net/http: consider reusing buffers replacing io.Copy calls with io.CopyBuffer

[artyom]

As been shown in #12450 by crude patch to make io.Copy reuse its underlying 32k buffer(s) workloads intensively (ab)using net/http can see a significant reduce in GC pressure. Since changes to io.Copy has already been proposed and rejected, I'd like to discuss the prospect of replacing at least some of the io.Copy calls in net/http by io.CopyBuffer with reused byte slices.

The net/http uses io.Copy the most in standard library, and I believe updating it to use recently introduced io.CopyBuffer is worth doing.

go (master) $ find src/net/http -name \*.go -type f ! -name \*_test.go -exec cat '{}' \+ | grep -wc io.Copy
19

At least some of those calls are not subject to usual io.ReaderFrom/io.WriterTo optimizations.

I'm ready to implement changes in the upcoming days and propose a CL.

[bradfitz]

I don't have an opinion until I see a concrete proposal. Which io.Copy calls?

[artyom]

@bradfitz one candidate is fallback path in net/http.response.ReadFrom, another is io.Copy used on liveSwitchReader when one makes use of http.CloseNotifier — httputil.ReverseProxy would gain most benefit from this.

benchmark                    old ns/op     new ns/op     delta
BenchmarkCloseNotified-4     309569        273466        -11.66%

benchmark                    old allocs     new allocs     delta
BenchmarkCloseNotified-4     50             48             -4.00%

benchmark                    old bytes     new bytes     delta
BenchmarkCloseNotified-4     36006         3089          -91.42%

[artyom]

Ref: https://go-review.googlesource.com/14177

[gopherbot]

CL https://golang.org/cl/14177 mentions this issue.