CPU Dhry/Whet benchmark decreasing when upgrading to 4.2x from prior versions
18 Feb 2005 4:55:52 UTC
Topic 187694
(moderation:
I've noticed since I went from 4.19 to 4.2x that BOINC seems to have changed the way it does its benchmarks.
AMD AthlonXP 3200+
4.19: 2064.52 MFlOps/Sec 4737.07 MIntOps/Sec
4.2x: 1998.40 MFlOps/Sec 3428.30 MintOps/Sec
Same system, Memory timings haven't changed from 2-2-2-5 @ DDR400. This pc is *not* overclocked in any way. Still running same applications this past month since the upgrade; no new games, browsers, drivers (short ATI Catylist 5.2), peripherals. I was able to get my old CPU bench stats from LHC@Home since they haven't yet started issuing WUs this year. Is anyone else who has migrated to the 4.2x devel-branch experienced this problem? It seems to have affected the amount of credit my pc is claiming.. my client has consistently claimed about 20-50 points less than other systems both faster and slower.
I'm going to tinker around with some things on my end - will post any results..
CPU Dhry/Whet benchmark decreasing when upgrading to 4.2x from p
)
> Is anyone else who has
> migrated to the 4.2x devel-branch experienced this problem? It seems to have
> affected the amount of credit my pc is claiming.. my client has consistently
> claimed about 20-50 points less than other systems both faster and slower.
It isn't a problem - it's a deliberate fix to stop the Windows compiler "cheating" and producing artificially inflated benchmarks. ;)
Be lucky,
Neil
> It isn't a problem - it's a
)
> It isn't a problem - it's a deliberate fix to stop the Windows compiler
> "cheating" and producing artificially inflated benchmarks. ;)
Hmm.. if so when E@H goes live it would only make fair sense to *require* 4.2x/4.6x branch clients else if I were a greedy stiffler (which I'm not) I'd use an older client just to "artificially inflate" my stats.. :o)
[Edit]
Come to think of it, the time predicted to complete a given workunit is also waaaay more accurate as well. Actual time is within ~5 minutes of predicted time...
[Edit]
I found this link to the boinc_dev mailing list as the only reference to someone else noticing this but no further documentation about it so far..
The benchmark bug has been
)
The benchmark bug has been corrected in the development version since November.
From BOINC CVS checkin notes in November 2004:
David 2 Nov 2004
- split Dhrystone source into 2 files (from Peter Smithson).
This supposedly prevents compiler optimizations on Win
that give inflated results
From BOINC_dev mailing List (Peter Smithson):
The original dhrystone source for revison 2.0 was split intentionally to
avoid over optimisation (see comments in original dhrystone test source
code). The BOINC uses merged source which causes the benchmark to run
unusually quickly when compiled with MSVC 7.1 due to better optimisation
possible. Presumably the optimiser is taking out most of the test and not
giving a result consistent with a real application. MSVC 6.0 did not do
this.
Link
I don't know if someone has tested to see if now the Windows and Linux Benchmarks are similar.
Professor Desty Nova
Researching Karma the Hard Way