ABP1 validate error with OS X partners

M. Schmitt
M. Schmitt
Joined: 27 Jun 05
Posts: 478
Credit: 15,872,262
RAC: 0
Topic 194622

My host AMD X2 running Linux got a wingman running OS X and the validator found a difference in the results and so sent out the task to another host running OS X too. As you can expect the result of my host is invalid.
In general this is not a big deal, but I remember some MAC users complaining about invalid results with the current ABP1 app.
I also recognised the OS X app is by far the fastest ABP1 app at the moment.
Is it possible that the OS X app is compiled with SSE2 support, because afaik this is the default compiler swich on all Intel MACs? In this case this might cause the differences in the results that the validator is finding.

Michael

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,522
Credit: 694,490,872
RAC: 112,986

ABP1 validate error with OS X partners

Quote:

My host AMD X2 running Linux got a wingman running OS X and the validator found a difference in the results and so sent out the task to another host running OS X too. As you can expect the result of my host is invalid.
In general this is not a big deal, but I remember some MAC users complaining about invalid results with the current ABP1 app.
I also recognised the OS X app is by far the fastest ABP1 app at the moment.
Is it possible that the OS X app is compiled with SSE2 support, because afaik this is the default compiler swich on all Intel MACs? In this case this might cause the differences in the results that the validator is finding.

Michael

Hi Michael,

Yes, the Mac OS app uses SSE2 support which causes the higher performance and most likely also the occasional (very rare) validation problems.

What gets sent back to the server is a list of the top N (100?) candidates for signals sorted by some metric of significance. If the values differ even by the tiniest difference, it's possible with some bad luck that there will be differences in the ranking or some candidates make it to the list on one host but not on the other.

For Mac users, this should be offset by the higher performance. For others that happen to get two Macs as wingmen "against" them (which is very rare).... sorry :-(

Bikeman

Martin Ryba
Martin Ryba
Joined: 9 Apr 09
Posts: 48
Credit: 158,802,371
RAC: 41,678

Hmmm, since I imagine over

Hmmm, since I imagine over 95% of the Windows PCs running ABP1 likely have SSE2, can't we run that too? I know the LIGO apps automatically detect the SSE(2?) support and launch the correct child application. Apparently that's on your "to do" list for the ABP1 app. Meanwhile, I wouldn't mind doing what I do on Milkyway which is to run a custom app which has SSE2 support. It's a bit more maintenance work (I have to upgrade the app by hand instead of relying on BOINC) but I'd like the higher processing speed. My 8 year old Dell is slow enough already.

"Better is the enemy of the good." - Voltaire (should be memorized by every requirements lead)

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,522
Credit: 694,490,872
RAC: 112,986

Hi all! A switcher could

Hi all!

A switcher could be an option, but it's not without complications (the notorious "too many exits" error on Windows might be related to this IIRC).

Anyway, while there are still some Pentium IIIs and Athlon XPs around, I wonder whether it would at least be safe to make the CUDA version SSE2 only. The CUDA app doesn't run on most on-main-board GPUs, just on regular cards, and I think there are only PCI Express cards for CUDA GPUs. I think the number of PCs with PCI Express bus but equipped with pre-SSE2-CPUs must be extremely small, right?

Bikeman

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,143
Credit: 2,925,340,747
RAC: 831,081

RE: A switcher could be an

Message 95460 in response to message 95459

Quote:
A switcher could be an option, but it's not without complications (the notorious "too many exits" error on Windows might be related to this IIRC).


????

AFAIK, the LIGO apps have what might more realistically be called a "launcher", not a "switcher": it detects the CPU capabilities, and selects the most appropriate worker application. From that point forward, it's a single app that performs all aspects of the computation - I don't see why one app should cause more exits than another.

And I'm not convinced by the attribution of the validation failures to the use of SSE2, either. At SETI, the third-party optimisers offer separate apps for SSE2, SSE3, SSSE3, and SSE4.1. All have been carefully bench-tested, and validate against all other versions (including stock applications with no SIMD optimisation at all). OTOH, there were problems with an early Linux AP optimisation, which failed cross-validation: I think that was eventually tracked down to injudicious use of compiler switches - once they were corrected, the app validated correctly. It would take a bit of research, but I could track down the references if you're interested.

Bottom line: the apps should validate against each other. If they don't, it's a bug - probably in compilation - which could and should be tracked down.

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1,079
Credit: 341,280
RAC: 0

RE: ...and I think there

Message 95461 in response to message 95459

Quote:
...and I think there are only PCI Express cards for CUDA GPUs.


Nope, there are CUDA-capable cards for PCI slots (without express :-), used to "pimp" older PCs.

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,522
Credit: 694,490,872
RAC: 112,986

RE: RE: ...and I think

Message 95462 in response to message 95461

Quote:
Quote:
...and I think there are only PCI Express cards for CUDA GPUs.

Nope, there are CUDA-capable cards for PCI slots (without express :-), used to "pimp" older PCs.

Gruß,
Gundolf

That's interesting, I considered that once but never found one at that time. Any reviews available/recommendations?

Bikeman

Gundolf Jahn
Gundolf Jahn
Joined: 1 Mar 05
Posts: 1,079
Credit: 341,280
RAC: 0

RE: RE: Nope, there are

Message 95463 in response to message 95462

Quote:
Quote:
Nope, there are CUDA-capable cards for PCI slots (without express :-), used to "pimp" older PCs.

That's interesting, I considered that once but never found one at that time. Any reviews available/recommendations?

Bikeman


As always, I can only refer to SETI threads I've read, for lack of the respective hardware at my place :-)
http://setiathome.berkeley.edu/forum_thread.php?id=56170

Gruß,
Gundolf

Computer sind nicht alles im Leben. (Kleiner Scherz)

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,522
Credit: 694,490,872
RAC: 112,986

Thanks for the

Thanks for the link!
Cheers
Bikeman

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.