harmony-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Alexey Varlamov" <alexey.v.varla...@gmail.com>
Subject Re: [drlvm][gc_gen] gc_gen is 2x slower in tests...
Date Fri, 18 May 2007 09:19:50 GMT
2007/5/16, Xiao-Feng Li <xiaofeng.li@gmail.com>:
> On 5/16/07, Alexey Varlamov <alexey.v.varlamov@gmail.com> wrote:
> > I checked my perceived changes in pre-commit testing time after
> > Harmony switched to gcv5, comparing with old gc_cc on the following
> > test:
> >
> > > sh build.sh -DBUILD_CFG=<cfg>  -Dtest.mode=jet test -Dtest.vmargs="-XX:vm.dlls=<gc>"
> > Some smoke tests (outofmemory.*, perf.SeveralThreads, etc)
> > intermittently fail/hang on gc_gen, so I've run each configuration 2-3
> > times. All machines are multicore (>= 2 CPUs) server boxes.
> > Ant-reported build time records are listed below:
> >
> > avg time in sec, gc_cc / gc_gen
> > --------------------------------------------
> > Win32 release:    340 / 520
> > Win32 debug:      620 / 1400
> > Linux32 release:  300 /  1200
> > Linux32 debug:    440 / 1000
> > Linux64 debug:    550 / 1300
> >
> > Indeed the time for testing is increased at least twice with gc_gen!
> >
> > Taking into account that gc_gen demonstrated very decent scores on
> > enterprise benchmarks, how to interpret these results? Heavy
> > initialization or client-unfriendly default settings of GC or smth
> > else?
>
> It's interesting. It will be greatly appreciated if you could do some
> initial investigations.

Xiao-Feng,
I'm not sure I can invest time into detailed investigation now, so
just created HARMONY-3907 to track this.

Thanks!
--
Alexey

Mime
View raw message