Most of our servers for years have been hooked to FC (fiber channel) attached disk with /tmp down localdisk. This has worked well but now due to the complexity and number of servers involved it's easier to just buy servers with localdisk (12 disk 15,000k SAS disk). We have been configuring them with 1 large array for data and a separate non raided disk for /tmp. We've noticed early on that having /tmp in the same volume as /data is very damaging as it will easily IO bound a busy server and the sequential writes of /tmp don't play well with the random IO of /data. The problem now is that even with /data and /tmp being on separate disks, heavy IO down /tmp (large sort files) can still IO bound the controllers (3ware, MegaRaid LSI) and cause contention to /data. Any thoughts on the best way to avoid this situation? I'm half tempted to somehow plug the one disk directly into an onboard scsi controller (if available) and not going through the raid controller so I can use it for /tmp. The other option is to buy a flash PCI card and use it for /tmp. It seems to me that the SCSI controllers should be able to handle this without affecting the other controllers. I tried turning off writeback caching on the /tmp disk but the behavior was unchanged. Any other ideas?
No announcement yet.
/tmp performance on RAID cardPage Title Module
Move Remove Collapse