New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
cmd/compile, x/build/cmd/relui: update compiler's PGO profile as part of rc1 relui flow #60234
Comments
We discussed this in the release team meeting today. The plan is to add this as a relui step, to be performed as part of creating rc1. We can also run the step as a one-off if necessary (e.g., after any major compiler refactoring). Ideally we can shrink the profile a little. Currently it's ~350K, which is just big enough to be a little annoying to update regularly in git. cc @golang/release |
We should discuss whether to try to get it done for this cycle. On the one hand it's not very necessary, but on the other it might be best to strike while the iron is hot. |
When I did it I put GOROOT in /tmp so the file paths are short. We could probably also trim cold nodes from the profile to reduce size. |
Change https://go.dev/cl/546337 mentions this issue: |
Since the default.pgo profile is collected, there has been a lot of development and refactoring in the compiler. It's time to refresh the compiler's PGO profile. As we are in the freeze, there will probably not be many major changes to the compiler, so this should be good for Go 1.22. Profile collected by running the cmd/compile/profile.sh script on the linux-amd64-perf gomote. Benchmark results on Linux/AMD64: │ nopgo.txt │ old.txt │ new.txt │ │ sec/op │ sec/op vs base │ sec/op vs base │ Template 127.4m ± 0% 125.3m ± 0% -1.65% (p=0.000 n=20) 123.5m ± 0% -3.07% (p=0.000 n=20) Unicode 105.94m ± 0% 100.77m ± 0% -4.87% (p=0.000 n=20) 99.26m ± 0% -6.30% (p=0.000 n=20) GoTypes 798.2m ± 1% 784.4m ± 0% -1.73% (p=0.000 n=20) 769.4m ± 1% -3.61% (p=0.000 n=20) Compiler 85.10m ± 0% 82.03m ± 1% -3.61% (p=0.000 n=20) 79.77m ± 1% -6.26% (p=0.000 n=20) SSA 6.054 ± 0% 5.945 ± 0% -1.81% (p=0.000 n=20) 5.811 ± 0% -4.03% (p=0.000 n=20) Flate 75.37m ± 1% 74.11m ± 0% -1.66% (p=0.000 n=20) 73.02m ± 1% -3.12% (p=0.000 n=20) GoParser 144.6m ± 1% 141.7m ± 1% -1.95% (p=0.000 n=20) 140.5m ± 1% -2.83% (p=0.000 n=20) Reflect 312.9m ± 1% 304.9m ± 1% -2.56% (p=0.000 n=20) 300.4m ± 0% -4.00% (p=0.000 n=20) Tar 115.8m ± 0% 113.6m ± 0% -1.88% (p=0.000 n=20) 111.9m ± 1% -3.33% (p=0.000 n=20) XML 145.9m ± 1% 143.6m ± 0% -1.58% (p=0.000 n=20) 141.3m ± 1% -3.15% (p=0.000 n=20) geomean 224.1m 218.9m -2.34% 215.2m -3.98% This also shows that a stale profile (the old one) still brings a speedup, but smaller, as expected. As before, the profile is transferrable. Benchmark results on Darwin/ARM64: │ nopgo.txt │ old.txt │ new.txt │ │ sec/op │ sec/op vs base │ sec/op vs base │ Template 71.31m ± 34% 68.97m ± 1% -3.28% (p=0.000 n=20) 68.66m ± 38% -3.71% (p=0.030 n=20) Unicode 58.70m ± 1% 55.64m ± 1% -5.20% (p=0.000 n=20) 53.05m ± 1% -9.61% (p=0.000 n=20) GoTypes 422.9m ± 0% 412.7m ± 0% -2.43% (p=0.000 n=20) 406.0m ± 1% -4.01% (p=0.000 n=20) Compiler 45.89m ± 1% 42.67m ± 2% -7.00% (p=0.000 n=20) 42.96m ± 1% -6.38% (p=0.000 n=20) SSA 3.190 ± 0% 3.090 ± 0% -3.15% (p=0.000 n=20) 3.007 ± 1% -5.74% (p=0.000 n=20) Flate 42.02m ± 1% 41.09m ± 1% -2.20% (p=0.000 n=20) 40.58m ± 1% -3.43% (p=0.000 n=20) GoParser 75.75m ± 1% 73.84m ± 1% -2.52% (p=0.000 n=20) 72.66m ± 1% -4.08% (p=0.000 n=20) Reflect 163.7m ± 1% 158.1m ± 1% -3.39% (p=0.000 n=20) 155.3m ± 1% -5.10% (p=0.000 n=20) Tar 62.03m ± 1% 60.15m ± 0% -3.02% (p=0.000 n=20) 59.74m ± 1% -3.69% (p=0.000 n=20) XML 79.38m ± 1% 77.32m ± 1% -2.59% (p=0.000 n=20) 75.05m ± 2% -5.45% (p=0.000 n=20) geomean 120.9m 116.6m -3.49% 114.6m -5.14% For #60234. Change-Id: I4ead6158f835c2e69c0f51ea24d044b82a7320cb Reviewed-on: https://go-review.googlesource.com/c/go/+/546337 Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org> Reviewed-by: Michael Pratt <mpratt@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Since the default.pgo profile is collected, there has been a lot of development and refactoring in the compiler. It's time to refresh the compiler's PGO profile. As we are in the freeze, there will probably not be many major changes to the compiler, so this should be good for Go 1.22. Profile collected by running the cmd/compile/profile.sh script on the linux-amd64-perf gomote. Benchmark results on Linux/AMD64: │ nopgo.txt │ old.txt │ new.txt │ │ sec/op │ sec/op vs base │ sec/op vs base │ Template 127.4m ± 0% 125.3m ± 0% -1.65% (p=0.000 n=20) 123.5m ± 0% -3.07% (p=0.000 n=20) Unicode 105.94m ± 0% 100.77m ± 0% -4.87% (p=0.000 n=20) 99.26m ± 0% -6.30% (p=0.000 n=20) GoTypes 798.2m ± 1% 784.4m ± 0% -1.73% (p=0.000 n=20) 769.4m ± 1% -3.61% (p=0.000 n=20) Compiler 85.10m ± 0% 82.03m ± 1% -3.61% (p=0.000 n=20) 79.77m ± 1% -6.26% (p=0.000 n=20) SSA 6.054 ± 0% 5.945 ± 0% -1.81% (p=0.000 n=20) 5.811 ± 0% -4.03% (p=0.000 n=20) Flate 75.37m ± 1% 74.11m ± 0% -1.66% (p=0.000 n=20) 73.02m ± 1% -3.12% (p=0.000 n=20) GoParser 144.6m ± 1% 141.7m ± 1% -1.95% (p=0.000 n=20) 140.5m ± 1% -2.83% (p=0.000 n=20) Reflect 312.9m ± 1% 304.9m ± 1% -2.56% (p=0.000 n=20) 300.4m ± 0% -4.00% (p=0.000 n=20) Tar 115.8m ± 0% 113.6m ± 0% -1.88% (p=0.000 n=20) 111.9m ± 1% -3.33% (p=0.000 n=20) XML 145.9m ± 1% 143.6m ± 0% -1.58% (p=0.000 n=20) 141.3m ± 1% -3.15% (p=0.000 n=20) geomean 224.1m 218.9m -2.34% 215.2m -3.98% This also shows that a stale profile (the old one) still brings a speedup, but smaller, as expected. As before, the profile is transferrable. Benchmark results on Darwin/ARM64: │ nopgo.txt │ old.txt │ new.txt │ │ sec/op │ sec/op vs base │ sec/op vs base │ Template 71.31m ± 34% 68.97m ± 1% -3.28% (p=0.000 n=20) 68.66m ± 38% -3.71% (p=0.030 n=20) Unicode 58.70m ± 1% 55.64m ± 1% -5.20% (p=0.000 n=20) 53.05m ± 1% -9.61% (p=0.000 n=20) GoTypes 422.9m ± 0% 412.7m ± 0% -2.43% (p=0.000 n=20) 406.0m ± 1% -4.01% (p=0.000 n=20) Compiler 45.89m ± 1% 42.67m ± 2% -7.00% (p=0.000 n=20) 42.96m ± 1% -6.38% (p=0.000 n=20) SSA 3.190 ± 0% 3.090 ± 0% -3.15% (p=0.000 n=20) 3.007 ± 1% -5.74% (p=0.000 n=20) Flate 42.02m ± 1% 41.09m ± 1% -2.20% (p=0.000 n=20) 40.58m ± 1% -3.43% (p=0.000 n=20) GoParser 75.75m ± 1% 73.84m ± 1% -2.52% (p=0.000 n=20) 72.66m ± 1% -4.08% (p=0.000 n=20) Reflect 163.7m ± 1% 158.1m ± 1% -3.39% (p=0.000 n=20) 155.3m ± 1% -5.10% (p=0.000 n=20) Tar 62.03m ± 1% 60.15m ± 0% -3.02% (p=0.000 n=20) 59.74m ± 1% -3.69% (p=0.000 n=20) XML 79.38m ± 1% 77.32m ± 1% -2.59% (p=0.000 n=20) 75.05m ± 2% -5.45% (p=0.000 n=20) geomean 120.9m 116.6m -3.49% 114.6m -5.14% For golang#60234. Change-Id: I4ead6158f835c2e69c0f51ea24d044b82a7320cb Reviewed-on: https://go-review.googlesource.com/c/go/+/546337 Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org> Reviewed-by: Michael Pratt <mpratt@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
CL https://golang.org/cl/451292 adds a default profile of the compiler, so we can build the compiler with PGO. As the compiler's source code evolves, we may want to periodically update the profile.
cc @prattmic @aclements
The text was updated successfully, but these errors were encountered: