perf: optimize MarshalFrame

[mccutchen]

Context

Some small optimizations for MarshalFrame (and therefore also WriteFrame) that hopefully do not harm readability/clarity of code. See also #50 for ReadFrame optimizations, from which the we learned some of the tricks below.

Summary

Note

👉 With all of these combined, it appears that we've improved WriteFrame's throughput by ~27%!

Each commit represents a separate small optimization, summarized here:

1. small optimization for writing first two bytes to pre-allocated buffer in 73e023c
   • ~4.5% improvement in throughput (this was a bit surprising to me!)
   • benchmark results
   • (This also dropped the marshaledSize() helper in favor of pre-allocating worst-case buffer size, but that was restored in 00024a3 without meaningful impact on performance)
2. replace modulo division with bit shifting in 0bbc064
   • Why? As discovered in perf: optimize ReadFrame #50, i % N == i & (N - 1) when N is a power of 2 and bitwise AND is usually a single CPU cycle vs multiple cycles for modulo division
   • Additional ~4% improvement over (1)
   • benchmark results
3. use manually-optimized applyMask() from perf: optimize ReadFrame #50 in 502dcc3
   • Additional ~14% improvement over (2) 🔥
   • benchmark results
4. *use copy() instead of append() for writing mask and payload in b082698
   • Additional ~3.5% improvement over (3)
   • benchmark results

Final results comparing 6526127 (on main) vs b082698

Copied from workflow summary.

goos: linux
goarch: amd64
pkg: github.com/mccutchen/websocket
cpu: AMD EPYC 7763 64-Core Processor                
                  │ ./baseline/bench-results.txt │      ./head/bench-results.txt       │
                  │            sec/op            │   sec/op     vs base                │
ReadFrame/1KiB-4                     911.2n ± 1%   888.6n ± 3%   -2.49% (p=0.009 n=10)
ReadFrame/1MiB-4                     567.4µ ± 1%   542.2µ ± 1%   -4.44% (p=0.000 n=10)
WriteFrame/1KiB-4                   1329.5n ± 0%   944.6n ± 0%  -28.95% (p=0.000 n=10)
WriteFrame/1MiB-4                   1020.1µ ± 0%   584.4µ ± 1%  -42.71% (p=0.000 n=10)
geomean                              28.94µ        22.71µ       -21.52%

                  │ ./baseline/bench-results.txt │       ./head/bench-results.txt        │
                  │             B/s              │      B/s       vs base                │
ReadFrame/1KiB-4                    1.056Gi ± 1%    1.083Gi ± 3%   +2.55% (p=0.009 n=10)
ReadFrame/1MiB-4                    1.721Gi ± 1%    1.801Gi ± 1%   +4.65% (p=0.000 n=10)
WriteFrame/1KiB-4                   740.9Mi ± 0%   1042.9Mi ± 0%  +40.75% (p=0.000 n=10)
WriteFrame/1MiB-4                   980.4Mi ± 0%   1711.1Mi ± 1%  +74.54% (p=0.000 n=10)
geomean                             1.059Gi         1.350Gi       +27.42%

                  │ ./baseline/bench-results.txt │       ./head/bench-results.txt        │
                  │             B/op             │     B/op      vs base                 │
ReadFrame/1KiB-4                    1.164Ki ± 0%   1.164Ki ± 0%       ~ (p=1.000 n=10) ¹
ReadFrame/1MiB-4                    1.008Mi ± 0%   1.008Mi ± 0%  +0.00% (p=0.016 n=10)
WriteFrame/1KiB-4                   1.125Ki ± 0%   1.125Ki ± 0%       ~ (p=1.000 n=10) ¹
WriteFrame/1MiB-4                   1.008Mi ± 0%   1.008Mi ± 0%  -0.00% (p=0.001 n=10)
geomean                             34.37Ki        34.37Ki       -0.00%
¹ all samples are equal

                  │ ./baseline/bench-results.txt │      ./head/bench-results.txt       │
                  │          allocs/op           │ allocs/op   vs base                 │
ReadFrame/1KiB-4                      5.000 ± 0%   5.000 ± 0%       ~ (p=1.000 n=10) ¹
ReadFrame/1MiB-4                      5.000 ± 0%   5.000 ± 0%       ~ (p=1.000 n=10) ¹
WriteFrame/1KiB-4                     1.000 ± 0%   1.000 ± 0%       ~ (p=1.000 n=10) ¹
WriteFrame/1MiB-4                     1.000 ± 0%   1.000 ± 0%       ~ (p=1.000 n=10) ¹
geomean                               2.236        2.236       +0.00%
¹ all samples are equal

Note: All intermediate results are combined into the edit history of #55 (comment), which shows only the latest results.

[github-actions]

🔥 Run benchmarks comparing b082698 against main:

gh workflow run bench.yaml -f pr_number=55

Note: this comment will update with each new commit.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.94%. Comparing base (6526127) to head (b082698).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #55      +/-   ##
==========================================
+ Coverage   92.81%   92.94%   +0.13%     
==========================================
  Files           2        2              
  Lines         487      496       +9     
==========================================
+ Hits          452      461       +9     
  Misses         28       28              
  Partials        7        7              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

benchstats: 6526127...b082698

View full benchmark output on the workflow summary.

goos: linux
goarch: amd64
pkg: github.com/mccutchen/websocket
cpu: AMD EPYC 7763 64-Core Processor                
                  │ ./baseline/bench-results.txt │      ./head/bench-results.txt       │
                  │            sec/op            │   sec/op     vs base                │
ReadFrame/1KiB-4                     911.2n ± 1%   888.6n ± 3%   -2.49% (p=0.009 n=10)
ReadFrame/1MiB-4                     567.4µ ± 1%   542.2µ ± 1%   -4.44% (p=0.000 n=10)
WriteFrame/1KiB-4                   1329.5n ± 0%   944.6n ± 0%  -28.95% (p=0.000 n=10)
WriteFrame/1MiB-4                   1020.1µ ± 0%   584.4µ ± 1%  -42.71% (p=0.000 n=10)
geomean                              28.94µ        22.71µ       -21.52%

                  │ ./baseline/bench-results.txt │       ./head/bench-results.txt        │
                  │             B/s              │      B/s       vs base                │
ReadFrame/1KiB-4                    1.056Gi ± 1%    1.083Gi ± 3%   +2.55% (p=0.009 n=10)
ReadFrame/1MiB-4                    1.721Gi ± 1%    1.801Gi ± 1%   +4.65% (p=0.000 n=10)
WriteFrame/1KiB-4                   740.9Mi ± 0%   1042.9Mi ± 0%  +40.75% (p=0.000 n=10)
WriteFrame/1MiB-4                   980.4Mi ± 0%   1711.1Mi ± 1%  +74.54% (p=0.000 n=10)
geomean                             1.059Gi         1.350Gi       +27.42%

                  │ ./baseline/bench-results.txt │       ./head/bench-results.txt        │
                  │             B/op             │     B/op      vs base                 │
ReadFrame/1KiB-4                    1.164Ki ± 0%   1.164Ki ± 0%       ~ (p=1.000 n=10) ¹
ReadFrame/1MiB-4                    1.008Mi ± 0%   1.008Mi ± 0%  +0.00% (p=0.016 n=10)
WriteFrame/1KiB-4                   1.125Ki ± 0%   1.125Ki ± 0%       ~ (p=1.000 n=10) ¹
WriteFrame/1MiB-4                   1.008Mi ± 0%   1.008Mi ± 0%  -0.00% (p=0.001 n=10)
geomean                             34.37Ki        34.37Ki       -0.00%
¹ all samples are equal

                  │ ./baseline/bench-results.txt │      ./head/bench-results.txt       │
                  │          allocs/op           │ allocs/op   vs base                 │
ReadFrame/1KiB-4                      5.000 ± 0%   5.000 ± 0%       ~ (p=1.000 n=10) ¹
ReadFrame/1MiB-4                      5.000 ± 0%   5.000 ± 0%       ~ (p=1.000 n=10) ¹
WriteFrame/1KiB-4                     1.000 ± 0%   1.000 ± 0%       ~ (p=1.000 n=10) ¹
WriteFrame/1MiB-4                     1.000 ± 0%   1.000 ± 0%       ~ (p=1.000 n=10) ¹
geomean                               2.236        2.236       +0.00%
¹ all samples are equal