Not enough CUDA jobs to keep busy
23 Sep 2012 16:17:07 UTC
Topic 196533
(moderation:
I have 2 computers that seem to run out of GPU jobs and I'm not sure if it's a problem on my part or hitting some sort of daily limit.
As far as I can tell WU are completing and being validated at a normal rate.
I did just bump these two to 3 tasks/GPU which makes me think about a limit.
I'd prefer to learn how to figure this out but if it takes privileges to view the details would someone please look at:
http://einsteinathome.org/host/5771544
and
http://einsteinathome.org/host/5860433
Thanks,
Joe
Not enough CUDA jobs to keep busy
)
More info on limits or daily quota can be found in this post by Bikeman:
http://einsteinathome.org/node/196400&nowrap=true#117984
Do these hosts run work for
)
Do these hosts run work for other projects?
The project balancing activity of the schedulers can lead to non-request for work in situations which seem rather odd to me.
RE: Do these hosts run work
)
No, actually e@h is my only project.
Joe
RE: More info on limits or
)
thank you for the link. I posted these cpus in that thread hopefully someone will take a look for me.
Joe
Please don't post the same
)
Please don't post the same problem in two different threads.
I saw the other one first and replied there. It would have been better to have only posted here with all responses in the one place.
Cheers,
Gary.
RE: Please don't post the
)
Sorry Gary,
I posted the question and was referred the thread that was directly on point.
What would be the appropriate way to respond to Bernd's request to be notified if anyone else saw the issue after his change?
Joe
RE: I posted the question
)
Being referred to another thread for the purposes of getting relevant information is hardly a demand for you to post there. Unfortunately, the link given to you didn't quite point to the best message for information about daily quotas. The later post by Bikeman gave the new formula being used (32 per CPU core + 160 per GPU) which in your case equates to 256+160 = 416 tasks per day - unless the quota has been diminished by errors. In the links to your hosts you supplied, you can actually see the daily CPU core limit for each machine (both at the max of 32) so there is no problem with errors. You can be confident you have the full quota and your problem must lie elsewhere.
So really, the other thread wasn't "directly on point" at all, for the purpose of working out why you weren't getting adequate work.
First of all, be really sure that the quotas weren't behaving as intended. If you're not really sure, stick with the existing thread and wait for responses. If it becomes clear that the quotas really are screwed, send Bernd a PM, and post your analysis in the other thread too, if you wish. I'm sure Bernd would have no objection to receiving a PM about a genuine bug/problem with the server code.
Please be aware that there was no criticism intended in my earlier response - nor is there any in this one. I have a 'thing' about trying to keep related information in the one place so that others are likely to find it all without extra effort. I would be really pleased if (when you identify the problem) you post your final conclusions here. That way, anyone else having trouble getting CUDA tasks should find suggestions with an appropriate search.
Cheers,
Gary.
Thanks Gary! I'm still
)
Thanks Gary!
I'm still trying to track this down.
I will post here if I find anything pertinent.
I do have a mixture of boinc versions and am using all my locations but these two machines are the only ones in the "school" location because they seem to be able to run 3 jobs per GPU.
Joe
I found it and it was an
)
I found it and it was an embarrassingly stupid mistake on my part.
I moved these two computers to a separate location so they could have their own project preferences because I wanted to try 3 CUDA tasks per GPU.
I'm not sure how this works exactly because I did get 3 tasks running, up from 2 in the previous location.
However, I completely missed the attribute "Use NVIDIA GPU" which was set to NO.
The only thing I can think of is that I had WUs in the queue and these got processed 3 at a time but the parameter stopped it from requesting new tasks.
Setting the project parameter to use the GPU and updating the project and I'm back heating my office at full blast.
Joe