Performance Improvement with --no-aggregate-callees

[JansonLv]

Based on release practice, it has been demonstrated that using the --no—aggregate-callees flag is more likely to bring about performance improvements.

We have two web services:

one of these services involves a lot of I/O processing and has high latency. In this case, PGO did not result in any noticeable improvement.

For the other web service, which only has a small number of I/O requests, when we initially deployed it with the --aggregate-callees flag, there was almost no observable improvement. However, during the second deployment with the --no—aggregate-callees flag, the results were significantly better, with CPU load decreasing by about 15%.

This is a comparison before (--aggregate-callees) and after (--no—aggregate-callees) the deployment:

Theoretically, the flags --no-aggregate-callees and --aggregate-callees should not affect the PGO effect.

Go1.22.7
Pyroscope: 1.7

#3638

compilation log:
pgo file
no-agg.txt
agg.txt

[JansonLv]

@kolesnikovae ,hi ,I noticed alternative profile sources in the Go official documentation. Are these consistent with what you think? Can they help you?

[kolesnikovae]

Hi @JansonLv! Yes, the resulting profiles are consistent with the requirements.

Unfortunately, I didn't get a chance to look into the issue last week; I hope I'll get luckier this week.