SlideShare a Scribd company logo

Why my Go program is slow?

•
•
Inada Naoki
Inada Naoki

The document discusses why a Go program may be slow and provides tips for profiling and optimizing Go code performance. It introduces the pprof CPU profiler and how to use it to analyze performance bottlenecks. Some factors that can slow down Go programs include garbage collection, memory copying, and function calls. The document hopes that future versions of Go will include features like concurrent garbage collection to improve speed.

1 of 17
Downloaded 21 times
Why my Go program is slow? GoCon 2014 autumn, INADA Naoki
INADA Naoki • @methane • KLab Inc. • Python • mysqlclient, PyMySQL, msgpack
And I’m gopher • I love Go. It’s more Pythonic than Python! • Explicit is better than implicit. • Simple is better than complex. • Readability counts. • There should be one-- and preferably only one --obvious way to do it. • Although that way may not be obvious at first unless you're Dutch (Gopher).
Today • CPU profiling with pprof • Random notes about Go’s performance. • What I hope Go 1.5~ will have
CPU profiling with pprof
runtime/pprof • Sampling profiler. • net/http/pprof provides HTTP API for pprof • github.com/davecheney/profile provice easy API for CLI program.
NOTE for Mac • Go’s CPU profiling is not work on Mac • Russ Cox (rsc) explains why: http://research.swtch.com/macpprof • And he provides kernel patch: http://godoc.org/code.google.com/p/rsc/ cmd/pprof_mac_fix • Activity monitor or Instruments provides another sampling profiler
pprof command • Included in google perftools • ~Go 1.3 bundles CLI (perl) • Go 1.4 has reimplement it with Go!
Embed net/http/pprof import ( “net/http” _ “net/http/pprof” ) func main() { go http.ListenAndServe(“:5000”, nil) }
Let’s use it • top • web (svg) • list, disasm, and weblist
Random notes about Go’s performance
Things makes Go slower • GC • memcpy • function call
GC • GODEBUG=gctrace=1 • heap profile with pprof • sync.Pool • Provide size hint for slice and map. • make([]int, 0, hint) • make(map[int]int, hint) • GOGC=400
memcpy • Choose carefully string or []byte • Especially for constants
function call • Go’s calling convention is much slower than C. • No pass by register • All register are volatile • hook for runtime. • But Go has inlining • small leaf function may be inlined • Not inlined func may not have hook, too.
compare • calls1.go http://play.golang.org/p/N8oz-eyFII • calls2.go http://play.golang.org/p/s_Uv0vWirZ 13022001ns vs 10698225ns
What I hope Go 1.5~ will have • Concurrent GC will be come in Go 1.5! • But more speed is welcome anytime • fastcall for leaf function? • pass by register • volatile register

More Related Content

Why my Go program is slow?

  • 1. Why my Go program is slow? GoCon 2014 autumn, INADA Naoki
  • 2. INADA Naoki • @methane • KLab Inc. • Python • mysqlclient, PyMySQL, msgpack
  • 3. And I’m gopher • I love Go. It’s more Pythonic than Python! • Explicit is better than implicit. • Simple is better than complex. • Readability counts. • There should be one-- and preferably only one --obvious way to do it. • Although that way may not be obvious at first unless you're Dutch (Gopher).
  • 4. Today • CPU profiling with pprof • Random notes about Go’s performance. • What I hope Go 1.5~ will have
  • 6. runtime/pprof • Sampling profiler. • net/http/pprof provides HTTP API for pprof • github.com/davecheney/profile provice easy API for CLI program.
  • 7. NOTE for Mac • Go’s CPU profiling is not work on Mac • Russ Cox (rsc) explains why: http://research.swtch.com/macpprof • And he provides kernel patch: http://godoc.org/code.google.com/p/rsc/ cmd/pprof_mac_fix • Activity monitor or Instruments provides another sampling profiler
  • 8. pprof command • Included in google perftools • ~Go 1.3 bundles CLI (perl) • Go 1.4 has reimplement it with Go!
  • 9. Embed net/http/pprof import ( “net/http” _ “net/http/pprof” ) func main() { go http.ListenAndServe(“:5000”, nil) }
  • 10. Let’s use it • top • web (svg) • list, disasm, and weblist
  • 11. Random notes about Go’s performance
  • 12. Things makes Go slower • GC • memcpy • function call
  • 13. GC • GODEBUG=gctrace=1 • heap profile with pprof • sync.Pool • Provide size hint for slice and map. • make([]int, 0, hint) • make(map[int]int, hint) • GOGC=400
  • 14. memcpy • Choose carefully string or []byte • Especially for constants
  • 15. function call • Go’s calling convention is much slower than C. • No pass by register • All register are volatile • hook for runtime. • But Go has inlining • small leaf function may be inlined • Not inlined func may not have hook, too.
  • 16. compare • calls1.go http://play.golang.org/p/N8oz-eyFII • calls2.go http://play.golang.org/p/s_Uv0vWirZ 13022001ns vs 10698225ns
  • 17. What I hope Go 1.5~ will have • Concurrent GC will be come in Go 1.5! • But more speed is welcome anytime • fastcall for leaf function? • pass by register • volatile register