Azul's Real-Time Profiling & Monitoring (RTPM) Tool
10Google Tech Talks
February, 13 2008
ABSTRACT
I present an unscripted live demo of RTPM, Azul's profiling
& monitoring tool (meaning: I'll take suggestions from the audience on
what buttons to push during the demo). JVM's routinely collect a lot
of very useful information internally and the Azul engineers were long
frustrated that this information was not readily available outside the
VM: with RTPM it now is.
Azul works with some of the largest Java programs around. We
routinely debug performance problems in programs with hundreds of
running threads (not just runnable threads) using garbage-collected
heaps with 100's of gigabytes. RTPM is a big selling point with our
customers, and our secret weapon for telling you what your program is
doing.
* Always-on (low overhead) everything
* All stats viewable from standard web browers
* Contended monitors, i.e. "hot locks" - and call stacks for
threads blocked on the lock
* Live counts of all sys-calls, i/o, file caches, RPC's, native calls
* Live stats on GC: heap size, GC cycle times, generation sizes,
app pause times, app allocation rates, etc
* Live stats of live stuff in the heap;
* Aggregate "points-to" on all heap objects. Suspect a leak? By
looking at the suspicious objects' points-to data you can quickly
zero-in on the root of the leak.
* Live hot-code profiling
* JIT asm dumps, annotated with live hardware perf counters
* Live stack traces of running threads (in 1-line-per Java call
format, or 1-line-per-Java-local)
* "Peek" into the live heap, much like a debugger can - straight
from live stack trace dumps
* Surf around the loaded class hierarchy, or the JIT'd code or all
HotSpot flags or...
* Works with 1000's of runnable threads
* Works with 100's of Gigabyte heaps
Speaker: Cliff Click (Azul Systems)