Interesting stuff.
A while back I had a K6-2 450 and K6-2 400 folding, and the 400 had 64Mb EDO and the the 450 had 64Mb SDRAM, but the 400 was running rings round the 450. It was doing 1.5 units to the 450s 1. Now the only advantage the 400 had was a 15G UDMA66 drive on a UDMA33 interface, vs a DMA mode 1 drive on the 450, I think it was something like 3MB/sec transfer vs about 20MB/sec transfer. So I guess storage speed can make quite a bit of difference.
Now while it's a good idea to blam in the RAM in older machines to use a RAM disk and get the production up, you may find some socket 7 based systems run slower. This is because some of them can only cache 64Mb. This is usually only a problem for boards that can only run 66/75/83 FSB, anything with 100Mhz FSB can cache at least 256Mb I think. This does not apply with K6-3/+ and K6-2+ chips though which have onboard L2 and cache independantly of the mobo. So try it and see what happens for old rigs, but keep a careful eye to make sure that production isn't hurting instead of increasing.
Beware of folding directly off USB drives, or other flash media. They have a limited number of write cycles. The early ones had life expectancies of only 10,000 writes, this improved greatly a couple of years ago to 100,000 and has gone into the millions for brand name flash parts. However some of the cheaper ones around may have low life expectancies, so folding off them directly may "use up" the part in a short space of time. A fast machine could easily change the log file 1000 times in a day, so for earlier type flash parts, you might only expect 10 days use out of it before you get errors. For medium old parts, and cheaper current parts that would be about 3 months. That's if the log file is the most frequently written file. If it writes anything else more frequently, you might see a real short lifetime for the part. So setting USB drives to run F@H from RAM and to back up to the USB drive every hour or so is a better idea. If you really want to stretch the lifetime of your USB drive, then make the backups in sequential directories until the drive is full, then start overwriting the earlier ones. That way you don't just hammer a single area of the drive all the time.
If we accept that storage performance has a significant impact on folding speed, but that the RAM drive option isn't viable for all machines. Then one might consider for "spare part" and older dedicated rigs, the use of RAID 0 arrays to enhance storage performance. I'm not saying you should go out and buy $150 worth of RAID card and $200 worth of drives for your P3 rig of course. Merely that one could consider applying older hardware in a more efficient manner. If for instance you've got rigs running off old 2G and 4G HDDs, you might consider doubling them up for better performance, with a cheap UDMA33 RAID card, that you can pick up for peanuts. Under linux the software RAID works quite well I beleive, to use off the internal interfaces. Also when buying older socket A boards for folding rigs, you might want to consider the RAID versions of A7Vs and some Abit boards. The prices on those, used, are rarely much more than boards without RAID, now they're getting longer in the tooth. Should you not find a cheap older RAID card available, you might even find performance improved by using 2 drives still, boot OS and swap on one drive, and the folding dir on a second drive on the other interface. So that the drive doesn't have to keep re-seeking between the swap file and the folding dir. Another hint is that size isn't everything. I've got some 420Mb drives that are faster than 2G and 4G drives I've got. Of course there are faster 2G and 4G ones around, but just don't necessarily assume your larger drives are faster. There's actually some blazing fast 4G-6G-8G UDMA33 drives around that will cream anything short of a good 40G or so UDMA66 drive with 2Mb Cache.
I'm thinking actually, that it might not really be sustained transfer speed that matters much. Not sure that F@H is moving large chunks most of the time, I think it's just writing and reading a lot of small pieces of data quite frequently, the larger reads happen at beginning and end of frames most likely. It's probably more down to seek performance, latency and burst transfers. In seek performance, there's quite a few older drives around with 9ms full seek times whereas even some new drives these days have 12ms times.
Anyhoo, whether you can find and use extra RAM for RAM drives, find ways of increasing your storage speed, or find storage that works better for F@H hiding in your closet, good luck with the optimisation, and fold on,
Road Warrior