Advanced search

Message boards : Number crunching : Have I biased the GPU-Grid servers against my GPU?

Author Message
dirkmittler
Send message
Joined: 13 Mar 12
Posts: 21
Credit: 8,773,573
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwatwat
Message 49367 - Posted: 28 Apr 2018 | 19:32:17 UTC

Hello.

It has been a long time, since I did any number-crunching. In the meantime I have switched the PC I had for doing so, from the Windows 7 O/S, to Debian 9, aka Debian / Stretch. The GPU on this machine is a GeForce GTX460, which, AFAICT, still works perfectly.

I have two observations about my recent history - barely 2 days old - on GPU-Grid:

1) When I recently upgraded my Linux-based graphics-driver, from the Open-Source ('Nouveau') driver, to the proprietary driver, I had quite a mess of hickups, before I decided that my present setup is stable again. During this recent setup interval, at one point I felt that the only decent way to proceed, was actually to abandon a batch of GPU-Grid jobs, precisely so that other BOINC-clients could still complete them.

2) Now that I have set my BOIN-client parameters as they should be for the long haul, I no longer seem to be receiving jobs, that use my GPU.

Now, there could be several reasons, why GPU-Grid is no longer asking for my GPU:

- Presently, I'm limiting BOINC to using only 2 CPU-cores. It has only been since yesterday that I've noticed, some BOINC jobs actually need 4 CPU-cores, to run 1 job. This is very different from how it once was. Is it the case that only the 4-CPU-core jobs, also make use of the GPU, in the modern era?

- Or, did the sight of the numerous jobs abandoned by the current machine, on April 27, get misinterpreted by the admins to mean, that this machine somehow fails to carry out GPU-computing, since I seem to recall that many of the abandoned jobs were in fact using my GPU - briefly.

I think that I can assert, that what happened on April 27 was not the fault of my setup or drivers, but was just the fault if you will, of how I clumsily installed my software, and the fault of not remembering quickly enough, how to configure BOINC-clients, still on April 27.

It will remain a constraint of mine, to limit BOINC to using 2 CPU-cores. But at the same time, I'm curious to find out for the first time, whether my setup will be capable of actually completing jobs that use the GPU - since my GPU has only been CUDA-capable since yesterday - as long as I don't just delete them again. :)

Dirk

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 1942
Credit: 12,280,400,519
RAC: 2,868,528
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 49368 - Posted: 28 Apr 2018 | 20:02:07 UTC - in response to Message 49367.
Last modified: 28 Apr 2018 | 20:04:54 UTC

The GPU on this machine is a GeForce GTX460
This GPU is not supported by the GPUGrid GPU app anymore.

- Presently, I'm limiting BOINC to using only 2 CPU-cores. It has only been since yesterday that I've noticed, some BOINC jobs actually need 4 CPU-cores, to run 1 job. This is very different from how it once was. Is it the case that only the 4-CPU-core jobs, also make use of the GPU, in the modern era?
There are two different apps on GPUGrid now:
1. The GPU app which uses CUDA 8.0
2. The CPU app which needs 4 CPU cores, and only one such workunit should start at once (due to a bug in the app)
You've received only the latter, (Quantum Chemistry tasks), because your GPU is too old for GPUGrid GPU apps.

Jim1348
Send message
Joined: 28 Jul 12
Posts: 614
Credit: 1,199,451,727
RAC: 134,958
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 49369 - Posted: 28 Apr 2018 | 20:31:27 UTC
Last modified: 28 Apr 2018 | 20:36:08 UTC

It appears that you are now running the QC jobs OK, but it should be noted that the quantum chemistry application will run on less than 4 cores; that is just the maximum it uses per work unit by default.

If it does not now operate on two cores, you can use an app_config.xml file placed in the "www.gpugrid.net" project folder. In Ubuntu, it is located in the "/var/lib/boinc-client/projects" folder; I don't know about Debian/Stretch, but it could be the same.

This one will limit it to run on two cores, and also limit it to run only one work unit at a time, thereby eliminating the startup problem:

<app_config>

<app>
<name>QC</name>
<max_concurrent>1</max_concurrent>
</app>

<app_version>
<app_name>QC</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>2</avg_ncpus>
<cmdline>--nthreads 2</cmdline>
</app_version>

</app_config>


If you have not used an app_config before, you can create it in a text editor, and then use the "save as" function to save it as an ".xml" file.
To activate it, you should read in the configuration files (or do a reboot). You can change those numbers as you choose.

PS - It actually increase the computation efficiency a little to use fewer than 4 cores; I often run on 1 core per work unit. But if you are memory-limited, the MT application makes more efficient use of memory for running on multiple cores.

dirkmittler
Send message
Joined: 13 Mar 12
Posts: 21
Credit: 8,773,573
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwatwat
Message 49370 - Posted: 28 Apr 2018 | 21:14:37 UTC
Last modified: 28 Apr 2018 | 21:16:56 UTC

I think that people have answered my questions perfectly.

I also never had any problems, with whether by BOINC client assigns the number of cores, which it's supposed to assign.

What did happen, is that I had erroneously set the BOINC client to allow 4 CPU cores to run work units, which immediately resulted in numerous WUs downloading, all of which did state a requirement for 4 cores. But immediately after seeing that, I chose not to be as generous in my settings.

The result of this can be, that all the WUs assigned to me, would never have gotten a chance to run again, because by that time I was only allowing for 2 CPU-cores to be used. And so, the fastest way to receive new WUs seemed to be at the time, to eject the ones I had received.

I guess I'll just have to accept, that my old GPU doesn't run CUDA 8.

Dirk

Post to thread

Message boards : Number crunching : Have I biased the GPU-Grid servers against my GPU?