net/http: ServeMux benchmark regression

[quasilyte]

There is notable difference in Go1.10.2 and tip BenchmarkServeMux results:

name        old time/op    new time/op    delta
ServeMux-4     127µs ± 1%     156µs ± 1%  +23.04%  (p=0.008 n=5+5)

On other machine:
ServeMux-4       143µs ±11%      165µs ± 7%    +15.44%  (p=0.008 n=5+5)

And another machine:
ServeMux-8    85.1µs ± 3%    95.0µs ± 6%  +11.61%  (p=0.008 n=5+5)

The most significant impact comes from additional synchronization added in 2fd1b52. The commit message lacks benchmarks info, so I think it's worthwhile to report regression anyway and decide whether it can be addressed or not (we can't just remove locking if it's needed there, but there can be other options).

There are other causes of slowdown as commit right behind 2fd1b52 is still slower (but the delta is more tolerable):

name        old time/op    new time/op    delta
ServeMux-4     127µs ± 1%     135µs ± 1%  +6.43%  (p=0.008 n=5+5)

[mvdan]

It seems like the commit you're pointing to was a race bugfix, so it's not the type of performance regression that can be easily fixed or reverted. Do you have any ideas on how to get the numbers back to 1.10 levels?

[quasilyte]

@mvdan, I'm not familiar with net/http at all, so I can't really propose any solution or evaluate submitted one. This is why I've added this to description:

The commit message lacks benchmarks info, so I think it's worthwhile to report regression anyway and decide whether it can be addressed or not (we can't just remove locking if it's needed there, but there can be other options).

I've reported it for completeness as it showed pretty high delta.
It's better to know the impact than not.
Usually, it's better when CLs that affects performance in this way includes benchstat info, but this one didn't, hence here is the missing link.

[OneOfOne]

@mvdan it's the defer, some goto hackary should make it pretty close to the old numbers.

If that's acceptable, I can push a PR sometime this week (or even better if you wanna do ;)).

[gopherbot]

Change https://golang.org/cl/113996 mentions this issue: net/http: avoid defer unlock in ServeMux.shouldRedirect

[gopherbot]

Change https://golang.org/cl/116378 mentions this issue: net/http: remove an allocation in ServeMux

[bradfitz]

@odeke-em, want to post an update on this now that 116378 is submitted? Or bump to Go 1.12 or Unplanned?

[odeke-em]

@bradfitz thanks for the ping.

So after CL 116378, we still have that regression from the newly introduced defer. I think we meant 116378 for #25482 instead :)

CL 113996 removes the defer that is the subject of this issue but it doesn't seem like you like its approach