Performance optimization is an integral part of software development. Whether in web services, data processing systems, or real-time communications, good performance is critical. The Profile Guided Optimization (PGO) mechanism introduced by Golang from version 1.20 can help better optimize the performance of Go programs.
The previous article explained in detail the knowledge of collecting valid sample data and compiling. This article explains in detail the questions you may have about PGO.
Does PGO optimization work for packages in the Golang standard library?
The answer is yes, PGO is valid for the entire application, and all packages will be recompiled based on the PGO file, including packages in the standard library.
Does PGO optimization work for packages in dependent modules?
The answer is yes, PGO is valid for the entire application, all packages will be recompiled based on the PGO file, including packages in dependent modules. Therefore, special usage of dependent modules will affect the optimization effect.
Will programs optimized with PGO perform worse if the adopted PGO profile does not reflect typical usage scenarios?
probably not. A PGO configuration file that does not reflect typical usage scenarios will optimize the part of the logic code in the application that is not frequently run, but will not degrade the performance of the part of the code that is frequently run.
If an application needs to be built for different operating systems and different system architectures, can the same PGO configuration file be used?
The answer is yes, the format of the configuration file is the same for versions of different operating systems and different system architectures. For example, sample files collected from linux/arm64 systems can be used in the build process for windows/amd64 systems.
Does enabling PGO affect build times?
Enabling PGO may result in a significant increase in build time. Because PGO works on all packages, each package will be rebuilt the first time PGO is used. Since these builds are also cached, subsequent incremental builds using the same PGO profile do not require a full recompilation.
How does enabling PGO affect the size of the built binary?
Binaries built with PGO enabled will be slightly larger due to the extra function inlining.
Improvements to PGO in Go 1.21
PGO has been introduced as an experimental function since Go 1.20. It needs to be enabled by specifying the pgo parameter when compiling, such as -pgo=auto or -pgo=auto=xxx.pprof. Starting from Go 1.21, pgo has become an official function. go build will automatically detect whether there is a default.pgo file in the main directory of the program, and if detected, it will enable PGO to build.
Regarding the performance improvement of PGO, the official data given by Golang is that in Go 1.21, a set of benchmark tests of a representative Go program shows that using PGO to build can improve performance by about 2-7%. With Golang's continuous optimization of PGO and the continuous improvement of the representativeness of the collected PGO configuration files, it is believed that PGO's role in improving performance will become greater and greater.