 0 {
+	sanfrancisco.PeerTagger.wait(sanfrancisco.Engine.tagQueued)
+	if sanfrancisco.PeerTagger.count(sanfrancisco.Engine.tagQueued) != 0 {
 		t.Fatal("Peers should be untagged but weren't")
 	}
 }
+
+func TestTaggingUseful(t *testing.T) {
+	oldShortTerm := shortTerm
+	shortTerm = 1 * time.Millisecond
+	defer func() { shortTerm = oldShortTerm }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Second)
+	defer cancel()
+	me := newEngine(ctx, "engine")
+	friend := peer.ID("friend")
+
+	block := blocks.NewBlock([]byte("foobar"))
+	msg := message.New(false)
+	msg.AddBlock(block)
+
+	for i := 0; i < 3; i++ {
+		if me.PeerTagger.count(me.Engine.tagUseful) != 0 {
+			t.Fatal("Peers should be untagged but weren't")
+		}
+		me.Engine.MessageSent(friend, msg)
+		time.Sleep(shortTerm * 2)
+		if me.PeerTagger.count(me.Engine.tagUseful) != 1 {
+			t.Fatal("Peers should be tagged but weren't")
+		}
+		time.Sleep(shortTerm * 8)
+	}
+
+	if me.PeerTagger.count(me.Engine.tagUseful) == 0 {
+		t.Fatal("peers should still be tagged due to long-term usefulness")
+	}
+	time.Sleep(shortTerm * 2)
+	if me.PeerTagger.count(me.Engine.tagUseful) == 0 {
+		t.Fatal("peers should still be tagged due to long-term usefulness")
+	}
+	time.Sleep(shortTerm * 10)
+	if me.PeerTagger.count(me.Engine.tagUseful) != 0 {
+		t.Fatal("peers should finally be untagged")
+	}
+}
+
 func partnerWants(e *Engine, keys []string, partner peer.ID) {
 	add := message.New(false)
 	for i, letter := range keys {

decision/ewma.go

@@ -0,0 +1,5 @@
+package decision
+
+func ewma(old, new, alpha float64) float64 {
+	return new*alpha + (1-alpha)*old
+}

decision/ledger.go

@@ -12,9 +12,8 @@ import (
 
 func newLedger(p peer.ID) *ledger {
 	return &ledger{
-		wantList:   wl.New(),
-		Partner:    p,
-		sentToPeer: make(map[string]time.Time),
+		wantList: wl.New(),
+		Partner:  p,
 	}
 }
 
@@ -30,16 +29,19 @@ type ledger struct {
 	// lastExchange is the time of the last data exchange.
 	lastExchange time.Time
 
+	// These scores keep track of how useful we think this peer is. Short
+	// tracks short-term usefulness and long tracks long-term usefulness.
+	shortScore, longScore float64
+	// Score keeps track of the score used in the peer tagger. We track it
+	// here to avoid unnecessarily updating the tags in the connection manager.
+	score int
+
 	// exchangeCount is the number of exchanges with this peer
 	exchangeCount uint64
 
 	// wantList is a (bounded, small) set of keys that Partner desires.
 	wantList *wl.Wantlist
 
-	// sentToPeer is a set of keys to ensure we dont send duplicate blocks
-	// to a given peer
-	sentToPeer map[string]time.Time
-
 	// ref is the reference count for this ledger, its used to ensure we
 	// don't drop the reference to this ledger in multi-connection scenarios
 	ref int
@@ -63,10 +65,19 @@ type debtRatio struct {
 	BytesRecv uint64
 }
 
+// Value returns the debt ratio, sent:receive.
 func (dr *debtRatio) Value() float64 {
 	return float64(dr.BytesSent) / float64(dr.BytesRecv+1)
 }
 
+// Score returns the debt _score_ on a 0-1 scale.
+func (dr *debtRatio) Score() float64 {
+	if dr.BytesRecv == 0 {
+		return 0
+	}
+	return float64(dr.BytesRecv) / float64(dr.BytesRecv+dr.BytesSent)
+}
+
 func (l *ledger) SentBytes(n int) {
 	l.exchangeCount++
 	l.lastExchange = time.Now()

go.mod

@@ -7,7 +7,6 @@ require (
 	github.com/gogo/protobuf v1.2.1
 	github.com/golang/protobuf v1.3.1 // indirect
 	github.com/google/uuid v1.1.1
-	github.com/hashicorp/golang-lru v0.5.1
 	github.com/ipfs/go-block-format v0.0.2
 	github.com/ipfs/go-cid v0.0.2
 	github.com/ipfs/go-datastore v0.0.5
@@ -38,3 +37,5 @@ require (
 	golang.org/x/text v0.3.2 // indirect
 	gopkg.in/yaml.v2 v2.2.2 // indirect
 )
+
+go 1.12