hans.boehm at hp.com
Fri Aug 8 14:27:54 PDT 2008
Parallel marking is not enabled by default. With the autoconf build machinery, --enable-parallel-mark should enable it on platforms on which it is believed to work, e.g. Linux.
It's a small loss on a uniprocessor, perhaps small enough that we should turn it on by default. Based on limited experiments, it's a significant benefit for medium scale multiprocessors, e.g. current mainstream multicores. It's likely to scale less well as processor count goes much above 4,
but current versions seem to run OK with 16, the most I've tested. It would be useful to systematicslly run some real benchmarks.
The effect is to decrease mark times, up to 4 processors often roughly in proportion to processor count. That both improves throughput and reduces stop-the-world pause times.
Parallel collection and incremental collection currently do not work well together. Parallel marking support for Windows would be nice, and is probably just a small matter of programming.
> -----Original Message-----
> From: gc-bounces at napali.hpl.hp.com
> [mailto:gc-bounces at napali.hpl.hp.com] On Behalf Of Nicolas Cannasse
> Sent: Thursday, August 07, 2008 3:23 AM
> To: gc at napali.hpl.hp.com
> Subject: [Gc] Multicore
> Another question wrt. performances on multicore CPU.
> Is parallel marking available ?
> If it is, is it default, or does it need some specific
> compilation flags ? What are the gains expected of using it ?
> Gc mailing list
> Gc at linux.hpl.hp.com
More information about the Gc