EthBerlinZwei: Profiling Libp2p Gossipsub, Golang version
Hackathon submission by @protolambda, learning libp2p with a non-networking background.
Short presentation: https://docs.google.com/presentation/d/1lL4QiPgbdDVIp9sSp7KL2oZ6Oquis-rCGU3MUkTX9BU/edit?usp=sharing
The bounty problem
Find and fix bottlenecks and performance hotspots in the Go implementation of gossipsub.
See this issue on
And so there it starts; read up on libp2p knowledge, read the Gossipsub spec and then trial-and-error throughout the hackathon. I started late however, since I worked on other Eth 2 issues too. Thanks to @raulk for getting me up to speed to work on this so fast.
Note that this is a hack, produced with a "see it work first" mindset, not a research paper. You are welcome to fork and improve the profiling.
To profile anything at all, some kind of test-run is necessary. One that stresses go-libp2p with a high throughput, with a good amount of peers and topics. Then, a PPROF profile can be made of the test-run, and help identify hotspots to optimize.
Why not use benchmarking?
Since the task is not to benchmark libp2p (discussed options here with @raulk however), but to profile and find (and fix) the hotspots, a more practical test-run with the actual overhead of opening a connection and not sharing memory for messages helps identify hotspots.
Also, the message-interval and size parameters are less strict: they can definitely affect speed, but there are only so many extremes to find hotspots for.
Benchmarking of the isolated gossipsub logic would be better if done with a mock net, something like this This however hides the overhead introduced by passing messages to a real socket, skewing the priorities in what to optimize for. If practical issues are solved, one could then use Perf to profile a Go benchmark with, and look into the memory allocations and flamegraph of the remaining calls. The bigger picture found in call-graphs in a non-benchmark setting does not show gossipsub code itself to be the bottleneck in practice however, hence not going the benchmarking route.
Common settings for the produced hackathon results (not claiming perfectness, time constraints to for pretty parametrization apply):
// total hosts hostCount := 100 // peers per host (randomly assigned) degree := 10 // pubsub topic chances: "/libp2p/example/berlin/protolambda/foo": 0.7, "/libp2p/example/berlin/protolambda/bar": 0.4, "/libp2p/example/berlin/protolambda/quix": 0.8, // A no-op logger is used during benchmarking for speed. logger := zwei.NewDebugLogger(nil) // For debugging this can be changed to: // logger := zwei.NewDebugLogger(log.New(os.Stdout, "experiment: ", log.Lmicroseconds)) // message size // big: 8 - 15 KB minMsgByteLen := 8 << 10 maxMsgByteLen := 16 << 10 // small: 10 bytes minMsgByteLen := 10 maxMsgByteLen := 10 // publish interval range for each simulated host (publish on 1 random topic) minSleepMs := 100 maxSleepMs := 300 // libp2p settings // transport: libp2p.Transport(tcp.NewTCPTransport), // mux choice: libp2p.Muxer("/yamux/1.0.0", yamux.DefaultTransport), //libp2p.Muxer("/mplex/6.7.0", mplex.DefaultTransport), // for some later profiles with mplex // security: libp2p.Security(secio.ID, secio.New), // GossipSub settings // Initially true, signing with Secp256k1. // Later disabled, since this was the biggest practical bottleneck, and obfuscates the smaller differences. pubsub.WithMessageSigning(true) // loopback through localhost, with no artificial latency libp2p.ListenAddrStrings( "/ip4/127.0.0.1/tcp/0", // 0: gets a random port assigned on localhost ), // There also are options to change the RNG seed for both initialization and the testrun itself, // but libp2p (interaction with machine itself, and go-routine scheduling) is not deterministic enough // to make the results fully reproducible.
zwei.Experimentis created with these. Message length and interval can be changed in the experiment code, if required.
- PPROF CPU-Profiling starts after setting up the experiment (starting hosts, starting gossipsub, and subscribing to topics)
- Start experiment
- Wait for stop-signal
- Stop profiling, save results, see log output for profiling output location.
- Stop libp2p tasks and close resources with
To generate a call-graph:
go tool pprof -web /tmp/profile......../cpu.pprof
Early results settings: Yamux, signed GossipSub, small 10 byte messages: signature verification is the clear bottleneck.
This test run published 31k messages, 1500k were received. A 145 seconds run. Full callgraph SVG
Then, I disabled GossipSub signatures (
pubsub.WithMessageSigning(false)) to see what was left.
For small messages, it is Yamux triggering secio encryption, which then writes to a socket connection provided by the kernel, which also forms a bottleneck.
This test run published 8k messages, 600k were received. A 20 seconds run. Full callgraph SVG
Raul then recommended to increase the message size, so repeat this with random 8 - 16 KB messages:
This test run published 5k messages, 200k were received. A 90 seconds run. Note the significantly lower throughput. Full callgraph SVG
For larger messages, SHA-256 calls by Yamux become the bottleneck. However, it looks like it is already using the excellent Sha-256 SIMD library for speed, so there is not much to gain unless something is being hashed twice and can be cached.
Now try again with Mplex:
This test run published 8k messages, 357k were received. A 33 seconds run. Note the significantly lower throughput. Full callgraph SVG
SHA-256 (and general secio crypto) is still by far the biggest bottleneck.
GossipSub itself is primarily limited by the crypto necessary to verify and encrypt the messages, and the data-structures used in its implementation do not seem to be worth optimizing at this time.
There seems to be some interesting difference in mplex vs. yamux to look into at a later moment, if it is not a usage problem from my side.
Some initial code was adapted from the go-libp2p examples repository,
also licensed with MIT.
Log in or sign up for Devpost to join the conversation.