CPU / GPU mixed processing problem


log in

Advanced search

Questions and Answers : GPU applications : CPU / GPU mixed processing problem

Author Message
CAHess-Den
Send message
Joined: 20 May 99
Posts: 21
Credit: 1,552,744
RAC: 457
United States
Message 920179 - Posted: 22 Jul 2009, 0:42:14 UTC

Short problem description: Machine ceases to process MP WUs in CPUs with GPUs enabled.

Setup: Quad-core Intel CPU w/ prefs = 2(50%) CPUs max, 50% CPU-time max set to be "always available"; Dual Nvidia 9600GT OCs defaulted to "when not in use," but sometimes force-run via local Preferences setting

Problem specifics:
I have USUALLY had four total WUs running: 2 in CPUs and 2 in my GPUs. This has worked fine for many weeks now. (Since at least BOINC v6.4.7 was released.)

I noticed that sometimes one of the CPUs will NOT pick up new work, despite loads of it in my queue. I determined that this was happening whenever a CPU was crunching on an AP WU: once the other CPU was free, it was no longer used to process any of the remaining available jobs.

This happened on at least four different occasions. (I wondered to myself whether this was intentional, but chalked it up to "mildly strange" and let it be. Until now. Read on.)

Now, however something different has happened: Once the AP/CPU job was completed, and the other MP/CPU completed... NO MORE CPU JOBS WILL RUN AT ALL. (Oh I've fiddled with settings - settings that didn't seem to make a differnce at all before, when they were running simultanously fairly well. Also, I have no more pending AP WUs.)

My queue of a dozen-odd MPs are only being processed by my GPUs now! [Which is a problem because, despite the fact that the GPUs run the jobs MUCH faster, they do not run NEARLY as often/regularly as my two available CPUs do!]

So now that NEITHER of my available CPUs are doing any work I though it was high time to raise this issue here.

Anomaly: The CUDA FAQ (#5) says that CPU and GPU work cannot be done simultaneously (or if out-of-date that they weren't at some time.) But Mine have been running fine since the first (working!) CUDA-enabled version of BOINC (6.4.7) came out! (And the work was accepted just fine!)
____________

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2632
Credit: 5,979,951
RAC: 3,827
Bulgaria
Message 920196 - Posted: 22 Jul 2009, 2:04:47 UTC - in response to Message 920179.
Last modified: 22 Jul 2009, 2:47:17 UTC

In the BOINC Manager, Tasks tab, Application column - what is the version of downloaded MB WUs - 6.03 (CPU) or 6.08 (GPU)?


You can try "ReSchedule" (I can't - I have no CUDA GPU):
http://setiathome.berkeley.edu/forum_thread.php?id=54447&nowrap=true#913445
http://setiathome.berkeley.edu/forum_thread.php?id=54491#914554
http://setiweb.ssl.berkeley.edu/forum_thread.php?id=54502&nowrap=true#914873


You can try also Lunatics' Unified Installer (the easy way to the Optimized SETI Applications):
http://setiathome.berkeley.edu/forum_thread.php?id=54288
http://lunatics.kwsn.net/index.php?module=Downloads;catd=9


To make the BOINC Manager to download more tasks change the preferences:
"Additional work buffer": 10 days
____________



- ALF - "Find out what you don't do well ..... then don't do it!" :)

CAHess-Den
Send message
Joined: 20 May 99
Posts: 21
Credit: 1,552,744
RAC: 457
United States
Message 920252 - Posted: 22 Jul 2009, 7:47:25 UTC - in response to Message 920196.
Last modified: 22 Jul 2009, 8:16:23 UTC

Well, nice catch: I have the "Application" column collapsed to near nothing, so never saw the reams of "6.08 (cuda)" MBs in my queue!

So this sheds light on what I'm thinking is a significant defect in SETI project's "work planning:" What IT thinks I want it to do and what I want it to do are drastically different, and the only means that I'd have to control that, would be to stop doing CUDA MBs. Which is the worst possible scenario for the project.

So instead we've got some SETI@home "app" to manage the problem from some other way?

Too bad, all that, really.

Anyway, I guess I'll give Reschedule a try then, and see how it does at letting me balance my CPU/GPU MB work by hand.... &^}

Thx!

*edit* Well, that seems to have possibly done the trick! I've got a bunch of CPU units running now!

It's going to take a little trial & error to work out how Reschedule7 makes its decisions, tho....
____________

Fred W
Volunteer tester
Send message
Joined: 13 Jun 99
Posts: 2524
Credit: 11,954,210
RAC: 0
United Kingdom
Message 920289 - Posted: 22 Jul 2009, 11:01:37 UTC - in response to Message 920252.

Well, nice catch: I have the "Application" column collapsed to near nothing, so never saw the reams of "6.08 (cuda)" MBs in my queue!

So this sheds light on what I'm thinking is a significant defect in SETI project's "work planning:" What IT thinks I want it to do and what I want it to do are drastically different, and the only means that I'd have to control that, would be to stop doing CUDA MBs. Which is the worst possible scenario for the project.

So instead we've got some SETI@home "app" to manage the problem from some other way?

Too bad, all that, really.

Anyway, I guess I'll give Reschedule a try then, and see how it does at letting me balance my CPU/GPU MB work by hand.... &^}

Thx!

*edit* Well, that seems to have possibly done the trick! I've got a bunch of CPU units running now!

It's going to take a little trial & error to work out how Reschedule7 makes its decisions, tho....

Your computer(s) are hidden so we can't see what version of Boinc you are running. Boinc 6.6.xx keep separate queues for CPU and GPU (which 6.4.x didn't - I suspect the root of your problem).

Perhaps investigate a Boinc version upgrade?

F.
____________

CAHess-Den
Send message
Joined: 20 May 99
Posts: 21
Credit: 1,552,744
RAC: 457
United States
Message 920733 - Posted: 23 Jul 2009, 18:00:10 UTC - in response to Message 920289.



I'm using BOINC v6.6.31. (I should have included this in my "Setup" description. My bad.)

Actually, it sounds like I should've stuck with 6.4.X then, NOT "an upgrade" to the more-current 6.6.xx versions that separate them out!

Due to my "staying current" I have now created MORE work for ME to manage my queue manually with Reschedule7....
____________

Profile BilBg
Volunteer tester
Avatar
Send message
Joined: 27 May 07
Posts: 2632
Credit: 5,979,951
RAC: 3,827
Bulgaria
Message 920922 - Posted: 24 Jul 2009, 3:14:12 UTC - in response to Message 920733.
Last modified: 24 Jul 2009, 3:19:06 UTC



I'm using BOINC v6.6.31. (I should have included this in my "Setup" description. My bad.)

Actually, it sounds like I should've stuck with 6.4.X then, NOT "an upgrade" to the more-current 6.6.xx versions that separate them out!

Due to my "staying current" I have now created MORE work for ME to manage my queue manually with Reschedule7....



BOINC v6.6.31 has a BUG
- when you exit BOINC Manager it does not exit (shutdown) the connected client (BOINC.exe).

I suggest you to upgrade to v6.6.36

If you want to you can also downgrade to 6.4.7 from here:
http://boinc.berkeley.edu/download_all.php
____________



- ALF - "Find out what you don't do well ..... then don't do it!" :)

CAHess-Den
Send message
Joined: 20 May 99
Posts: 21
Credit: 1,552,744
RAC: 457
United States
Message 921172 - Posted: 25 Jul 2009, 0:41:52 UTC - in response to Message 920922.


[...]

BOINC v6.6.31 has a BUG
- when you exit BOINC Manager it does not exit (shutdown) the connected client (BOINC.exe).

I suggest you to upgrade to v6.6.36

If you want to you can also downgrade to 6.4.7 from here:
http://boinc.berkeley.edu/download_all.php


Thx & wilco.

I didn't know that it'd been fixed!

____________

Profile Questor
Volunteer tester
Send message
Joined: 3 Sep 04
Posts: 362
Credit: 148,364,591
RAC: 57,357
United Kingdom
Message 921389 - Posted: 26 Jul 2009, 2:59:47 UTC - in response to Message 921172.

While in the mood for upgrades there is a newer version of Reschedule1.9 also available on the Lunatics site.
____________
GPU Users Group



Questions and Answers : GPU applications : CPU / GPU mixed processing problem

Copyright © 2014 University of California