search

jonhoo / volley

Volley is a benchmarking tool for measuring the performance of server networking stacks.
MIT License
123 stars 12 forks source link

Improve Go server performance #1

Closed tsenart closed 9 years ago

tsenart commented 9 years ago

This change set improves the Go server performance by reducing memory allocations and not relying on run-time reflection which the binary.{Read,Write} functions use under the hood.

jonhoo commented 9 years ago

Great, thanks! I'll test this change on Monday to see how much it impacts performance. I suspect this is not the major cause of slowdown for the Go code though. It seems like something more fundamental is wrong...

tsenart commented 9 years ago

We could easily profile one of these tests and see where the bottleneck is.

jonhoo commented 9 years ago

I profiled it a while ago, and the main culprit was epollwait. Unfortunately, that doesn't help us much in terms of improving the performance.. I'll check again with your merged changes on Monday.

jonhoo commented 9 years ago

Hmm, the performance improvement appears to be fairly marginal:

benchmark go vs tsenart/go

I'll merge the changes regardless, and post profiling results in servers/go/profile.svg.

tsenart commented 9 years ago

I have the theory that by increasing the number of OS threads that the Go scheduler can run on, we'll get better results.

The GOMAXPROCS variable limits the number of operating system threads that can execute user-level Go code simultaneously. There is no limit to the number of threads that can be blocked in system calls on behalf of Go code; those do not count against the GOMAXPROCS limit. This package's GOMAXPROCS function queries and changes the limit.

On Monday, June 15, 2015, Jon Gjengset notifications@github.com wrote:

Merged #1 https://github.com/jonhoo/volley/pull/1.

— Reply to this email directly or view it on GitHub https://github.com/jonhoo/volley/pull/1#event-331331155.

Sent from Gmail Mobile

jonhoo commented 9 years ago

I'm not sure what you mean. Since there is no limit to the number of threads Go can have blocking in system calls at any given point in time, Go should presumably manage this on its own? The best we can do is set GOMAXPROCS to an appropriate value. Are you proposing we increase the value of GOMAXPROCS further?

tsenart commented 9 years ago

Yes, that's what I'm proposing, for the sake of experimentation. It'd also be nice to know what version of Go is being used.

jonhoo commented 9 years ago 
$ go version
go version go1.4.1 linux/amd64

I'll run a benchmark with a higher GOMAXPROCS and see if it makes a difference. Could you open a ticket for this so we can continue the discussion there? Please also link back to this PR from the ticket.

tsenart commented 9 years ago

I'll do that tomorrow as I'm limited to a mobile phone until then.

On Monday, June 15, 2015, Jon Gjengset notifications@github.com wrote:

$ go version go version go1.4.1 linux/amd64

I'll run a benchmark with a higher GOMAXPROCS and see if it makes a difference. Could you open a ticket for this so we can continue the discussion there? Please also link back to this PR from the ticket.

— Reply to this email directly or view it on GitHub https://github.com/jonhoo/volley/pull/1#issuecomment-112195877.

Sent from Gmail Mobile

jonhoo commented 9 years ago

Never mind, I filed it as #2.