TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

Profile-Guided Optimization (PGO) for Go Services

2 posts · May 2023 – Jun 2024

What this trail covers

Grab has adopted Profile-Guided Optimization (PGO) for its Go services to improve performance and resource utilization. This involves collecting CPU profile data from production services and feeding it back into the Go compiler during the build process. Initial experiments on services like TalariaDB have shown significant reductions in CPU usage (around 10%), memory usage (around 30%), and persistent volume usage (around 38%). The effectiveness of PGO varies across services, with some requiring more effort than the gains provided, such as in the Catwalk service. Future improvements to `go-build` and PGO support for monorepo services are expected to drive greater adoption.

All Grab trails
  1. Profile-guided optimisation (PGO) on Grab servicesThis post details the adoption of Profile-Guided Optimization (PGO) for Grab's Go services. It explains how to enable PGO by profiling services to obtain pprof files and then incorporating these profiles into the Go build process. The post presents case studies for TalariaDB, demonstrating significant reductions in CPU (10%), memory (30%), and volume usage (38%), along with an increase in ingested event count per CPU. It also discusses an experiment with the Catwalk service, where the effort to enable PGO outweighed the observed 5% improvement. The post highlights the challenges of enabling PGO for monorepo services and concludes that its applicability varies across services.
  2. 2.3x faster using the Go plugin to replace Lua virtual machineThis post details the replacement of Lua VM with Go plugins for custom data processing in Talaria, achieving a 2.3x performance improvement and reduced memory usage. It includes benchmarks comparing Lua VM, Go native functions, and Go plugin calls, and explains the integration of Go plugins as a loader within Talaria's architecture.