Megaraid with brakes on?

[Ben Scott]

On 1/4/07, Dan Pritts <danno at internet2.edu> wrote:
> This is no less reliable than using RAID0 and my guess is it will work
> better than RAID0 for your workload.

  It depends.

  *If* you have enough physical disk units to offset the overhead of
distributing the I/O load, *and* you have enough I/O bandwidth, *and*
your disk controller(s) can keep up, *and* the kernel's scheduling
algorithms can take advantage of the increased I/O availability, *and*
your userland software can take advantage of the increased I/O
availability, then distributing the I/O across physical units can give
you a very real benefit.  That's a lot of "if's", of course, but
real-world scenarios are still not uncommon.

  If it's a single user reading all those small files in a random
pattern, then I would not expect much benefit.  A single user likely
means every file is only read once (per lifetime in cache), and the
random pattern means read-ahead won't help.

  If it's multiple users reading all those small files in a random
pattern, then things might work out better.  One physical unit can be
serving up the file for one user, while a different physical unit can
be serving up the file for a different user.  In theory, in this
situation, distributed I/O could help quite a bit.  I really don't
know enough about the workload to say for sure.

-- Ben