mirror of
https://github.com/kubernetes/kubernetes.git
synced 2026-11-06 05:17:08 +00:00
Patrick Ohly
3e1a76e344
This is a tool generated by AI which refuted an earlier theory that increased stop-the-world events slowed down scheduling performance. The opposite seems to be true: reducing memory pressure reduced the impact of garbage collection.
traceparse
traceparse summarizes GC activity from Go execution trace files (.trace.out)
produced by go test -trace or the -perf-trace flag in scheduler_perf.
Usage
go -C hack/tools build -o /tmp/traceparse ./traceparse
/tmp/traceparse <trace-file> [<trace-file> ...]
Output
For each file the tool prints a table of named range events sorted by total wall-clock duration, followed by the average live heap size:
=== path/to/foo-trace.out ===
range count total duration
----- ----- --------------
GC concurrent mark phase 47 5.662s
GC mark assist 14294 743.936ms
GC incremental sweep 161077 321.769ms
stop-the-world (GC sweep termination) 47 7.568ms
stop-the-world (read mem stats) 42 5.446ms
stop-the-world (GC mark termination) 47 2.221ms
stop-the-world (start trace) 1 1.792µs
avg live heap: 796 MB (3827825 samples)
Key metrics for scheduler_perf analysis
| Range | What it tells you |
|---|---|
GC concurrent mark phase |
Total GC work; fewer cycles / less time = less allocation pressure |
GC mark assist |
Most sensitive indicator: goroutines interrupted mid-scheduling to help mark because allocation outpaces background GC; high counts mean hot-path allocations are limiting throughput |
stop-the-world (GC sweep termination) |
STW pauses that stop all goroutines; directly affects scheduling tail latency |
stop-the-world (GC mark termination) |
Second STW pause per cycle; usually short |
avg live heap |
Working set size; drives GC trigger frequency |