by blerton
I'm a long time fan of PCLinuxOS - I now have more than 70 machines running it. They contribute to distributed computing projects - Einstein@Home, Milkyway@Home and LHC@Home. The first two (for some time now) have applications that run on GPUs - both nVidia and AMD/ATI.
In particular, I have 12 machines with HD4850 GPUs that have crunched for MW since early 2009 using Windows XP and the Catalyst 12.1 drivers, which were the last ones with the OpenCL support needed for the MW GPU tasks. I have long treasured the hope that one day I could work out how to run the same tasks under PCLinuxOS. I worked out how to do that about two weeks ago, and have so far converted 9 of the 12 without problem.
I've been a very infrequent lurker on these forums for several years and by pure chance, I stumbled across this quote: "When you get a Synaptic fglrx update notice you will want to install either the fglrx-legacy packages or the fglrx-current packages depending on your video card." It caught my attention. I had seen the Synaptic update notice about fglrx-current and fglrx-legacy some time ago and that had prompted me to go hunting on the AMD website for more information. From that, I knew I needed the legacy driver, and I read that to get OpenCL support for the GPU, I needed to install the catalyst driver first and then the AMD APP SDK. The only problem was that whilst fglrx-legacy was in the repo, there was no SDK that I could find.
While I was contemplating what to do, one of the 12 WinXP machines had a disk failure. I decided to install the latest KDE-minime, set up the fglrx-legacy driver, and test things out. No problem doing that, but trying to run a MW task gave an immediate failure. So I downloaded the SDK (Version 2.7) from the AMD website, unpacked it and ran the install script that was included. The messages looked all very promising, so after a quick reboot, I tried another task. I was very pleased to see immediate success. I also found the aticonfig command, and have used it to enable overdrive, and then set core and memory frequencies to my liking. In the nine machines I've converted so far, the GPU crunch times are comparable to, or even slightly better than, those achieved under WinXP.
The biggest difference is in the crunch times of Einstein tasks running on the CPUs. Ten of the 12 machines have dual core CPUs -- mainly e6300 Wolfdales -- and the two CPU tasks running alongside the GPU task are significantly slowed (>30% longer to crunch) under WinXP, compared to a similar machines with no GPU. Now, under PCLinuxOS, there is no (or very little) such slowdown so this is another big win. Needless to say, I'm quite delighted with the outcome of a hard disk failure.
I just wanted to thank all those responsible for PCLinuxOS. I started using it in 2007, and it has always been a very good fit with the way I like to do things. I've sampled other distros, but have never found anything else that can even get close enough to stand a chance of luring me away! I've been contemplating purchasing a couple of HD7870s, so I'll probably do that now that I know that the OpenCL runtime can be had by installing the SDK from AMD. Is there any likelihood that OpenCL support might be included in the repo sometime soon? I've been using the CUDA support I found there for nVidia cards for some time now. I actually have around 20 machines, mainly with GTX650s, all running PCLinuxOS of course!
|