net/http/internal/http2: pre-allocate header map in server's newWriterAndRequest

[kidnamedcudilovedrake-dot]

In src/net/http/internal/http2/server.go, the function newWriterAndRequest constructs a new request and its header map when parsing an incoming HTTP/2 request frame.
Currently, the map is created using make(Header) without specifying an initial capacity:

header := make(Header)
rp.Header = header
for _, hf := range f.RegularFields() {
    header.Add(sc.canonicalHeader(hf.Name), hf.Value)
}

[kidnamedcudilovedrake-dot]

Since the number of regular header fields (f.RegularFields()) is known beforehand, creating the map without capacity leads to dynamic map resizing and re-allocations as headers are added. This causes unnecessary CPU overhead and heap allocation pressure.

We can optimize this by pre-allocating the map with make(Header, len(regularFields)) and caching the RegularFields() slice, similar to how it is done in transport.go for the client.

I have uploaded a CL for this optimization: https://go-review.googlesource.com/c/go/+/794260

[gopherbot]

Change https://go.dev/cl/794260 mentions this issue: net/http/internal/http2: pre-size request header map in newWriterAndRequest

[gabyhelp]

Related Issues

• net/http: Add map preallocate #57121 (closed)
• x/net/http2: reduce allocations of (*clientConnReadLoop).handleReponse by 10% #37853 (closed)

Related Code Changes

• net/http/internal/http2: pre-size request header map in newWriterAndRequest
• net/http: Add map preallocate

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[kidnamedcudilovedrake-dot]

Here are the benchmark results comparing the base code and the optimization using BenchmarkClientRequestHeaders:

goarch: amd64
pkg: net/http/internal/http2
cpu: Intel(R) Core(TM) i5-14600KF

                                     │ bench_req_unopt.txt │          bench_req_opt.txt           │
                                     │       sec/op        │    sec/op     vs base                │
ClientRequestHeaders/___0_Headers-20          56.56µ ± 73%   76.21µ ± 32%        ~ (p=0.315 n=10)
ClientRequestHeaders/__10_Headers-20          57.03µ ±  3%   59.93µ ±  2%   +5.08% (p=0.004 n=10)
ClientRequestHeaders/_100_Headers-20         167.35µ ± 79%   87.32µ ±  4%  -47.82% (p=0.000 n=10)
ClientRequestHeaders/1000_Headers-20          829.4µ ±  6%   765.0µ ±  6%   -7.76% (p=0.001 n=10)
geomean                                       145.5µ         132.2µ         -9.14%

                                     │ bench_req_unopt.txt │          bench_req_opt.txt           │
                                     │        B/op         │     B/op      vs base                │
ClientRequestHeaders/___0_Headers-20          4.599Ki ± 0%   4.604Ki ± 0%   +0.11% (p=0.003 n=10)
ClientRequestHeaders/__10_Headers-20          6.275Ki ± 0%   5.927Ki ± 0%   -5.54% (p=0.000 n=10)
ClientRequestHeaders/_100_Headers-20          28.83Ki ± 0%   23.63Ki ± 0%  -18.03% (p=0.000 n=10)
ClientRequestHeaders/1000_Headers-20          321.7Ki ± 0%   228.3Ki ± 0%  -29.03% (p=0.000 n=10)
geomean                                       22.74Ki        19.59Ki       -13.88%

                                     │ bench_req_unopt.txt │          bench_req_opt.txt           │
                                     │      allocs/op      │  allocs/op   vs base                 │
ClientRequestHeaders/___0_Headers-20            50.00 ± 0%    50.00 ± 0%       ~ (p=1.000 n=10) ¹
ClientRequestHeaders/__10_Headers-20            74.00 ± 0%    73.00 ± 0%  -1.35% (p=0.000 n=10)
ClientRequestHeaders/_100_Headers-20            344.0 ± 0%    337.0 ± 0%  -2.03% (p=0.000 n=10)
ClientRequestHeaders/1000_Headers-20           4.066k ± 0%   4.048k ± 0%  -0.44% (p=0.000 n=10)
geomean                                         268.2         265.6       -0.96%
¹ all samples are equal

[dmitshur]

CC @neild, @nicholashusin.