Questions and Answers :
GPU applications :
How to set single host resource share? (CUDA!)
Message board moderation
Previous · 1 · 2
Author | Message |
---|---|
ML1 Send message Joined: 25 Nov 01 Posts: 20389 Credit: 7,508,002 RAC: 20 |
It has no s@h work. OK... So I aborted the suspended CPDN work and... a few work fetch requests later and a few s@h WUs have been downloaded and are now being worked on. So... Looks like a quirk of the scheduler logic. If the CPU is considered maxed out on current (and suspended) work, then the GPU can be starved. Yet "EDF" does not consider suspended work and you can push the client out of EDF by suspending WUs. Also, the (expected) CPU fraction for feeding the GPU is preset in the xml... Reality varies and is different! A good question is how multiple resources should be balanced for the user specified resource share when there are more than one or there are alternate bottlenecks to performance (CPU maxed out in this case and the GPU utilisation is dependant on the CPU). And there is still my original question of how you can set unique resource shares for a host rather than being limited to the default-home-school-work values? Happy crunchin', Martin See new freedom: Mageia Linux Take a look for yourself: Linux Format The Future is what We all make IT (GPLv3) |
Jord Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 |
And there is still my original question of how you can set unique resource shares for a host rather than being limited to the default-home-school-work values? BAM! has user made venues. But else, you're out of luck, I think. |
John McLeod VII Send message Joined: 15 Jul 99 Posts: 24806 Credit: 790,712 RAC: 0 |
Any project that has a contact backoff will not be eligible. Yes. If it has a backoff, it is not allowed to fetch work. BOINC WIKI |
Jord Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 |
Since we just learned that the anonymous platform mechanism is expecting to find CPU information only in the app_info.xml file, you think that that's at the start of Martin's problem? (BOINC 6.6 does not support CUDA through the app_info.xml) |
ML1 Send message Joined: 25 Nov 01 Posts: 20389 Credit: 7,508,002 RAC: 20 |
Since we just learned that the anonymous platform mechanism is expecting to find CPU information only in the app_info.xml file, you think that that's at the start of Martin's problem? Now that gets even more confusing because it appears to be getting WUs and returning results via CUDA... For example: Task 1178537956 Happy crunchin', Martin See new freedom: Mageia Linux Take a look for yourself: Linux Format The Future is what We all make IT (GPLv3) |
ML1 Send message Joined: 25 Nov 01 Posts: 20389 Credit: 7,508,002 RAC: 20 |
Aside: Note that the original problem is in the confusion in how work is added up by the scheduler for CPU tasks and GPU tasks and for considering tasks that are user suspended. The refusal to request more work was cleared by aborting the queued up CPDN tasks. Suspending those tasks merely removed the EDF status. Suspending the CPDN tasks wouldn't allow any s@h WUs to be downloaded. There is still the anomalous server-side messages about not completing work in time even though Boinc gets 100% CPU time! A local resource share override would be rather nice ;-) Cheers, Martin See new freedom: Mageia Linux Take a look for yourself: Linux Format The Future is what We all make IT (GPLv3) |
John McLeod VII Send message Joined: 15 Jul 99 Posts: 24806 Credit: 790,712 RAC: 0 |
Since we just learned that the anonymous platform mechanism is expecting to find CPU information only in the app_info.xml file, you think that that's at the start of Martin's problem? The latest word is that the client does, but the server does not really support it yet. It should only take a couple of days to get the CUDA anonymous platform support to work correctly at the server. BOINC WIKI |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.