On Fri, 30 Jul 1999, Tom Livingston wrote:
> Are there drawbacks to setting your stripe size to be 4k ? The FAQ I
> used when setting up my 10 drive raid5 array suggested 128k as a
> starting point, which is what I used.
the only possible drawback: the disk/controller might not handle properly
such high-frequency 'skip,read,skip,read,...' SCSI-command workloads. Most
controllers/disks have no problems with that. (any decent disk has a more
than 40k readahead buffer - 40k == 4k * 10, the 'skipping period' for your
config.) Some disks/controllers though dont like the higher amount of SCSI
commands - so this can result in a slowdown. (i have seen such a slowdown
on older disks with crappy SCSI tagged-queueing firmware)
> Others on this list seem to be confused as to how to calculate an optimum
> stripe size. Would this be a good thing to "get the word out" on, or is
> there a more appropriate way to determine it?
no generic rule i'm afraid - but testing this isnt difficult. Use the
--dangerous-no-resync mkraid flag (available since the latest release) to
set up a test-array with various chunksizes - this way you dont have to
wait for resync to finish. (CAUTION, the flag should not be used for any
production array that is freshly created) Of course if the array is
already created this cannot be changed anymore. (there were ideas floating
around to allow background resizing of RAID5 arrays - but there is no code
yet.)
if the controller and disk firmware handles higher intensity tagged
queueing correctly, then you'll get the best performance with 4k
chunksize.
-- mingo
