gctuner

gctuner

Introduction

Inspired by How We Saved 70K Cores Across 30 Mission-Critical Services (Large-Scale, Semi-Automated Go GC Tuning @Uber) .

 _______________  => limit: host/cgroup memory hard limit
|               |
|---------------| => gc_trigger: heap_live + heap_live * GCPercent / 100
|               |
|---------------|
|   heap_live   |
|_______________|

Go runtime trigger GC when hit gc_trigger which affected by GCPercent and heap_live.

Assuming that we have stable traffic, our application will always have 100 MB live heap, so the runtime will trigger GC once heap hits 200 MB(by default GOGC=100). The heap size will be changed like: 100MB => 200MB => 100MB => 200MB => ....

But in real production, our application may have 4 GB memory resources, so no need to GC so frequently.

The gctuner helps to change the GOGC(GCPercent) dynamically at runtime, set the appropriate GCPercent according to current memory usage.

How it works

 _______________  => limit: host/cgroup memory hard limit
|               |
|---------------| => threshold: increase GCPercent when gc_trigger < threshold
|               |
|---------------| => gc_trigger: heap_live + heap_live * GCPercent / 100
|               |
|---------------|
|   heap_live   |
|_______________|

threshold = inuse + inuse * (gcPercent / 100)
=> gcPercent = (threshold - inuse) / inuse * 100

if threshold < 2*inuse, so gcPercent < 100, and GC positively to avoid OOM
if threshold > 2*inuse, so gcPercent > 100, and GC negatively to reduce GC times

Usage

The recommended threshold is 70% of the memory limit.

// Get mem limit from the host machine or cgroup file.
limit := 4 * 1024 * 1024 * 1024
threshold := limit * 0.7

gctuner.Tuning(threshold)

// Friendly input
gctuner.TuningWithFromHuman("1g")

// Auto
// There may be problems with multiple services in one pod.
gctuner.TuningWithAuto(false) // Is it a container? Incoming Boolean