[svlug] New server plans moving forward

Rick Moen rick at linuxmafia.com
Sun Jan 21 18:59:18 PST 2007

Quoting Daniel Gimpelevich (daniel at gimpelevich.san-francisco.ca.us):

> If you know that for a fact, then the way we've set it up is indeed
> optimal.

Let's talk about that.

> Huh? Doesn't RAID5 (parity) _reduce_ usable space, given the same
> combination of drives?

In RAID5, you lose to parity overhead the storage equivalent of one
drive among those included in the stripe set -- as opposed to the
current lossage to parity overhead of the storage equivalent of half the
eight drives deployed. 

In a nutshell:

Current setup:  one 144GB volume
8-drive RAID5:  one 252GB volume
7-drive RAID5:  one 216GB volume (plus one 36GB hot-spare, as the 8th)

> This was the purpose of the mirroring. We made a distinction between a
> mirror of stripes (RAID01) and a stripe of mirrors (RAID10).

Yes, they have different, but equally problematic, failure scenarios.
Also, they have in common that 50% (144GB) parity overhead.

RAID 0+1 or RAID 1+0 also yields faster disk I/O, but we don't really
benefit from that (because Web/mail servers are typically much more
constrained by network I/0 or RAM, than by disk I/0).  It's likewise
less CPU overhead, but that doesn't really matter for us, either
(because Web/mail servers are almost never CPU-bound, either).

The only time you typically even feel the Linux "md" driver's overhead
at all in RAID5 or RAID4 mode is when remirroring after drive
replacement -- unless you _are_ in fact signficantly CPU-bottlenecked,
which is (again) highly unlikely on Web/mail servers.

> I am well aware of how cheap those dongles have become. I believe I
> have seen them at affordable prices at Central Computer, Halted, and
> Microcenter. However, I believe that no non-SCA drives ever existed
> which could physically fit inside the model 9008 in the presence of
> one of these converters.

Actually, I believe otherwise, and can recall having demonstrated same
to Paul, in person.

> In addition, the converters require an external power connection,
> which the 9008 does not provide.

No, you're thinking of the converter that goes the _other_ way, that
lets you adapt an SCA drive for use on a regular 68pin chain.  68pin 
hard drives take a D-shell ribbon data connector plus a Molex power
connector (and have SCSI ID jumpers); SCA hard drives connect data +
power + SCSI ID signals through the single SCA connector.

> We used only 36GB drives in the 9008. He asked for another 36GB to be
> donated because it would hopefully have a sufficient number of blocks to
> be able to mirror the remaining drive in the last mirrored pair, which is
> a 36GB drive, like the rest. That drive, like the dead one, but unlike the
> rest of the drives, is a Quantum Atlas V, if that's any indication of its
> exact number of blocks (this varies with the model).

Whatever it is, I have the means (and desire) to re-check it.  If you

More information about the svlug mailing list