GRPS#3 with Nvidia stops computing?
29 Mar 2014 21:04:42 UTC
Topic 197501
(moderation:
Hello, I'm currently running Windows 8.1 with BOINC version 7.2.42. I have been running Einstein Gamma-ray Pulsar Search #3 using my NVidia card with driver version 335.23. I have noticed when the workunit gets to around 91%, the unit stops crunching. The GPU temperature drops from the normal operating temperature of 50C (while workunits are crunching) to its idle temperature of 31-32C... it gets stuck. If I allow it to run, it just sits. As far as I can tell its an intermittent problem, as some workunits have crunched properly and been validated. I am using the NVidia video card just for crunching, as I have my integrated Intel graphics running my computer graphics. Has anyone else experienced this problem before or is it just me? I hope its not indicative of a hardware problem.
GRPS#3 with Nvidia stops computing?
)
I think you are seeing normal behavior, and wasting resources by terminating by hand jobs which left to themselves would finish normally.
As I understand it, late in the GRP3 run there is a different portion of the computation. In the current code, that portion does not use the GPU materially. also the computation required by that portion varies substantially from WU to WU, and lastly no useful completion percentage progress is reported.
Suggestion: The next time you think have a case of this, leave it running until it has at least doubled the elapsed time at which it changed behavior. Let us know what you see.
On the other hand, the GRP3 work as currently configured does not make heavy use of your GPU compared to the other Einstein GPU applications. Many of your peers have chosen either to exclude GRP3 entirely from running on their GPU-equipped hosts, or else only to allow CPU GRP3 jobs by setting the "GPU utilization factor of FGRP apps" item in Einstein preferences for the location (aka venue) of the host to -1.
Thank you for clarifying the
)
Thank you for clarifying the issue. You are 100% correct. The computing pause the GPU experiences is by design and the workunits are completing correctly.