... Is the PERC 5/i dual channel? If so, are 1/2 the drives on one channel and the other half on the other channel? I find this helps RAID10 performance when the mirrored pairs are on separate channels. ... With the SAS controller (PERC 5/i), every drive gets it's own 3 GB/s port. ... Your transfer rate seems pretty good for Dell hardware, but I'm not experienced with SAS drives to know if those numbers are good in an absolute sense. Also, which driver picked up the SAS controller? amr(4) or aac(4) or some other? That makes a big difference too. I think the amr driver is "better" than the aac driver. .. The internals of the current SAS drives are similar to the U320's they replaced in terms of read/write/seek performance, however the benefit is the SAS bus, which helps eliminate some of the U320 limitations (e.g. with Perc4, you only get 160 MB/s per channel as you mentioned). It's using the mfi driver... Here's some simplistic performance numbers: time bash -c "(dd if=/dev/zero of=bigfile count=125000 bs=8k && sync)" Raid0 x 2 (2 spindles) ~138 MB/s on BSD Raid5 x 4 ~160 MB/s BSD, ~274 MB/s Knoppix (ext2) Raid5 x 6 ~255 MB/s BSD, 265 MB/s Knoppix (ext3) Raid10 x 4 ~25 MB/s BSD Raid50 x 6 ~144 MB/s BSD, 271 MB/s Knoppix * BSD is 6.1-RELEASE amd64 with UFS + Soft updates, Knoppix is 5.1 (ext2 didn't like the > 1TB partition for the 6 disk RAID 5, hence ext3) Seems to me the PERC5 has issues with layered raid (10, 50) as others have suggested on this list is a common problem with lower end raid cards. For now, I'm going with the RAID 5 option, however if I have time, I would like to test having the hardware do raid 0 and doing raid 1 in the os, or vice versa, as proposed in other posts. Also, I ran a pgbench -s 50 -c 10 -t 1000 on a completely default BSD 6.1 and PG 8.1.4 install with RAID5 x 6 disks, and got 442 tps on a fresh run (the numbers climb very rapidly due to caching after running simultaneous tests without reinitializing the test db. I'm guessing this is due to OS caching since the default postgresql.conf is pretty limited in terms of resource use). I probably need to up the scaling factor significantly so the whole data set doesn't get cached in RAM if I want realistic results from simultaneous tests, but it seems quicker to just reinit each time at this point. On to some kernel tweaks and some adjustments to postgresql.conf... - Bucky
