swarm: implement smart dialing logic (#2260)

* implement smart dialing

* add more comments and tests

* change address ranking logic to dial one quic address before others

* add randomized worker loop tests

* simplify priority queue implementation

* improve DialRanker docs

* one more test

* add explanatory comments and rename variables

* fix allocations in dialQueue

* fix allocations in dialRanker

* Apply suggestions from code review

Co-authored-by: Marten Seemann <martenseemann@gmail.com>

* fix comments

* add logging

* add holepunching test

* add metrics for tracking dial prioritisation impact

clean up redundant address filtering committed

* add test for webtransport filtering

* update changelog

* fix flaky test

* update dashboard

* update dial ranking delay dashboard to use pie chart

* change <=1ms label to 'No delay' in dashboard

* add defensive check to map presence

---------

Co-authored-by: Marten Seemann <martenseemann@gmail.com>

CHANGELOG.md

@@ -1,4 +1,5 @@
 # Table Of Contents <!-- omit in toc -->
+- [v0.28.0](#v0280)
 - [v0.27.0](#v0270)
 - [v0.26.4](#v0264)
 - [v0.26.3](#v0263)
@@ -8,6 +9,17 @@
 - [v0.25.1](#v0251)
 - [v0.25.0](#v0250)
 
+# [v0.28.0]()
+
+## 🔦 Highlights <!-- omit in toc -->
+
+### Smart Dialing <!-- omit in toc -->
+* When connecting to a peer we now do [happy eyeballs](https://www.rfc-editor.org/rfc/rfc8305) like dial prioritisation to prefer QUIC addresses over TCP addresses. We dial the QUIC address first and wait 250ms to dial the TCP address of the peer.
+* In our experiments we've seen little impact on latencies up to 80th percentile. 90th and 95th percentile latencies are impacted. For details see discussion on the [PR](https://github.com/libp2p/go-libp2p/pull/2260#issuecomment-1528848170).
+* For details of the address ranking logic see godoc for `swarm.DefaultDialRanker`.
+* To disable smart dialing and keep the old behaviour use the 
+`libp2p.NoDelayNetworkDialRanker` option.
+
 # [v0.27.0](https://github.com/libp2p/go-libp2p/releases/tag/v0.27.0)
 
 ### Breaking Changes <!-- omit in toc -->

config/config.go

@@ -123,6 +123,8 @@ type Config struct {
 
 	DisableMetrics       bool
 	PrometheusRegisterer prometheus.Registerer
+
+	NoDelayNetworkDialRanker bool
 }
 
 func (cfg *Config) makeSwarm(eventBus event.Bus, enableMetrics bool) (*swarm.Swarm, error) {
@@ -173,6 +175,9 @@ func (cfg *Config) makeSwarm(eventBus event.Bus, enableMetrics bool) (*swarm.Swa
 	if cfg.MultiaddrResolver != nil {
 		opts = append(opts, swarm.WithMultiaddrResolver(cfg.MultiaddrResolver))
 	}
+	if cfg.NoDelayNetworkDialRanker {
+		opts = append(opts, swarm.WithNoDialDelay())
+	}
 	if enableMetrics {
 		opts = append(opts,
 			swarm.WithMetricsTracer(swarm.NewMetricsTracer(swarm.WithRegisterer(cfg.PrometheusRegisterer))))

core/network/network.go

@@ -187,6 +187,16 @@ type Dialer interface {
 	StopNotify(Notifiee)
 }
 
+// AddrDelay provides an address along with the delay after which the address
+// should be dialed
+type AddrDelay struct {
+	Addr  ma.Multiaddr
+	Delay time.Duration
+}
+
+// DialRanker provides a schedule of dialing the provided addresses
+type DialRanker func([]ma.Multiaddr) []AddrDelay
+
 // DedupAddrs deduplicates addresses in place, leave only unique addresses.
 // It doesn't allocate.
 func DedupAddrs(addrs []ma.Multiaddr) []ma.Multiaddr {