No Java profiler is perfectly accurate, since JDK does not provide sufficient means to find where CPU time is exactly spent. Even "honest" profilers based on private HotSpot APIs will not tell the whole truth. Hardware counters and kernel functions could probably help, but unfortunately they are not aware of Java code at all. We will discuss several approaches to Java profiling: JVM TI, AsyncGetCallTrace, perf_events and Flame Graphs. Their principles and limitations will be considered. We will find a way to combine the advantages of all approaches together. Finally, we will see how Odnoklassniki performs full-stack profiling in production: from Java code down to Linux kernel.
Download presentation