Raid 10 question

Hello there, Question on Raid 10: Our database has one big table that represents around 50% of IO. The rest is evenly divided between logs and the other tables. We have 10 physical disks to play with. After mirroring, we are left with 5 logical disks. We have 2 choices: 1. Create one wide stripe set (5 disks) and put everything there. 2. Create two stripe sets (3 disks + 2 disks) and put the offending table on one set and everything else on the other. Solution 1 gives better throughput, but the offending table is never read sequentially. It is mostly read by Read Ahead (happily running at 64 / 32). The problem with this solution is too much traveling by the disk heads. Solution 2 has less throughput, but the disk heads will travel less. I know only a real test will give the correct answer, but I do not have that luxury. I have to make a recommendation in the blind. Thanks for any opinion. -- Bashar Chalabi CTL, London

Personally I'd go with the single wide stripe or make that a stripe of 4 pairs and use the last pair for logs, temp and root. Hmm 3-stripe and one pair for root & temp and the other for logs? Oh foo now I've made your decision twice as hard...... Hmmmm. Single 5-pair stripe with everything on it or 3-pair stripe with two singletons. You can also put all the relatively static tables onto the root and log disks to minimize load on the stripe. Art S. Kagel Bashar Chalabi wrote: > > Hello there, > > Question on Raid 10: > > Our database has one big table that represents around 50% of IO. The rest is > evenly divided between logs and the other tables. We have 10 physical disks > to play with. After mirroring, we are left with 5 logical disks. > > We have 2 choices: > > 1. Create one wide stripe set (5 disks) and put everything there. > 2. Create two stripe sets (3 disks + 2 disks) and put the offending table on > one set and everything else on the other. > > Solution 1 gives better throughput, but the offending table is never read > sequentially. It is mostly read by Read Ahead (happily running at 64 / 32). > The problem with this solution is too much traveling by the disk heads. > > Solution 2 has less throughput, but the disk heads will travel less. > > I know only a real test will give the correct answer, but I do not have that > luxury. I have to make a recommendation in the blind. > > Thanks for any opinion. > > -- > Bashar Chalabi > CTL, London