Running 2 CUDA Instances

[Eroc]

I am curious about running 2 CUDA instances on my cards. I have managed to figure out how and got 2 running on my GTX460 and it works great. Usage went from 60-80% to closer to 95-100%.

I think I read somewhere that the Fermi cards benefit from 2 instances simultaneously, but was wondering if the same applied to non-Fermi cards (i.e. GTX260)

I guess I could try it out and see. Has anyone else had experience with this?

Also, I noticed that nowhere in here is a sticky for setting up 2 instances on a GPU. I had to investigate and even then basically just had to mess with my App_info.xml file until I got it to work.

Have I missed a thread on this?

 

[klear]

I would also be curious to know if running multiple instances per gpu is more productive, especially on g92 and gt200 cards.

Also, should I increase the default .04 CPUs to feed the gpu faster?

 

[Eroc]

Good question. I don't think that matters as much as long as you free up a core or two to feed the gpu. Watch you cpu usage. On the core i freed up for just one gpu, it is fairly idle.

 

[Tyrinon]

A while back I had attempted to run two wu's on a gtx 260, but it did not work out very well. I believe it was after this that I learned that you needed a 400 series card to run two wu's. As for my 460's, all running two wu's at the same time is double the time to run them instead of one. No idea if points change, have not checked that yet as I just set my 460s to run one wu each. Hopefully this gives them a break to cool down some. As it is now, I am only running two of them in my i7 b/c the third (top) card was pushing upper 90's. I have been reading about water cooling gpu's, but not sure it is worth the hassle and added expense.

As for the two wu's at once, I believe the info was in a thread, but I do not know which one. Perhaps someone can put it in the sticky for easy reference?

 

[Eroc]

Yeah. My 460 runs 2 WUs no prob. Very similar times as a single (slightly longer). I think it is a Fermi thing.

 

[eaglescouter]

When the Fermi optimizations are finished by the Lunatics folks we should see some significant production gains from the FERMI CUDA cards. The current software is not really optimized.

My GTX570 is now running 2 units at a time, I have also run 3 units at a time.

Issues:
1. I needed to free up one core of my CPU to feed the GPU, this increased the RAC for the machine.
2. How many units your GPU can run is partially based upon the amount of ram on the card.

 

[Eroc]

Now I have a question on running multiple video cards in the same comp. I just moved a 260 into my main rig so I can work on upgrading my second rig. They both ran the new versions of Lunatics fine when in separate boxes.

Now that they are in the same box the 260 (or gpu0 as the program sees it) states: (not used) next to the 260.

How can I get it to run on both cards. I have heard of people doing this with mixed cards I thought, but I am stuck atm.

Edit: Figured it out. Had to add cc_config.xml to my boinc folder with the correct options. If anyone else has questions about doing that let me know!

 

[razorface]

I started seeing the "setiathome_enhanced 6.10 (cuda_fermi)" wu's after installing the new 3.8 unified installer, but not with the 3.6. Is 3.8 the fermi optimized app? Or if not, I wonder why these work units weren't seen before? The readme for the new version, whice adds ATI support, is here

 

[Eroc]

Yes 3.8 is fermi optimised. It does work with older cards as well. My 260 seems to be running faster as well. Even doing 2 instances along side my 460. Still have not been able to fill my que on that box since the outage 2+ weeks ago.

 

[razorface]

Still have not been able to fill my que on that box since the outage 2+ weeks ago.

Same here, new cards and no work for them

 

[CryptokiD]

would my gtx 550 and gtx 560 benefit from running multipule work units on each gpu? also am i better off disabling cpu crunching and let the cpu core feed the gpu?

 

[Eroc]

Well, the 560 will definitely benefit and the 550 probably as well. It has very few cores but it is a Fermi and can thread well. I'd need to check out it's architecture more.

If they are in the same box you have to run the same amt of units on both cards (i.e 1 each or 2 each). Even if the 550 does not do as well the 560 will more than make up for it.

If they are in different boxes definitely do 2 on 560, as it has architecture similar to and better than the 460 and the 460 does well with it. Then see what happens on the 550.

I have 2 cards in my comp and I have just one cpu not crunching and it works fine.

 

[eaglescouter]

I reserve one core to feed FERMI GPU's.

 

[Eroc]

Crunching all CPU since I cannot get enough GPU units. When I do get a GPU unit it knocks it out so quickly, even with the CPUs crunching, that it doesn't affect them that much.

 

[Wicked_Pixie]

Edit: Figured it out. Had to add cc_config.xml to my boinc folder with the correct options. If anyone else has questions about doing that let me know!

I'd like to know how. I can't even find any .xml files in the \ProgramData\BOINC\projects\setiathome.berkeley. edu folders or any folders for that matter. I already changed the settings to show all hidden files...

Does that mean I should just create one and drop it in?

Thanks

 

[Eroc]

I'd like to know how. I can't even find any .xml files in the \ProgramData\BOINC\projects\setiathome.berkeley. edu folders or any folders for that matter. I already changed the settings to show all hidden files...

Does that mean I should just create one and drop it in?

Thanks

Yeah you would need to create the cc_config.xml with the necessary instructions...BUT before you do that...Are you trying to run seti on multiple video cards in one box, multiple seti instances for one video card, or both?

 

[Wicked_Pixie]

It's for running multiple seti instances on 2 GPUs (HD6970) on one box.

Here's my observation so far:

1. With nothing changed GPU usage is @ 40% on one and 22% on the other.
2. When I freed up one core, both GPU usage is in the mid-40%
3. When I freed up 2 cores, GPU usage is in the mid-70%.

Also, GPU usage is highly dependent on WU. On AP WUs, GPU usage drops to mid 40's.

So, will I benefit from running 2 instances/per card? By allocating a core for each card, I'm seeing 1/3 reduction in time to complete.

All running stock atm, will oc it when I figure out how to optimize both...

 

[Eroc]

I would say that 2 per gpu is probably better, but I am not 100% sure on ATI cards. But doing the math if it takes 1/3 longer to do 2 times the wu's I'd say do 2 per gpu.

i.e. 10mins at 1/gpu or 10(1+1/3)= 13.33mins at 2/gpu. That means in a little more than 26mins you will have 4 units done at 2/gpu. In that same time with 1/gpu you will have 2 done and a little over 1/2 through the 3rd.

As far as the cc_config.xml goes it needs to read:
<cc_config>
<options>
<use_all_gpus>0</use_all_gpus>
</options>
</cc_config>

To run multi units per gpu you need the app_info.xml file under the Lunatics ati app to read (default is 1):
<count>0.5</count>

...I think. I have an Nvidia gpus and I am not familiar with the ATI wording in their optimizer.

 

[Wicked_Pixie]

Thanks eroc! Will drop in app_info and cc_config files later. Thanks for doing the math for me. I was never good at it.
Hopefully, it is linear...

Yeah, it seems more work involved with optimizing AMD cards for S@H.
http://setiathome.berkeley.edu/forum_thread.php?id=64588#1120079

Oh, and didn't mean to hijack the thread. I'll start a new thread for AMD cards, if there are any changes.

Thanks again for your help.

 

[Eroc]

Glad to help. An AMD thread would be great.

Also, as long as you have Lunatics installed, you should have the app_info.xml already in your seti folder (apps dir). You just need to edit it.