OK, here's my update.
I finished that 364 point WU tonight and got a crappy 2126 to replace it. Meanwhile, I think it's taking about 8-9 minutes per frame (100 total frames) for the two 3025 Work Units I've gotten so far on the Linux SMP client running under VMware, which is free for the downloading as Hotdog
mentioned in the original post. (Since he's awesome.
That's 587 points in probably less than 15 total hours, or 939.2 points per day. I'm running an AMD Opteron 165 (1800 MHz stock) upclocked to ~2581 MHz and stable enough to use for work. It will run faster, but eventually at some point it will mess up the file system somehow and not want to boot, and then I have to run expensive software to retrieve such things as email files and text notes I haven't backed up. Anyway, this is an incredible increase in daily point production.
And what's more, I can leave my Windows Folding clients running at the same time. They slow down to a crawl, but they keep going. As Hotdog
points out above, the VMware system gets somewhere in the range of 80-some percent of the CPU. And you have to give it enough RAM to run the folding software speedily. While I won't run it while I'm working on the machine, in the evenings and over night I do run it, now. I'm running it as I type this in Windows, as a matter of fact. My first WU, which I stopped for most of a day to let another finish and because I can't run it during working hours, finished with only 40% of the alotted time remaining for it. That's because they want the WU back in 1 day at best, and 2 days at worst. Since getting the VMware client going is fairly simple to get up and running, and you can make scripts to quickly run the folding client, it's not much effort to get moving on this. There are stupid things you will want to play around with, though. I had to change the number of processors VMware thought I had, and also since I don't have a floppy drive installed I changed that to using an image file or some such thing to emulate one. I should probably have kept a log of the things I did, but I was dubious when I started this experiment. Dummy.
If you have a machine that can run this software, which includes any AMD dual-core and all the new Intel C2Ds and some others, you will do your own and your team's output a great improvement if you begin to run it. If you have any dedicated boxes meeting that requirement, and now that I check the memory
utility on that "box" (heh) it shows it's using about 450 MB for itself and the rest in cache, so next I'll be experimenting with how little RAM I can allow the VMware software while maintaining good performance, and it may well be that machines with under 1 GB can run it well. I did give it 8 GB on a spare drive I have, but so far it has used under 2 GB of that, and I'm sure a less full-featured Linux client would have an even smaller footprint.
The only odd thing is that the Linux clock seems to lose almost 1/5 of the time passing. Its clock is now 25-26 minutes behind the real time after running for 1 hour and 56 minutes. Oh well. So don't trust its frame times, use Windows' clock instead. I do run it fullscreen over night since I assume it gets more of the CPU that way, though, so I can't time that except by going back and calculating the actual time used. I just know the thing offers killer points, though, so I'm running it.