[Gc] powerpc64 problems
hans.boehm at hp.com
Wed Nov 23 11:15:02 PST 2005
This seems to be getting slightly off subject, but since pauses usually
seem to be blamed on garbage collectors :-) ...
> What I actually found was the data scattered all over the
> place. Some of the slower tests seem to hog the CPU, Linux
> never seemed to schedule more than about 5 processes, and the
> min/max times were spread over a huge range .. ;(
It's actually not clear to me whether that's a bug or a feature. We're
starting to see machines with on the order of 100,000 cache lines. At
100 nsecs to reload one, that comes out to potentially 10 msecs to
reload the whole cache. If you do that once per 10 msec time slice, it
may start to look attractive to be a little less fair, and a little
smarter about cache reuse.
I do know that a fair amount of thought and benchmark testing went into
the linux scheduler.
More information about the Gc