Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
You, latency and profiling @ GopherCon India 2017
Search
Filippo Valsorda
February 25, 2017
Programming
4.3k
13
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
You, latency and profiling @ GopherCon India 2017
Filippo Valsorda
February 25, 2017
More Decks by Filippo Valsorda
See All by Filippo Valsorda
Asynchronous networking @ GopherCon 2018
filosottile
3
2.2k
Le note cifrate di Antonio Marzi
filosottile
1
460
Why cgo is slow @ CapitalGo 2018
filosottile
2
4.9k
Squeezing a key through a carry bit @ 34c3
filosottile
0
1.8k
Calling Rust from Go, without cgo @ GothamGo 2017
filosottile
1
3k
You, latency and profiling @ GolangUK 2017
filosottile
0
1.3k
Encrypting the Internet with Go @ GopherCon 2017
filosottile
9
2.8k
TLS 1.3 @ 33c3
filosottile
4
7k
Stealing Bitcoin With Math - HOPE XI
filosottile
13
30k
Other Decks in Programming
See All in Programming
AI駆動開発にグラフDBを重ねてみた
satoshi256kbyte
2
630
進化を続けるGo toolsの現在地 / The Current State of Ever-Evolving Go Tools
hond0413
0
290
DynamoDBの基礎を振り返りながらベクトル検索機能を理解する
musan
3
260
kubernetes コンポーネント開発入門 / 新卒N年目の勉強会&交流会!〜〇〇への誘い〜 #n_study
mazrean
0
140
[DroidKaigi 26] How We Moved from Android Studio to Slack
thisaay
0
150
思考垂れ流し開発 ~音声入力 × AIエージェント × 開発ハーネスによる試行錯誤~
npostring
0
710
Swift愛好会と私(ウホーイ) / Swift Fan Club and Uhooi
uhooi
0
120
30年振りにコンパイラの定数整数除算を改善した
herumi
9
4.4k
Webエンジニアなのにブラウザの仕組みがわからないので、Pythonで自作してみた
tatsuki12
4
1.6k
T3DD26: From RAGs to Riches
martinhelmich
0
130
ALB ログから Trace を気合で繋げる技術
fohte
7
840
LLMは4年分のCompose移行を再現できるのか?実プロダクト279件のXMLで探る自動化の境界線
makun
0
260
Featured
See All Featured
Helping Users Find Their Own Way: Creating Modern Search Experiences
danielanewman
31
3.3k
Visualization
eitanlees
152
17k
Imperfection Machines: The Place of Print at Facebook
scottboms
270
14k
Leveraging LLMs for student feedback in introductory data science courses - posit::conf(2025)
minecr
1
370
Data-driven link building: lessons from a $708K investment (BrightonSEO talk)
szymonslowik
1
1.3k
Building Experiences: Design Systems, User Experience, and Full Site Editing
marktimemedia
0
590
Scaling GitHub
holman
464
140k
The Power of CSS Pseudo Elements
geoffreycrofte
82
6.5k
Jamie Indigo - Trashchat’s Guide to Black Boxes: Technical SEO Tactics for LLMs
techseoconnect
PRO
0
640
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
333
23k
It's Worth the Effort
3n
188
29k
Agile that works and the tools we love
rasmusluckow
331
22k
Transcript
You, latency and profiling Filippo Valsorda — @FiloSottile
What is fast?
A fast hash… … can hash many MB per second
A fast regex… … can process many MB/s
A fast database… … can store many GB per second
A fast database… … can store many GB per second
… or replies to queries in a few milliseconds
A fast website… … can handle many requests at the
same time … or loads in a few milliseconds
A fast API… … can handle many clients at the
same time … or answers in a few milliseconds
Fast is… Throughput and Latency
CPU profiling -cpuprofile /debug/pprof/profile
CPU profiling SIGPROF
runtime/proc.go
CPU profiling http.Handler 1 Running on CPU http.Handler 2 I/O
I/O http.Handler 3 I/O http.Handler 4 I/O
CPU profiling Running on CPU I/O I/O I/O I/O http.Handler
1 http.Handler 2 http.Handler 3 http.Handler 4
CPU profiling I/O Observed latency http.Handler 1
func Write(data []byte) { for i := 0; i <
50; i++ { tmpfile, _ := ioutil.TempFile("", "ex") defer os.Remove(tmpfile.Name()) _, err = tmpfile.Write(data) tmpfile.Close() } } func Hash(data []byte) { for i := 0; i < 50; i++ { sha256.Sum256(data) } }
$ time curl http://127.0.0.1:12345/hash-and-write 9.831 total go tool pprof -web
slowserver cpu.pprof
None
$ time curl http://127.0.0.1:12345/hash-and-write 9.831 total $ time
curl http://127.0.0.1:12345/write-no-hash 7.692 total
Latency is not all about CPU
CPU profiling is for throughput The tracer is for latency
$ time curl http://127.0.0.1:12345/write-no-hash 7.692 total go tool trace -pprof=syscall
trace.out > syscall.pprof go tool pprof -web syscall.pprof
None
The tracer Detailed nanosecond-level log of execution events: • goroutine
scheduling, blocking • syscalls, network, I/O • garbage collection
The tracer • log files are BIG • 25% overhead
(after Go 1.7) • symbols are embedded (Go 1.7)
Trace vs. profile • Full event context • Goroutine metadata
and history • Discrete events, not samples
The tracer import "runtime/trace" import "net/http/pprof" /debug/pprof/trace?seconds=5
None
go tool trace -pprof=TYPE trace.out Supported profile types are: -
net: network blocking profile - sync: synchronization blocking profile - syscall: syscall blocking profile - sched: scheduler latency profile
go tool trace -pprof=syscall trace.out > syscall.pprof go tool pprof
-web syscall.pprof
go tool trace -pprof=syscall trace.out > syscall.pprof go-torch -b syscall.pprof
&& open torch.svg
None
None
-pprof=syscall func Write(data []byte) { for i := 0; i
< 50; i++ { tmpfile, _ := ioutil.TempFile("", "ex") defer os.Remove(tmpfile.Name()) _, err = tmpfile.Write(data) tmpfile.Close() } }
-pprof=sync // Block waits on a channel func Block() {
<-time.After(1 * time.Second) }
-pprof=sync
-pprof=sync
-pprof=net func Download() []byte { r, err := http.Get("https://golang.org/doc/go1.8") body,
err := ioutil.ReadAll(r.Body) r.Body.Close() return body }
-pprof=net
-pprof=net
None
Writing our own profile prof := make(map[uint64]pprof.Record) for _, ev
:= range events { if ev.Type != trace.EvGoBlockNet { continue } rec := prof[ev.StkID] rec.Stk = ev.Stk rec.N++ rec.Time += ev.Link.Ts - ev.Ts prof[ev.StkID] = rec } pprof.BuildProfile(prof).Write(os.Stdout)
var childG = make(map[uint64]struct{}) var lastGLen int for { for
_, ev := range events { if ev.Type != trace.EvGoCreate { continue } if _, ok := childG[ev.G]; !ok && !filterStack(ev.Stk, re) { continue } childG[ev.Args[0]] = struct{}{} } if len(childG) == lastGLen { break } lastGLen = len(childG) }
if _, ok := childG[ev.G]; !ok && !filterStack(ev.Stk, re) {
continue }
tracefocus tracefocus -filter=ServeHTTP trace.out > net.pprof github.com/FiloSottile/tracetools
None
None
Build more tools! • Focus on goroutine number • Aggregate
all blocking types • Visualizations • …
tracehist tracehist -filter=Close trace.out github.com/FiloSottile/tracetools
None
Build more tools! Profile latency!
Questions? Filippo Valsorda @FiloSottile filippo@cloudflare.com