This has me scratching my head a bit...
11 Feb 2005 20:44:08 UTC
Topic 187526
(moderation:
I've returned 2 WUs so far... In each case I've got an Athlon XP 1900+ that claimed lower credit (less computation time) then an Athlon MP of higher clock then mine...
http://einsteinathome.org/workunit/330921
http://einsteinathome.org/workunit/331092
I'm host 12051, and got
1105684 12051 10 Feb 2005 6:01:22 UTC 11 Feb 2005 20:31:49 UTC Over Success Done 28,923.59 66.09 pending
The other host is a dual CPU Athlon MP 2200+, but took longer to complete
1105682 3632 8 Feb 2005 13:06:05 UTC 9 Feb 2005 21:24:33 UTC Over Success Done 43,213.87 82.03 pending
The other unit is the same thing. Any ideas?
This has me scratching my head a bit...
)
Not looked into the results yet, but the claimed credit is computed from the CPU hours and the benchmark results. The latter vary a lot between client versions and operating systems, so I wouldn't worry too much.
BM
BM
> I've returned 2 WUs so
)
> I've returned 2 WUs so far... In each case I've got an Athlon XP 1900+ that
> claimed lower credit (less computation time) then an Athlon MP of higher clock
> then mine...
>
> http://einsteinathome.org/workunit/330921
> http://einsteinathome.org/workunit/331092
>
> I'm host 12051, and got
>
> 1105684 12051 10 Feb 2005 6:01:22 UTC 11 Feb 2005 20:31:49 UTC Over Success
> Done 28,923.59 66.09 pending
>
> The other host is a dual CPU Athlon MP 2200+, but took longer to complete
>
> 1105682 3632 8 Feb 2005 13:06:05 UTC 9 Feb 2005 21:24:33 UTC Over Success Done
> 43,213.87 82.03 pending
>
> The other unit is the same thing. Any ideas?
It's because the other host is Linux, and for reasons that we don't understand, our Linux code still doesn't run as fast as the Windows code. It seems that the Windows compilers to a better job. Or we haven't found the right compiler flags yet for Linux.
Bruce
Director, Einstein@Home
k, thx... Code efficiency
)
k, thx... Code efficiency could account for the seeming backward claim in credits given the CPU clocks...
Anyhow, they both validated and got credit, so I didn't have an early termination or anything...
> > I've returned 2 WUs so
)
> > I've returned 2 WUs so far... In each case I've got an Athlon XP 1900+
> that
> > claimed lower credit (less computation time) then an Athlon MP of higher
> clock
> > then mine...
> >
> > http://einsteinathome.org/workunit/330921
> > http://einsteinathome.org/workunit/331092
> >
> > I'm host 12051, and got
> >
> > 1105684 12051 10 Feb 2005 6:01:22 UTC 11 Feb 2005 20:31:49 UTC Over
> Success
> > Done 28,923.59 66.09 pending
> >
> > The other host is a dual CPU Athlon MP 2200+, but took longer to
> complete
> >
> > 1105682 3632 8 Feb 2005 13:06:05 UTC 9 Feb 2005 21:24:33 UTC Over Success
> Done
> > 43,213.87 82.03 pending
> >
> > The other unit is the same thing. Any ideas?
>
> It's because the other host is Linux, and for reasons that we don't
> understand, our Linux code still doesn't run as fast as the Windows code. It
> seems that the Windows compilers to a better job. Or we haven't found the
> right compiler flags yet for Linux.
>
> Bruce
There are optimizied versions of BOINC for Linux that will set your benchmarks closer to what they should be. Check out this thread on Seti that shows the better results for the optimized Linux version.
http://setiweb.ssl.berkeley.edu/forum_thread.php?id=11180
The site for the different BOINC versions is here
http://www.pperry.f2s.com/
> It's because the other host
)
> It's because the other host is Linux, and for reasons that we don't
> understand, our Linux code still doesn't run as fast as the Windows code. It
> seems that the Windows compilers to a better job. Or we haven't found the
> right compiler flags yet for Linux.
For the Boinc client, the following flags are about as good as you'll get:
-march=[cpu] -O3 -fomit-frame-pointer -funroll loops -ffast-math -fforce-addr -ftracer
where [cpu] is pentium2, i686, athlon, etc, as appropriate. It generally seems to have the least effect of all the flags so i586 would probably be OK for a general distribution.
For the work client, it is IMPERATIVE to remove the -ffast-math flag.
Be lucky,
Neil
> There are optimizied
)
> There are optimizied versions of BOINC for Linux that will set your benchmarks
> closer to what they should be.
If with "closer to what they should be." you mean closer to windows benchmark, then in reality they are higher then they should be. This is because all windows BOINC CC up to version 4.19 (this included) have a bug that gives higher benchmarks than it should be. Actually the standard Linux BOINC CC gives the "right" benchmarks.
This bug is corrected in the comming BOINC CC 4.2x series.
Professor Desty Nova
Researching Karma the Hard Way