>>31 My test setup:
VIA EPIA800 V, 128MB PC133 RAM
Probably the 2.6.4 kernel..
Silicon Image ATA133 controller; 200GB "Diamond Max 10" Maxtor, 7200 rpm, 8mb/cache. hdparm'd properly.
Apache 2, wu-ftpd.
Throughput tests involved single and multiple concurrent runs of bonnie+, iozone, tar -f - >/dev/null; and wget from other machines (over 100BaseT). I don't have the logbook with the raw timing data around, though. With the wget and tar tests, the files it spidered through were a fansub archive, typically 134-300MB each.
ext2 came out in front, although bad hdparm options affected it greatly (large readahead didn't improve things, in spite of just serving large files). ext3 was next, then reiser3 (I guess from it primarily being a read-only test), then the rest about a step further down that did poorly on the head-trashing tests.
We had a burn party just a few days ago, actually, and the machine set up by the abovementioned "XFS weenie" did
exceedingly poorly. It took me over 6 hours to leech ~80GB off that system, whereas the ext3 machines belonging to another guy had no problems maxing out. (The leeching record of the day was like 240GB in <3 hours. Yes, he had 1000BaseT..)