Golang GC Tuning Guide: GOMEMLIMIT Configuration and Heap Allocation Optimization Strategies

When performance degradation due to GC occurs in Golang, you must determine the need for tuning via Profiling. If runtime.findObject or runtime.greyObject appear at the top of profiling results, it is time for GC tuning.

Use GOMEMLIMIT Instead of GOGC

The core of GC tuning is minimizing Stop the World (STW) time. The existing GOGC is based on heap growth rate, making value configuration complex and posing OOM risks. However, specifying a memory upper limit with GOMEMLIMIT induces GC to run as late as possible, thereby minimizing STW. Since GOMEMLIMIT is a SoftLimit, consider that it may use slightly more memory than the configured value.

Heap Allocation Analysis and Tuning Points

In Heap Profile, the alloc_space and alloc_objects metrics are more important than inuse. This is because large allocation amounts accelerate GC Cycles and consume CPU for Marking tasks. You should identify functions with high object allocation using a Flame graph and determine the causes of variables escaping to the Heap with the -gcflags='-m -m' option.

Practical Optimization Tips

Specific methods to reduce Heap allocation are as follows:

• Beware of non-typed arguments: Using fmt.Println or interface{} causes variables to be allocated on the Heap. When choosing a Logger, zap is recommended over log or slog as it keeps Primitive types on the Stack.
• Avoid Pointers: Using pointers unconditionally allocates to the Heap, so CallByValue is often more efficient.
• Constantize Slice Capacity: Specifying Capacity as a constant when creating a Slice ensures objects under 64KB are allocated on the Stack and excluded from GC targets.
• Utilize Pool: sync.Pool reduces Heap allocation but may consume CPU for management costs, so it should be compared via Benchmark.
• Use slices.SortFunc: sort.Slice creates Heap objects on every call, whereas slices.SortFunc has no allocations.

Finally, verify performance improvements through Benchmark, but it is important to stop at an appropriate level as excessive tuning harms readability.