On Thu, 30 Jun 2005, Matus UHLAR - fantomas wrote:
BTW I think I've found another reason why RAID should not behave better with squid: even SCSI disks behave better on sequential reads. So, fetching a file sequentially from a drive (case of multiple filesystems on multiple drives) should be faster than fetching some parts (size of a stripe) from one disk and some parts from second disk
Luckily for RAID0/5 the stripe size often practically eleminates this drawback as the percentage of small reads crossing a stripe boundary is fairly small.
For larger files (but not so for Squid) with little or no other I/O activity at the same time the above really gives a benefit as you will then have sequencial reads from more than one drive in parallell, and similarily for very large writes.
RAID0 does not give any performance benefit for Squid as Squid already distributes the load among the cache_dirs. You only gain easier configuration thanks to a single large drive at the expence of loosing the whole content should one drive fail.
RAID5 of three drives with a good battery backed up HW raid controller is quite well balanced between performance and redundancy. But don'w confuse this with "fake" raid controllers without battery backup as these does not provide the same performance benefits and you get all the negative performance impacts of RAID5 only gaining redundancy.
Direct mounting of the drives is generally the absolutely fastest, via the raid controller battery backup is you have one. But lacks in redundancy should a drive fail. RAID1 of OS and configuration recommended for redundancy, but somewhat costly for cache drives..
Regards Henrik
