tejal29/slowjam

NOTE: This is not an officially supported Google product

SlowJam is a two-part tool for analyzing function latency within Go programs.

stacklog - a library for sampling the stack of a Go program during runtime
slowjam - a binary for visualizing the latency from recorded stack samples

SlowJam excels at finding optimization opportunities in automation workflows, as they tend to block on command-line execution or remote resources, rather than CPU resources. This tool was created from a hackathon hosted by Google's Container DevEx team, with the goal of finding ways to reduce the start-up latency within minikube.

Features

Hybrid Gantt/Flamegraph visualizations
Automated sampling of all function calls
Trivial to integrate
Zero overhead when inactive, <1% overhead when activated

Screenshot

See example/minikube.html for example output.

Requirements

Go v1.14 or higher

Usage

Recording

Add this to the main() method of your program, or within any other function you wish to record stack samples for:

s := stacklog.MustStartFromEnv("STACKLOG_PATH")
defer s.Stop()

This will invoke the stack sampler if the STACKLOG_PATH environment is set, and will write the stack samples to that location. If you prefer greater control over the configuration, you can also use:

cfg := stacklog.Config{
  Path: "out.slog",
  Poll: 100 * time.Millisecond,
	Quiet: true
}
  
s, err := stacklog.Start(cfg)
defer s.Stop()

By default, this will poll the stack every 125ms.

Visualization

Install slowjam:

go install github.com/google/slowjam/cmd/slowjam

Analyze a stacklog using the interactive webserver:

slowjam -http localhost:8080 /path/to/stack.slog

To output a Gantt/Flamegraph chart to out.html:

slowjam -html out.html /path/to/stack.slog

To output a text summary to out.txt:

slowjam -html out.txt /path/to/stack.slog

Real World Example

Here's an example PR to integrate SlowJam analysis into minikube: minikube#8329.

With this PR, anyone can generate a slowjam profile:

env STACKLOG_PATH=minikube.slog minikube start

You can then convert the data to various forms, as per the examples/ directory:

slowjam --goroutines 1 --pprof example/minikube.pprof example/minikube.slog slowjam --html example/minikube.html example/minikube.slog

What minikube contributors discovered with these results were:

Functions which could obviously be run in parallel were executed in serial.
Functions which we expected to be fast (<1s) were slow (10s). In many cases we were able to remove or rewrite these functions to do less work.

The net result was a 2.5X reduction in start-up latency: from ~66 seconds to ~26 seconds.

tejal29 / slowjam