Re: Machine Configuration

Edward Bidinotto wrote: > > Enio Roncarati wrote: > > > > Hi everybody. > > > >...we want to find the best configuration for the disk > > array... > > Each array has 6 HZ40 from DIGITAL, that support bolth > > raid level 3 and 5. Do keep in mind the fact that the HZ40 has an intelligent controller than self optimizes between raid 3 and 5, That's why it's called a raid 3/5 device. Therefore, you will have less to worry about than you would with other vendor's products. > We initially set up our Informix instance out on the raid devices and > turned on Raid level 5. Our database was not maintained for quite some > time and we had a very bad problem with our tables being extremely > fragmented...lots of extents... > > Anyway, this does not work well with Raid 5...mainly because you are > performing striping with parity across the Lun.. Striping w/parity across a lun is not itself the cause for performance problems, it's just using parity (since that is the extra disk write). Putting parity accross a lun actually allows for parity to be written independantly (since on another disk) of the data write. > Since all of the tables > were extremely fragmented we took tremendous hits during the writes. This is a performance problem outside using RAID, of course. I would recommend addressing fragmentation before implementing new components (RAID or otherwise). > After talking with the Raid vendor..here's what we may try if we go > back to Raid. > Use Raid 0+1 for most of the heavily hit tables, temp, root, logs, etc > etc.. Best Performance = Most Expensive = Most Reliable = Raid 0 + 1 > Then use Raid 5 for the rest. Well, that depends on what you're looking for. Generally, in a database, you'd rather have Raid 3, which is for bandwidth intensive environments. Raid 5 capitalizes on high I/O rates. Of course, the size of the disks themselves will play a large factor in this. (You can get more I/O out of 2 2GB drives than 1 4GB drive. You'd need to be very OLTP to want raid 5 over raid 3. > Raid 0+1 is mirroring and striping as > well so you'll have good recovery capability if a disk goes out. Actually, 0 + 1 is striping and mirroring, but I think you meant that. Optimum implementation is: Primary Mirror (not striped) (striped) 0--------m--------0 0 = physical disk 0--------m--------0 m = mirror pair 0--------m--------0 0--------m--------0 0--------m--------0 In this configuration, you could be online losing any single of the mirrored pairs and wouldn't go down unless both pairs of the mirror failed. > You > may also want to keep another raid disk around in case one fails. Of course. Otherwise you don't have nearly the capability for fault tolerance. This is where Digital fills in well, with the hot-swappable drives...which Digital Unix supports w/o a reboot (and NT does NOT). Of course, other vendors do too, I mention this because the original post involved a Digital system.