"idunno" wrote in message
om...
"Bill Todd" wrote in message
...

For small writes, a classic RAID-3 implementation may be slightly faster
because its spindles are synchronized (and the mirrored spindles in
RAID-10
typically aren't, so the write there will take take the longer of the
two
disk access times). But if the RAID-10 controller has stable write-back
cache such that it can queue up write requests and execute them in
optimal
order, any such advantage may be reduced or eliminated (as it will be
for
large writes in any event, where the access times become a smaller
percentage of the overall overhead).

But that's for purely serial small writes. For multiple small writes
requested in parallel, RAID-10 may well be able to process at least some
of
them in parallel, whereas RAID-3 will serialize them at the parity disk
(though should at least be able to queue-optimize their execution).

RAID-10, of course, provides potentially significantly better *read*
performance for a given usable capacity. It also provides somewhat
better
availability, since with RAID-3 (or -5) the loss of any two disks
results in
data loss whereas with RAID-10 data is lost only if the two disks happen
to
be partners.

- bill

Thank you. That helps. Are these performance differences maintained
when you enable spindle sync on a RAID 10 array with a large
write-back cache?

Hmmm. Most people with read-dominated workloads would likely consider
unsynched spindles a *feature* of RAID-10, since they'd get the faster of
the two potential read options (at least if the array were suave) at only
small expense in write performance. But if the array allows the option of
synching the spindles for RAID-10 operation, and if you're performing only
writes (so that the heads can be presumed to be in the same position on both
mirror partners before each write request - at least I hope you could
presume that), then I can't think of any situation off the top of my head in
which the RAID-10 array wouldn't equal the speed of the RAID-3 array, and
for parallel write operations it might well be superior (with the large
write-back cache, even somewhat larger writes might be able to achieve some
parallelism, though you're still going to have a hard limit of half the
aggregate streaming bandwidth of the disks - the same value as the limit for
a RAID-3 array of equal usable capacity, assuming that its parity generator
can keep up).

- bill