On Wed, May 22, 2013 at 6:50 PM, Bill Neubauer wrote:
I was reading in the release notes for Go 1.1 about performance enhancements
through the coupling of the net package and the scheduler and I'm curious
how that works. Could someone please give me a pointer to the file(s) I
should be looking at to better understand that?
Hi,
It consists of 2 parts:
Improved scheduler with polling support:
https://codereview.appspot.com/7314062/https://codereview.appspot.com/7448048/and the network poller itself:
https://codereview.appspot.com/7579044/These two changes together improve network performance ~10x on linux:
BenchmarkTCPPersistent 81670 7782 -90.47%
BenchmarkTCPPersistent-2 26598 4808 -81.92%
BenchmarkTCPPersistent-4 15633 3674 -76.50%
BenchmarkTCPPersistent-8 18093 2407 -86.70%
BenchmarkTCPPersistent-16 17472 1875 -89.27%
BenchmarkTCPPersistent-32 7679 1637 -78.68%
Why it is faster:
- global mutex for poller state is eliminated
- edge-triggered epoll is used instead of level-triggered
- it uses heap-based runtime timers instead of home-grown list-based timers
- moreover, timers are set only once as opposed to setting them on
each read/write
- the polling is done by runtime worker threads at appropriate
moments, as opposed to by a separate goroutine at inappropriate
moments
- poller injects batches of newly runnable goroutines directly into scheduler
- blocking/unblocking of goroutines is more efficient in runtime and
no need to allocate additional channels for that
- other minor improvements
If you want to look at the current code:
src/pkg/net/fd_unix.go (Read and Write functions)
src/pkg/net/fd_poll_runtime.go (interface to runtime poller)
src/pkg/runtime/netpoll.c (poller code)
src/pkg/runtime/netpoll_epoll.c (linux part)
src/pkg/runtime/proc.c (scheduler integration, search for netpoll)
