Re: [MLUG] software raid conifg - fedora

[Dave Lloyd <EMAIL:PROTECTED>]

On Thu, 10 Nov 2005, Mark Haidekker wrote:

> I would agree with that. If one drive fails, you'll realize the problem real 
> quick. There is still enough time to replace the drive before the second 
> drive also fails. I had that situation recently - the intact drive is still 
> intact. Furthermore, after drive replacement, no data was lost.

It's my experience that if you have two drives, do an offline 
spare.  If you have three, do mirroring plus an offline spare.  In
this case, offline spare means 'do a dump once a week to it and
take it off line.'

> As I mentioned above, disk failures do occur, and I had a RAID case recently. 
> I was glad I had RAID set up on that server, but I did not miss the spare 
> drive.

Yes, disk failures do occur.  But two failures on the same raid before
you can sync a replacement disk are very rare and usually due to
external environmental influences, a manufacturing defect with a batch
of drives or ignoring the failure for an extended period of time.

> IMO, if you have 3 drives, configure them as RAID5. You'll get 160 MB, 66% 
> efficiency instead of 50%.

That might be OK, but doing raid5 across partitions is a good way to get
shitty performance.  What I'm talking about here is to do something like
create three partitions on each disk (swap, home, root) and then have
three raid5 devs each using one partition from each disk.  Heavy i/o
to/from one dev then backs up i/o to/from the other devs.  You may not
see it on a pee-cee, but I've seen it often enough on real computers,
er, large systems.

> > Oh, and please, please do not mirror swap.  Mirroring really cuts write
> > performance, and if you start to write to swap, well, things will begin
> > to suck much more than a "normal" condition under which you write to
> > swap.
> 
> Right - but having one swap partition on each drive is great. Actually, some 
> people recommend RAID0 for swap to speed it up.

raid0 would be OK.  Or having two swap partitions.  If one spindle in
your mirror fails, you'll just lose any apps that were swapped out to
the swap part on that particular spindle. 

I think that you can play pretty fast and loose with anything up to
4cpus.  When you start getting to 16+cpus (or 512) on a system, you
really have to start paying attention to stuff like this much more.

--dlloyd

_______________________________________________
members mailing list
EMAIL:PROTECTED
http://mlug.missouri.edu/mailman/listinfo/members