I recently saw that my GTS 240 GPU was processing a "Gamma-ray pulsar search #2 1.12 (FGRPopencl-nvidia)" task, but the GPU load was 6%. After restarting BOINC, the GPU load wouldn't go above 0%. The task is slowly progressing, but I don't think it's getting any help from the GPU.
I've seen 1 report in the nVidia Forums, on the 331.58 driver feedback thread, where users were complaining about OpenCL performance, but... my situation here seems different.
Is anyone else noticing OpenCL performance issues on the latest drivers, possibly on older GPUs?
Windows 8.1 x64
Copyright © 2024 Einstein@Home. All rights reserved.
OpenCL tasks - Low GPU% on 331.58 drivers?
)
Time to test 331.65 drivers on that same task!
331.65 drivers are exhibiting
)
331.65 drivers are exhibiting the same behavior - only 0-12% GPU usage.
What I need to know is:
Is this an Albert application issue, an Albert task issue, or is this this an nVidia driver issue?
The task that completed
)
The task that completed successfully had a bunch of error info in the Std error output portion.
See:
http://albertathome.org/task/1179842
There were a lot of these lines:
Error in OpenCL context: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_WRITE_BUFFER on GeForce GTS 240 (Device 0).
Error during OpenCL host->device transfer (error: -4)
Any ideas?
I wanted to chime in to
)
I wanted to chime in to mention that, today, I spent about 4 hours testing every released beta/whql driver version back to 314.22, on my Windows 8.1 x64 machine. I was testing OpenCL performance of the Albert@Home "Gamma-ray pulsar search #2 1.12 (FGRPopencl-nvidia)" task, on my GTS 240, for each driver version.
The performance results are:
331.65: Bad
331.58: Bad
331.40: Bad
327.23: Very Bad
326.80: Very Bad
326.41: Very Bad
326.19: Very Bad
320.49: Very Bad
320.18: Very Bad
320.14: Very Bad
320.00: Very Bad
314.22: Very Bad
... where "Bad" means GPU Load % fluctuating between 0-25%, and "Very Bad" means GPU Load % fluctuating between 0-9%.
So, if this is a regression, it's not recent. It's entirely possible that my issue is with the task itself, and not within the drivers.
Thoughts?
Thanks,
Jacob
A GTS 240 GPU (G92a or G92b
)
A GTS 240 GPU (G92a or G92b chip, depending on version) is a very old and slow GPU (Q4 2009).
Most of the available performance will have been squeezed out of those years ago. Newer drivers will still be trying to leverage more performance out of:
GFxxx (Fermi) chips - 2010 onwards
GKxxx (Kepler) chips - 2012 onwards
GK110 (Titan-class) chips - 2013 onwards
And remember the huge driver architecture changes between Windows XP and Vista/7/8 (WDDM model).
Your GTS 240 would probably be happiest with Windows XP and a legacy driver - but in that system (and depending on the application - wait for the CUDA 6 tests), even a baby Kepler should still be showing improvement as the drivers are refined.
I didn't want to spam the boards with my stats - just milestone theads - but apparently signatures are no longer optional. Follow the link if you're interested.
http://www.boincsynergy.com/images/stats/comb-3475.jpg
Let me put it another
)
Let me put it another way:
What is the expected behavior of an "FGRPopencl-nvidia" Albert@Home task... on my GTX 660 Ti? What GPU Load % should I expect on that beefy GPU?
Here is what I'm currently seeing on the 331.65 drivers (monitoring with eVGA Precision-X at a 100ms polling interval):
GTX 660 Ti: Super-quick flickers, going from 0% to 33% back to 0%, about 4 times a second.
GTX 460: 23% most of the time, but brief surges downward to 15% for usually less than 2 seconds.
GTS 240: Bouncing around about once a second, between being at 13% and being at 26%.
Does this sound like correct behavior for those architectures, for that task type? :)
My CUDA tasks are usually 90% constant on the GPUs, which is why I thought the GPU Loads reported here looked suspicious. (And yes, I know CUDA is way different than OpenCL, but... are the low loads in this post really the expected loads for this task type?)
I did some testing a little
)
I did some testing a little while back and determined that a 660Ti Win7 x64 needs to run 4 concurrent tasks w/ 1 cpu core for each GPU task to keep properly fed. I forget how long they took but IIRC they only garner 70 pts. each ... call me a pt hore but I'm crunch other Einstein\ Albert GPU tasks until we see something new here.
I only have selected Perseus
)
I only have selected Perseus ARM only but occasionally get an FRGP.
Is this by design or an oddity of running beta?
TIA,
Steve
When this happens, getting
)
When this happens, getting the wrong kind of work, see if you can get the server log for that contact. To do this go to your list of computers and then in the rightmost column click on the datestamp for the last contact to display the log, post it here and maybe that can shed some light on why you are allocated work for an application you have not selected.
I apologize in advance for
)
I apologize in advance for the wall of text about to ensue as I'm not sure what precisely is relevant and there are references from the beginning to the end for 2 GAMMA WU buried in this. I do move hardware between rigs and I have aborted a couple of GAMMA WUs in the past. I did let these 2 run with mixed results. The first ran on a 7950 concurrently with 2 Milkyway WUS and experienced a computation error which is likely more to do with my card OC than anything else. The second ran to completion on a 7850 with an additional concurrent Perseus Arm task and is now pending validation. One last side note, the message board removes duplicate space characters but in the original the 31st line there was only a single space before the [CRITICAL] tag, all other lines have 4 spaces (likely a tab).
I'm not concerned about getting unselected gamma WU as I'll crunch what I get, I promise no more aborts, but maybe something in the scheduler needs a tweak?
- Steve