[conspire] Partitioning problem

[Tony Godshall]

According to Rick Moen,
> Quoting Tony Godshall (togo at of.net):
> 
> > I'd put whatever you expect to be the heavyest read-load on
> > (/var/www?) a 2-way or 3-way mirror RAID (to exploit read 
> > parallelism for best performance).
> 
> I actually don't know if the "md" driver supports 3-driver RAID1, 
> but hadn't considered that, because only two of the candidate drives are
> of similar size.  But yes, read optimisation is a nicely incidental
> gain you get for free with the redundancy.
> 
> (Karsten's about to post a big flippin' hole in my scheme, in the
> portion of this thread on the SVLUG list.  I plead that I merely listed
> as candidates for the mirror pair those parts of the file tree I back
> up, in hopes that people would _tell_ me why parts of them would be dumb
> to put there.  I'm about to get my wish.)

People often think of RAID primarily for redundancy, but the
performance gain for read access is just as real, and
is especially noticable when you run with noatime mounts.

Actually I find the idea of RAID for reliability a bit
flawed myself- if they are in the same case and share the
same power supply you are just as likely to fry two or three
drives at a time as just one.  

If reliability is the primary goal, I'd rather have fully
redundant systems and load-share across them- at least for
apps like webserver where it's easy to do so.  

> > Perhaps split swap across all three....
> 
> I've considered that.  It just seems oddly wrong, when the whole rest of
> the 18GB pair are devoted to mirrored data.  But it bears thinking.
> 
> In case it wasn't obvious, I've never used "md" software RAID before
> (only Mylex hardware RAID, which sucked ;->  ).

Well, if you want a high-performance system, you just give
it enough RAM to make swap performance fairly irrelvant 
anyway... 

TG