Hi,
I'm looking for a way to measure mutex contention in a Java app running
on Solaris 10. I started off by using the dvmti probes
monitor-contended-enter and monitor-contended-entered, but I found that
the act of observing these probes skewed my program performance to the
point that the measurements weren't useful.
I'm thinking of taking a lower-level approach to this. Is it correct to
assume that Java object monitors are going to map to a lwp_mutex at some
point in the call stack? I replaced my dvmti-based script with this:
syscall::lwp_mutex*:entry
{
self->ts[probefunc] = timestamp;
}
syscall::lwp_mutex*:return
/self->ts[probefunc] != 0/
{
duration = timestamp - self->ts[probefunc];
@times[ tid ] = quantize( duration );
self->ts[probefunc] = 0;
}
And the timings are pretty much what I'd expect, within an order of
magnitude. Can anybody suggest any improvements here, or am I totally
off-base?
Thanks,
Russ
_______________________________________________
dtrace-discuss mailing list
[email protected]
