From: <wk...@bn...> - 2011-05-11 05:53:44
|
Greetings: We are testing MFS and have been very impressed so far. One discussion that we are having internally is the desireability to use RAID (either Raid0, Raid1 or Raid10) on chunkservers. Obviously with a goal of 2 or more, its not a protection issue, which is the classic RAID scenario. as we see it, RAID on chunkservers has the following to recommend it. + stability: Assuming RAID1/10, if a drive dies the chunkserver doesn't immediately fall out of the cluster, and forcing a need to rebalance. We can fail it deliberately at a time of our choosing to replace the drive. + less disks: with RAID1/10 we might feel better using a goal of 2 instead of 3. + speed: RAID0 would be faster outright. RAID1/10 may provide faster on reads and slightly slower on writes (RAID 10 much better). The reasons against would be: - stability: with RAID0, the loss of a single drive kills the entire chunkserver. - stability: another layer to fail (the MD layer). - cost: increased power consumption and of course 2x or more the number of drives. We would be using Linux SoftRaid (MD driver) rather than hardware raid, so the card cost is not an issue. - speed: slightly slower writes on RAID1/10 So what is the consensus of the more experienced users? Are you using RAID (0,1,10 or others) on your chunkservers? Are we missing something on the above analysis? -bill |