2 WU's on one card?

Message boards : Number crunching : 2 WU's on one card?

Author	Message
Daniel Neely Send message Joined: 21 Feb 09 Posts: 5 Credit: 36,705,213 RAC: 0 Level Scientific publications	Message 20339 - Posted: 4 Feb 2011 \| 0:31:59 UTC
	I've noticed that my 260's are only running at 50-70% load and was wondering if anyone had tried using an app config file to run 2 tasks concurrently in order to increase throughput?
	ID: 20339 \| Rating: 0 \| rate: / Reply Quote

[B^S] thierry@home Send message Joined: 7 Sep 08 Posts: 7 Credit: 7,504,776 RAC: 0 Level Scientific publications	Message 20732 - Posted: 20 Mar 2011 \| 11:09:43 UTC - in response to Message 20339.
	I have the same question. I use a GTX 480 and the GPU load is only at 55%. With this card I run 4 Einstein WUs or 2 Seti WUs at a time. So it must be possible to run 2 GPUgrid WUs together. Does somebody has a app_info file to do this? Thanks
	ID: 20732 \| Rating: 0 \| rate: / Reply Quote

skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications	Message 20733 - Posted: 20 Mar 2011 \| 11:28:59 UTC - in response to Message 20732.
	Yes. A few people have done this. The problem is that it requires a lot of work on the crunchers part, to setup and maintain the app file. I have only tried it with short tasks and I would be concerned about doing this for long tasks (RAM amount and runtime/credit - can you complete 2 tasks inside 24h)? For GIANNI tasks there would be no benefit, in fact it might slow them down overall. For KASHIF tasks you might see a slight increase in turnover (tasks and credits), but not much. For IBUCH tasks I would expect you to see perhaps 10 to 15% increase, compaired to a good 1task per card setup. I cannot confirm that two long tasks will not use too much video memory, or that they would over-run past 24h, resulting in lower credit. You would also have to account for the affect on your card. Will it tax the card to the point that you start getting failures. If so then this would undo any benefit. In my opinion it would only be worth it on CC2.0 cards, if at all, and not cc2.1 cards.
	ID: 20733 \| Rating: 0 \| rate: / Reply Quote

[B^S] thierry@home Send message Joined: 7 Sep 08 Posts: 7 Credit: 7,504,776 RAC: 0 Level Scientific publications	Message 20734 - Posted: 20 Mar 2011 \| 13:50:00 UTC - in response to Message 20733. Last modified: 20 Mar 2011 \| 13:52:53 UTC
	Apparently (it's my first long run) a WU is crunched in ~15 hours. So if running two WUs slow down the process, 24 should be enough. As an example, I take 45' to crunch an Einstein WU, 60' for two, 75' for three and 90' for four, at the same time, with a max of 85% of GPU load for the last case. And concerning the card, when I run PrimeGrid, the load is at 99%, so it can be worst ;-) And 15% is 15%. It would be nice to try .....
	ID: 20734 \| Rating: 0 \| rate: / Reply Quote

Snow Crash Send message Joined: 4 Apr 09 Posts: 450 Credit: 539,316,349 RAC: 0 Level Scientific publications	Message 20752 - Posted: 21 Mar 2011 \| 22:20:43 UTC
	I'm going to give this a try soon. Perhaps we can get some real crunching going on with the GTX460 series. I'm thinking that there is so much unused resources on the cards themselves and waiting on Windows, specifically in Vista/Win7, that the overall % increase will be well worth the effort. Our team is particulary tech savy so it should not be diffiuclt to understand the implications that they will always need to keep up with files/ version maintenance of their app_info file. Also, we have a dedicated subforum for GPUGrid so we can all help each other keep up on these details. Does anyone have an app_info file that I could use as a base? I'm thinking I would like to test the "shorties" before delving into the realm of loooong WUs :-) TIA, Steve
	ID: 20752 \| Rating: 0 \| rate: / Reply Quote

Snow Crash Send message Joined: 4 Apr 09 Posts: 450 Credit: 539,316,349 RAC: 0 Level Scientific publications	Message 21222 - Posted: 22 May 2011 \| 13:28:36 UTC
	Like everyone else found, this multi WUs did not improve anything a 460. Now I would like to try with LONG Wus on a 480 but I can't seem to get the app_info correct ... if there is anyone who could provdie a little direction I would appreciate it, I think my problems are mainly on knowing the actual app name and class plan <app_info> <app> <name>acemd2</name> </app> <file_info> <name>acemdlong_6.13_windows_intelx86__cuda31.exe</name> <executable/> </file_info> <file_info> <name>cudart32_31_9.dll</name> <executable/> </file_info> <file_info> <name>cufft32_31_9.dll</name> <executable/> </file_info> <file_info> <name>tcl85.dll</name> <executable/> </file_info> <app_version> <app_name>acemd2</app_name> <version_num>613</version_num> <avg_ncpus>0.054780</avg_ncpus> <max_ncpus>0.054780</max_ncpus> <flops>17842995981.141441</flops> <plan_class>cuda32</plan_class> <coproc> <type>CUDA</type> <count>0.5</count> </coproc> <gpu_ram>249561088.000000</gpu_ram> <file_ref> <file_name>acemdlong_6.13_windows_intelx86__cuda31.exe</file_name> <main_program/> </file_ref> <file_ref> <file_name>cudart32_31_9.dll</file_name> </file_ref> <file_ref> <file_name>cufft32_31_9.dll</file_name> </file_ref> <file_ref> <file_name>tcl85.dll</file_name> </file_ref> </app_version> </app_info> ____________ Thanks - Steve
	ID: 21222 \| Rating: 0 \| rate: / Reply Quote

Snow Crash Send message Joined: 4 Apr 09 Posts: 450 Credit: 539,316,349 RAC: 0 Level Scientific publications	Message 21226 - Posted: 22 May 2011 \| 21:06:00 UTC - in response to Message 21222.
	anyone looking to build a custom app_info should make a copy of their client_state.xml file as that is where all the details necessary can be found ;-) ____________ Thanks - Steve
	ID: 21226 \| Rating: 0 \| rate: / Reply Quote

Post to thread

Message boards : Number crunching : 2 WU's on one card?

	About	Science	Volunteers	Performance	Forum	Join us	Donate