Advanced search

Message boards : Number crunching : Warnings (and recommendations) for Linux crunchers

Author Message
Profile Saenger
Avatar
Send message
Joined: 20 Jul 08
Posts: 134
Credit: 5,807,001
RAC: 33,530
Level
Ser
Scientific publications
watwatwatwatwatwat
Message 18352 - Posted: 16 Aug 2010 | 17:16:01 UTC

First:
This is just a warning about some non-standard behaviour, not about the project in general. I have Linux and run it, it's just that it isn't running as I think it's supposed to do (and as it states it does in the BM).

Here we go:

The project will use a full core of your CPU for the application plus some percentages (depending on the power of your GPU) for the Xorg process. The less power, the more Xorg is working.

The nice value of the CPU part of the application is 10 instead of the usual 19, so it gets a whole core.

As it signals BOINC that it will only use 15% of a core other applications are running in parallel to the one GPU, depending on the number of cores one each. Those get the remaining cores to crunch, on my quad that's 4 apps for 3 cores, on a duo it will be 2 on one core.

BOINC doesn't seem to handle this situation that good, as on my machine the CPU-usage fluctuates quite considerably. After I put the nice value of the GPU-task manually to 19, the usage was again a straight line at 100%. As my card isn't that fast I think I will see at least every third WU in time to change it by hand, on faster GPUs it would be probably quite challenging to keep track of it.

I tried to restrict the use of CPUs to 3, in the hope that GPUgrid will manage to "steal the remaining one and only 3 others will run in parallel, it didn't go well. There were only 3 others running, plus the one from GPUgrid, but they didn't put all 4 cores to proper work.

My personal Recommendations for users:

For Linux your GPU should be even faster than for Windows, as it will utilize a lot of CPU power otherwise and take it away from other projects without acknowledging it.

If possible put the nice value to 19 for the running WU to keep the other projects and the CPU happy.

Expect other projects to go down in it's productivity.

Recommendations for the project team:

Change the value for the used CPU in the WU from 15% (0.15 CPUs) to the real needed 100% (1 CPUs), so only n-1 other apps will run on a n-core machine.

Change the nice value for the CPU-part to 19, as is usual in BOINC.
____________
Gruesse vom Saenger

For questions about Boinc look in the BOINC-Wiki

Profile Saenger
Avatar
Send message
Joined: 20 Jul 08
Posts: 134
Credit: 5,807,001
RAC: 33,530
Level
Ser
Scientific publications
watwatwatwatwatwat
Message 18358 - Posted: 17 Aug 2010 | 15:35:23 UTC

Here's the difference of CPU-load with nice10 and nice19:

____________
Gruesse vom Saenger

For questions about Boinc look in the BOINC-Wiki

Profile Saenger
Avatar
Send message
Joined: 20 Jul 08
Posts: 134
Credit: 5,807,001
RAC: 33,530
Level
Ser
Scientific publications
watwatwatwatwatwat
Message 18387 - Posted: 23 Aug 2010 | 8:21:42 UTC
Last modified: 23 Aug 2010 | 8:52:45 UTC

I've noticed a change in the runtime and CPU-time with WUs on nice10 and nice19:

    - With nice10 the runtime and CPU-time is, as expected, nearly the same, for the 6803-credits-WUs on my machine ~10h each.
    - With nice19 the runtime increased, about 1.5h, but the CPU-time decreased about the same amount, so there is a significant difference between CPU-time and runtime now.


So the changing is not that good for the runtimes (clockwise) of GPUgrid, but it's much fairer to other projects running in parallel. I don't know whether the GPU-load increased or decreased, I don't know how to monitor that.
____________
Gruesse vom Saenger

For questions about Boinc look in the BOINC-Wiki

Post to thread

Message boards : Number crunching : Warnings (and recommendations) for Linux crunchers