Thread 'Multi-GPU multi-project scheduling oddity'

Message boards : BOINC client : Multi-GPU multi-project scheduling oddity
Message board moderation

To post messages, you must log in.

AuthorMessage
jjv1001
Avatar

Send message
Joined: 6 Dec 11
Posts: 8
Finland
Message 93484 - Posted: 2 Nov 2019, 11:45:25 UTC

I seem to be hitting some sort of a corner case with the scheduler. A quickish explanation:
6 projects, 3 of them have both CPU and GPU jobs. In order to maintain a balance between projects a limitation on concurrent jobs is imposed. Also project A is limited to GPU1 while project B is limited to GPU2. Project C is allowed on both. All GPU jobs are done fractionally to allow multiple jobs on a single GPU.
Now the situation is that project A decided to send me a ton of CPU jobs. This combined with the limit on concurrent jobs forces the project to compute CPU work only. Now project A is running on CPU only, project B is still using GPU2 happily but project C seems to insist on also using GPU2 leaving the entire GPU1 idle.

JJ
ID: 93484 · Report as offensive
Richard Haselgrove
Volunteer tester
Help desk expert

Send message
Joined: 5 Oct 06
Posts: 5133
United Kingdom
Message 93485 - Posted: 2 Nov 2019, 13:26:48 UTC - in response to Message 93484.  

Projects do not 'decide' to send you work: your BOINC client decides to 'Request' work.

To get to the bottom of this, you're probably going to have to do some intense work with the Event Log. Start by setting <sched_op_debug> - that adds comparatively little to the total volume of messages, but will get you started with understanding how the process works.

If you're running a recent version of BOINC (you mention neither that, not your operating system), you can set event log flags from Ctrl+Shift+F. Don't set too many options at once: too many massages can overwhelm the system.
ID: 93485 · Report as offensive
jjv1001
Avatar

Send message
Joined: 6 Dec 11
Posts: 8
Finland
Message 93487 - Posted: 2 Nov 2019, 15:31:23 UTC

Wrong turn of phrase. The request for work was of course initiated by my client. I think something is off with the amount of work received but that really is not the issue. I'm entirely okay with a project doing CPU only work for a while although in this case the amount of jobs in the queue will never finish before the deadline. Actually now that I look at it I have three pages worth of "not started by deadline" cancellations so the current situation will probably fix itself in short order.
The actual problem is the weird logic in stuffing the secondary GPU full of work while leaving the primary entirely empty.
I do believe that the problem is platform independent and have actually seen it occasionally appear over the years and versions. But the platform is Win x64 and the client is currently 7.16.3.
Should it matter the actual projects in question are A - Einstein, B - Milkyway and C - SETI
I'll keep monitoring the logs.

JJ
ID: 93487 · Report as offensive

Message boards : BOINC client : Multi-GPU multi-project scheduling oddity

Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.