Not Enough cuda_femi WUs? GPU Starvation!


log in

Advanced search

Message boards : Number crunching : Not Enough cuda_femi WUs? GPU Starvation!

Author Message
hancocka
Send message
Joined: 19 May 00
Posts: 10
Credit: 2,517,833
RAC: 23
United Kingdom
Message 1322334 - Posted: 30 Dec 2012, 19:05:49 UTC
Last modified: 30 Dec 2012, 19:06:08 UTC

I'm not getting enough WU Tasks downloaded to keep my GPUs crunching?

I've set my preferences to 10 days minimum, and 10 days additional work buffer.

is it something todo with the CPU, I've got enough for the CPU to be busy, with 10 threads, but my 2 x GTX590s, crunch 8 WUs in no time at all, and all the WUs downloaded are then set to ready to report.

e.g. currently got to

5 x Ready To Report: Local Setiathome_enhanced 6.03
20 x Ready To Report: Setiathome_enhanced 6.10 (cuda_femi)
67 x Ready to Start Local Setiathome_enhanced 6.03

Some Astropulse ones running at present and a few to report.

It seems, I'm often stuck in Ready to Report modes, or have I got the Balance wrong, and should turn OFF CPU!

Any advice?
____________

rob smithProject donor
Volunteer tester
Send message
Joined: 7 Mar 03
Posts: 8734
Credit: 61,635,402
RAC: 49,464
United Kingdom
Message 1322340 - Posted: 30 Dec 2012, 19:17:52 UTC - in response to Message 1322334.

Just now there are limits on the number of tasks "in progress", I think they are one hundred per processor. Now this may be 100 per processor type, so you are hard on that limit. If it is 100 per processor it would be reasonable to expect your big machine to report 500 "in progress". There are quite a few reports of upload & download problems, so don't panic as everyone is probably in the same situation with tasks not being transferred or reported as smoothly as one would expect/desire.
("In progress" means not actually being worked on, but have been assigned to, but not reported by, a particular machine)
____________
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

ClaggyProject donor
Volunteer tester
Send message
Joined: 5 Jul 99
Posts: 4209
Credit: 34,469,243
RAC: 18,525
United Kingdom
Message 1322347 - Posted: 30 Dec 2012, 19:28:50 UTC - in response to Message 1322334.
Last modified: 30 Dec 2012, 19:38:03 UTC

Firstly, Setiathome at present has a limit of 100 CPU and 100 GPU tasks, so setting 10 + 10 days won't get you extra work, and will likely get you less work because the High AR work only has a deadline of 14 days so won't be sent,

Secondly, Boinc 7 has a new Scheduler that operates differently to the scheduler in Boinc 6,
setting 10 days Minimum and 10 days Additional will make Boinc 7 try and fill up to 20 days work, then it won't ask for work again until it gets below 10 days work,
a Cache setting of 3 days minimum and 0.01 days additional is more appropiate, Boinc will fill to 3.01 days work, and every time it goes below 3 days work it'll ask for more,

BOINC 7.0

In finer detail
The CPU and GPU schedulers have been rewritten from the ground up, as is the new work fetch module. This means that work fetch will happen completely different from how you're used to it in BOINC 6. Where BOINC 6 would for instance report work and immediately ask a new task, BOINC 7.0 will try to store a couple of "ready to report" results, report them and immediately request work only when BOINC is under the value set by the Minimum work buffer.

This means that when you have left your old BOINC 6 values for "Connect to" + "Additional work" at, example given 0.1 and 1.0, that BOINC 7.0 will ask for 1.1 days worth of work and ONLY renew this cache when it's fallen under the 0.1 days worth of work limit.

Which means that it can happen that your BOINC runs empty, because 7.0 won't request new work before it has dropped below the 'minimum work' setting and will only ask for work up to the 'and additional' setting --and that only from the project that has the highest priority (worst REC to resource share ratio). Only if that project doesn't have work it will ask other projects in order of priority.


Claggy

hancocka
Send message
Joined: 19 May 00
Posts: 10
Credit: 2,517,833
RAC: 23
United Kingdom
Message 1322357 - Posted: 30 Dec 2012, 19:41:17 UTC - in response to Message 1322347.
Last modified: 30 Dec 2012, 19:42:26 UTC

okay, I'll adjust to 3 min and 0.01 days.

I did wonder, if it was just home desktop computers, had got too fast for Seti WU/s!
____________

Profile Paul D Harris
Volunteer tester
Send message
Joined: 1 Dec 99
Posts: 1123
Credit: 33,598,472
RAC: 0
United States
Message 1322499 - Posted: 31 Dec 2012, 2:17:12 UTC

Mine keeps about 100 wu cache for the cpu 8 at a time and 100 wu for the 2 gpu 2 at a time that is 4 wu at a time total for gpu. BOINC keeps it around 100 all the time for the cpu and the gpus.
____________

Message boards : Number crunching : Not Enough cuda_femi WUs? GPU Starvation!

Copyright © 2014 University of California