Nvidia Multi-GPU support

Come join the... uh... er... fold.

Moderators: just brew it!, farmpuma

Re: Nvidia Multi-GPU support

Postposted on Fri Jul 18, 2008 6:54 pm

I've been running the GPU client on two machines, one under Vista 64 and one under XP Pro , for a few days now and would like to offer a few observations. Under the XP setup , the client would completely lock up one core of my dual-core machine, whereas with the Vista 64 setup, the maximum load I would see is between 15% - 30% on one core of my quad, and that's with two GPU clients running (8800GT and 8500GT). Seems pretty obvious that Vista is much better at handling the GPU client than XP does. I'm going to try to dig up a copy of Vista 32 to experiment and get some more numbers, unless of course, someone on TR is already running Vista 32 with the GPU client and can offer his/her thoughts.
Fold! And I don't mean your clothes!

Do you have a favorite gerbil recipe? Please share with the TR community!
flybywire
Gerbil Jedi
 
Posts: 1883
Joined: Wed Jun 16, 2004 2:28 pm
Location: Springfield, VA - USA

Re: Nvidia Multi-GPU support

Postposted on Fri Jul 18, 2008 7:49 pm

The more I read the more I'm thinking v1.07 was pulled and is not public any more. They did have a direct download link but the file from Stanford is no longer hosted. My log files say v1.06 even after deleting the FahCore11.exe file and having it re-download the core. So trying to change affinity via Environment Variables isn't going to work just yet. Only once v1.07 is out.

But folding on multi GPUs works perfectly other than they share a CPU core (unless you manually change affinity via task manager as each WU comes in). Just a little lower ppd for now. No crashes of any kind. Been running for 2-3 days with this multi GPU setup on XP.
i3-530 | HR-01 Plus (passive) | DH55TC | 4GB Kingston DDR3 | Toshiba 250GB 2.5" HDD | Mini P180 | picoPSU 150-xt w/102w brick | 21w idle
Pegasus
Gerbil First Class
 
Posts: 161
Joined: Fri Jun 30, 2006 1:13 am

Re: Nvidia Multi-GPU support

Postposted on Fri Jul 18, 2008 11:49 pm

The SMP loss of which I write is the points you'd get running the SMP core. If you could get 1 CPU core to handle feeding both GPU2 clients, you might be able to run the SMP core. Apparently you'd have to be running Vista for that to succeed and add to your output, though.

I wonder why you're not getting more output from your 8800GT 256 cards, though. I'm getting ~4700 PPD from my one, and it's fed by a 2.0 GHz A64-3200. I mean to test what happens if I run the console client with it, which I'm betting will be a loss of output, but I've got other things ahead of that for now.

I think I read something about the v. 1.07 core and its disappearance or existence on the foldingforum.org site, but am not sure where. The gist was that, vague memory problem here, but there was something about the 1.07 not being updated in version number at least at first, and I thought Pande said that was going to be fixed. It may well have already been fixed before I ever downloaded the GPU2 client last week, though, since the thing had been out by some time in June.

I found this thread which says some stuff about it, but was closed by 7im, but he calls the thread a duplicate and implicitly points to a thread linked in it, which is this one, which scolds saying v. 1.07 isn't available now and whatever. Some others said the installer was updated in the first linked thread, though, but maybe you have to remember to delete files from \Documents and Settings\username\Application Data\Folding@Home-gpu area or something.

I'm more confused about things the more I read, because people espouse ideas and moderators or others don't always correct them, and then there are contradictions in other threads. Pretty much as you said.

In the end, were I in your situation I'd uninstall the client, then search for FahCore_11.exe and if any is found at minimum rename the file(s), and then download another copy of the installer and re-install from it, and see if it then begins reading the required environment variables (which in XP can be set in System Properties->Advanced), as mentioned in this (currently 6 page long) thread, A new Nvidia core (v1.07) for users w/ multi-gpu boards.

Flying Fox: The cores are launched by the executable I mentioned. Changing that thing's assigned CPU core is the only plausible way to do it. One could assumedly find or if need be write a utility to sleep and every minute or so wake up to make sure all cores are where they ought to be, but we'll cross that bridge when we come to it if the environment variables are officially not working.

Someone in a thread here in the last day or so said something about a program he found to assign affinity, but I'm out of steam now so others can find it.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: Nvidia Multi-GPU support

Postposted on Sat Jul 19, 2008 12:31 am

Ragnar Dan wrote:Flying Fox: The cores are launched by the executable I mentioned. Changing that thing's assigned CPU core is the only plausible way to do it. One could assumedly find or if need be write a utility to sleep and every minute or so wake up to make sure all cores are where they ought to be, but we'll cross that bridge when we come to it if the environment variables are officially not working.
I usually just manually set the worker EXE's directly without bothering the "main" executable. ;)

Ragnar Dan wrote:Someone in a thread here in the last day or so said something about a program he found to assign affinity, but I'm out of steam now so others can find it.
Fuzzhead found it, it's Process Lasso.
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24440
Joined: Mon May 24, 2004 2:19 am

Re: Nvidia Multi-GPU support

Postposted on Tue Jul 22, 2008 7:23 pm

That program works for the time being. Thanks for finding it fuzzhead.
Been using it for a few days and have it set so they run on both cores. So it keeps them from resetting to the same single core.
i3-530 | HR-01 Plus (passive) | DH55TC | 4GB Kingston DDR3 | Toshiba 250GB 2.5" HDD | Mini P180 | picoPSU 150-xt w/102w brick | 21w idle
Pegasus
Gerbil First Class
 
Posts: 161
Joined: Fri Jun 30, 2006 1:13 am

Re: Nvidia Multi-GPU support

Postposted on Tue Jul 22, 2008 10:27 pm

I was a bit tired when I wrote my bit and the part I wrote to FF is not quite logical. I'm still too lazy to look up whether there's a way to force a program's launched process to run on the same core, and it doesn't matter unless Stanford does it anyway.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: Nvidia Multi-GPU support

Postposted on Tue Jul 22, 2008 10:44 pm

Ragnar Dan wrote:I was a bit tired when I wrote my bit and the part I wrote to FF is not quite logical. I'm still too lazy to look up whether there's a way to force a program's launched process to run on the same core, and it doesn't matter unless Stanford does it anyway.

The Process Lasso program that Pegasus and I were talking about is a utility to force specified EXE's to specified affinity to make them run only on certain cores. And Pegasus said was that with GPU2 clients, they need 100% of a core to work. His problem was that his 2 GPU2 clients, without any intervention, always seem to want to go back to one core. That makes the 2 GPU2 clients fight for the CPU cycles and is reducing his ppd. He was looking for a tool that can force the 2 GPU2 clients to run on separate cores to fully maximize those ridiculous ppd that can be unleashed by the GPU2 clients. 8)
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24440
Joined: Mon May 24, 2004 2:19 am

Re: Nvidia Multi-GPU support

Postposted on Tue Jul 22, 2008 10:50 pm

The summary wasn't necessary, though it may help other readers. I just wasn't clear on what worked for assigning executable Fah cores to CPU cores (whether what Stanford already had with environment variables did anything). I've mostly lost interest for the time being, and for those who need it there's obviously the Process Lasso utility, so it doesn't matter much. I have to assume Stanford will eventually fix things if there's enough demand for it. I'd like to see what difference Vista makes, but I'm not on the edge of my seat over it. :wink:
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Previous

Return to TR Distributed Computing Effort

Who is online

Users browsing this forum: No registered users and 1 guest

cron