[HankB]

It's not working very hard yet.

Code:

Every 2.0s: nvidia-smi -a                                                        Fri Apr 15 10:33:22 2011


==============NVSMI LOG==============


Timestamp                       : Fri Apr 15 10:33:22 2011

Driver Version                  : 260.19.29


GPU 0:
        Product Name            : GeForce GTX 460
        PCI Device/Vendor ID    : e2210de
        PCI Location ID         : 0:1:0
        Board Serial            : 650381377
        Display                 : Connected
        Temperature             : 55 C
        Fan Speed               : 30%
        Utilization
            GPU                 : 25%
            Memory              : 3%

Apparently the new NVIDIA Fermi cards are not yet that well supported under Linux. I tried to get a SETI GPU app running but all it did was produce errors. As near as I can tell, the Linux CUDA app for SETI does not yet support Fermi. (I found some threads, but they all seem to end w/out any closure.) I don't have the resources to run XP in a VM to support FAH so that's out.

If you're aware of any other Lunux/GPU DC projects please speak up!

At least there's the possibility that GPUGRID supports Alzheimer's research. That's possibly a self serving choice. I think. Can't recall the details.

And there is an Overclocker's team (with one active member aside from myself. )


Edit: this is not really solved, but that's the only way I could post it.

I did find that if I suspend Rosetta, GPUGRID goes up to 8% CPU and keeps the GPU at about 90%. I'd like to find out how to alter scheduling so that GPUGRID gets 8% and the remainder of the system gets what's left (including about 390% for Rosetta.)

[Adak]

Sounds like some experimentation with priorities are needed. Perhaps raise the GPU client, (or core), and lower the Rosetta client (or core), one notch.

If you were to post this on the BOINC forum, you'd get a lot better answers as to where to use a linux GPU for one of their projects.

Congrats on getting it working though. I recall the work it took to get my Linux GPU folding in FAH, but I'm a near total noob in Linux.

[HankB]

Hi Adak,
Thanks for the tips. Rosetta is already at maximum 'nice' (lowest priority) and I did try boosting GPUGRID to get it to work better and with no improvement.

I've been searching and asking at a lot of forums with no useful answers. The best suggestion so far is to reduce Rosetta to three cores. Unfortunately that takes 100% of one core away from Rosetta to provide 8% of one core to GPUGRID. Not IMO a good tradeoff.

I also found that the additional GPU made my OC unstable so I'm back to stock speed until I can address that.

I looked at FAH and I recall there was no Linux client but that some ran it in a VM running XP. That didn't excite me at all. The thought of trying to balance CPU vs. GPU usage between a Linux host and Windows guest sounded like a fools errand to me. I also looked at SETI but as near as I can tell, there is no Linux client for Fermi cards yet. (I did get it running but every WU errored out in 40 seconds.) GPUGRID OTOH was brain-dead easy to get going, making it an easy choice.

[Adak]

You can definitely run a GPU FAH client, in native Linux. It's not supported by Stanford, but we've been doing it for years - even I do it, and I'm quite the noob in Linux, believe me.

Would it work with your Fermi card however? I have to plead ignorant on that question. I killed a GPU card with folding, and now that the manufacturer is out of the GPU business (yes, BFG), I can't get any further repairs or replacements. So much for "lifetime" warranties. I have retired my GPU from further folding. It's nowhere near as strong as your Fermi card, however.

Knocking down to a lower core number does sound off the mark.

I would expect that a project named GPUGRID, would have an easy way for GPU's to join in the folding. They'd be crazy not to, wouldn't they?

So what does GPUGRID projects work on?

[HankB]

Originally Posted by Adak

You can definitely run a GPU FAH client, in native Linux. It's not supported by Stanford, but we've been doing it for years - even I do it, and I'm quite the noob in Linux, believe me.

Can you point me to where I can find a Linux client? All I can find is the official line that "we hope to support Linux as well" I also found info on running under Wine, but not recent.

Would it work with your Fermi card however? I have to plead ignorant on that question. ...

Worth a try if I can find it.

I would expect that a project named GPUGRID, would have an easy way for GPU's to join in the folding. They'd be crazy not to, wouldn't they?

So what does GPUGRID projects work on?

Indeed! But not all DC projects even support Linux so I make no assumptions. They state "Volunteer computing for biomedicine." More info at http://www.gpugrid.net/science.php. (Yet Another Folding Project AFAICT.)

thanks,
hank

[HankB]

I'm happy to report that I have a partial solution to the GPU usage issue. I've switched to a real time kernel and with that, I can run 4 cores Rosetta (which get near 100% each) and a single instance of GPUGRID which keeps the GPU usage at about 44%. It's not fantastic, but it's better.

Now I wonder if I could run two instances of GPUGRID on the GPU and boost performance. Most comments about that suggested it would not provide a performance boost, but those are probably relevant to situations where the host OS is already keeping one instance busy. I guess I should let RAC stabilize a bit so I can evaluate the results of further changes.

Other good news: I boosted CPU-NB voltage two steps and that seems to have stabilized my overclock with the new video card. I'm back to crunching Rosetta at a higher speed.

Other not so good news. Nvidia has not yet provided overclocking support for their Fermi cards on Linux.

[Adak]

RT kernel and a solution for the overclock problem too? Well done!

I'm not sure how much another folding instance on the Fermi card is going to need from the OS. Something, surely. With the RT kernel, I can only say you're the test case, my friend.

I thought the unsupported Linux gpu set up was a done deal - but the newer cards and FAH cores for them, has turned over the whole apple cart. I read through most of this 23 page thread on it, and frankly, I would not try and fold on a Fermi card, with FAH, at this time. You're clearly more experienced with Linux (easily!), than I am, but this just sounded so iffy and brittle:
http://foldingforum.org/viewtopic.ph...4673&start=330

and they're still working with wine and wrappers and etc. --> uffda! That's not as good as I had expected.

Looks like GPUGRID is doing a lot of similar protein analysis, like Rosetta, but concentrating on the GPU niche of the market. They have a good web site and video's. The way that these molecules, move around to do their function - it's just amazing.

For high end graphics cards, the Linux world is a sparse place still.

Other not so good news. Nvidia has not yet provided overclocking support for their Fermi cards on Linux.

Keep bugging 'em! Squeaky wheel gets the grease.