[wplug] Hardware RAID tuning

[Matthew Zwier]

On Wed, Jun 15, 2011 at 10:17 PM, Drew from Zhrodague
<drewzhrodague at zhrodague.net> wrote:
>        Shouldn't you not share controllers with other arrays, and actually use
> multiple controllers per array? This is how we did things back in the
> day with old JBOD Suns -- stripe across disks and stripe across
> controllers. Modern architectures turn everything on their ears, so I
> could be missing something.

I doubt you're missing something, but we're out of slots in the
server, and out of space in the rack for a new one :)  Did I mention
this is a scientific computing cluster for a small research group?
Funds and space are limited.  Also, I can't correlate the performance
drop to a specific load on the other array.

>        If it helps, I've found that a straight scp in Amazon's cloud is faster
> than an scp -C or compressed rsync. The results may be different with a
> hardware environment, but testing is the only way to be sure.

Yeah, tested this yesterday.  Compression slows things down
considerably, Enumerating files (rsync) slows things down to kilobytes
per second.  A typical user has a million or so tiny files and a few
hundred (10 GB - 200 GB) very large files.  xfsdump over an
uncompressed ssh connection appears the way to go.

Interestingly, piping things through bar to measure throughput seems
to slow things down a *lot*, and in fact may be the greater reason for
the intermittent throughput -- I'm sustaining 4 MB/s - 50 MB/s,
seemingly correlated to what set of files xfsdump is working on, now
that I don't have bar in the pipe.  I'm wondering if I need to do
something *really* dumb, like xfsdump -J - | xfsrestore -J -
/some/nfs/mount.  That would seem to be going through one fewer
userland buffers.

Hmm...what's the bandwidth of a PCI bus?

MZ