runtime: performance regression on FreeBSD

[gopherbot]

by mark.vonminden:

What steps will reproduce the problem?

Run Dave Cheney's autobench tools on FreeBSD/amd64. See:
https://github.com/davecheney/autobench

Which compiler are you using (5g, 6g, 8g, gccgo)?

6g

Which operating system are you using?

FreeBSD

Which version are you using?  (run 'go version')

go version go1.1 freebsd/amd64

Please provide any additional information below.

Seeing a performance regression in the net/http package on FreeBSD. I was able to detect
it using Dave Cheney's autobench tools, which compares 1.03 and 1.1 using the built-in
benchmark tests (https://github.com/davecheney/autobench).

In particular, there seems to be a regression in the client/server performance.

Here is a snippet from the attached benchcmp output (comparing 1.03 and 1.1 on FreeBSD):

benchmark                         old ns/op    new ns/op    delta
BenchmarkHTTPClientServer            144770      8017199  +5437.89%
BenchmarkClientServer                156708      8017160  +5015.99%
BenchmarkClientServerParallel4       185736      2019450  +987.27%
BenchmarkClientServerParallel64      172426       209856  +21.71%
BenchmarkServer                      148844      4050047  +2621.00%

I also compared the http benchmarks on Linux and FreeBSD for Go 1.1 and saw that FreeBSD
was significantly slower. I've attached benchcmp output for that as well.

Attachments:

1. freebsd-amd64-d5666bad617d-vs-e570c2daeaca.txt (12181 bytes)
2. go1.1-http-bench-linux-vs-freebsd.txt (3961 bytes)

[robpike]

Comment 1:

Without looking at anything but the numbers, I can say that a delta like that is almost
certainly due to an unfortunate location for a stack segment boundary. We're working on
ways to avoid that in future, by arranging stacks in a different way.

[bradfitz]

Comment 3:

Seems more likely related to networking performance.  net/http's tests stress that
pretty hard where other tests do not.

Labels changed: added os-freebsd, removed priority-triage.

Status changed to Accepted.

[mikioh]

Comment 4:

On freebsd/amd64 virtual machine,
1.0.3 vs. 1.1.2:
BenchmarkClientServer                           97216     81947450  +84194.20%
BenchmarkClientServerParallel4                  97311     20397772  +20861.42%
BenchmarkClientServerParallel64                 97443      1638817  +1581.82%
BenchmarkServer                                 98145     41155672  +41833.54%
1.0.3 vs. tip:
BenchmarkClientServer                           97216        59846  -38.44%
BenchmarkClientServerParallel4                  97311        59673  -38.68%
BenchmarkClientServerParallel64                 97443        53200  -45.40%
BenchmarkServer                                 98145        89862   -8.44%
1.1.2 vs. tip:
BenchmarkClientServer                        81947450        59846  -99.93%
BenchmarkClientServerParallel4               20397772        59673  -99.71%
BenchmarkClientServerParallel64               1638817        53200  -96.75%
BenchmarkServer                              41155672        89862  -99.78%
Looks not so bad.

Status changed to Fixed.

Attachments:

1. freebsd-amd64-1.0.3-1.1.2.txt (4574 bytes)
2. freebsd-amd64-1.0.3-tip.txt (4563 bytes)
3. freebsd-amd64-1.1.2-tip.txt (4563 bytes)

[davecheney]

Comment 5:

Nice

[mikioh]

Comment 6:

This issue was closed by revision cb3b292.