Re: [tlug] SATA software RAID or SAS hardware RAID?

["Stephen J. Turnbull" <stephen@example.com>]

Curt Sampson writes:

 > Some servers, for example, have more than one PCI bus, spreading the
 > disk I/O load across the two buses with IDE controllers may well achieve
 > better performance than the maximum theoretical performance possible
 > with a single controller of any kind.

For me the killer fact is that people who depend on reliable servers
for their bread and butter (eg, Akamai, IIRC) still are willing to pay
for gold-plated SCSI/RAID solutions.  Do you think they're just being
excessively conservative?

Also, while I'm no expert, getting multiple anything to perform better
than a single piece of hardware seems to be non-trivial.  Eg, I was
surprised but not shocked to discover that a benchmark program that in
sequential mode takes 1.2 seconds for 1000 iterations takes over a
second *per iteration* in parallel mode[1] on a 2 x Dual Opteron
system.  My suspicion is that the bottleneck is master-slave messaging
from socket to the other (I know the bottleneck is intermittent, and
it is in the messaging, but I haven't confirmed that it matters which
pair of CPUs are involved).  The breakage is pretty spectacular, even
though the parallel monitor is very-well-respected software, and the
mobo (Tyan 2895) and CPUs are excellent.

N.B. I don't disagree with your main conclusions: I think almost
everybody is well-advised to start with software RAID on SATA which is
cheap and highly reliable, and only if that isn't good enough is it
worth thinking about gold plate.

Footnotes: 
[1]  Using the PVM package, which probably communicates via TCP/IP
over Unix-domain sockets.  MPIv2, several implementations of which
have good shared memory support, seems to be the way to go for real
work, but PVM comes with "batteries included" for demos, eg, a nice
XPVM visual monitor, and extending to heterogenous networked VMs is
trivial.