Divining Profiler Accuracy: An Approach to Approximate Profiler Accuracy Through(…)
summary
Paper approximates profiler ground truth by injecting CPU‑wasting movs at machine‑code basic‑block boundaries and using hardware counters to tune the slowdown, targeting 50%/100%/150% target slowdowns. Experiments on 14 Java benchmarks with the Growl compiler and compiler replay show median slowdowns within ~5–10% of targets and 13/14 benchmarks keep method‑time ratios within 5%. Adding 5s slowdowns to blocks exposes large accuracy differences between Async‑profiler, JFR, JProfiler, YourKit and highlights missing debug info as a frequent failure mode.
AI-written summary · about
chapters (10)
- 0:00Profiler accuracy challenges
- 1:08Profiler result variability
- 2:36Slowdown for testing profilers
- 3:40Slowdown at machine code level
- 4:43Experimental setup
- 6:52Target slowdown accuracy
- 7:35Slowdown side effects
- 9:23Profiler accuracy with slowdown
- 11:31Conclusion and implications
- 12:15Question and answer session
related talks
01h 00mThe Basics of Profiling
51mProduction Profiling: What, Why and How
50mJPoint 2023 TechRadar: Performance Tools
01h 21mPerformance Really Matters
01h 17mThe Basics of Profiling
01h 03mApplication Performance Optimisation in Practice
44m.NET Performance Investigation
47mPerformance optimisation: how do I go about it?