RAID (was: RE: [geeks] New Itanium machines from SGI)

[Gregory Leblanc]

On Wed, 2003-01-08 at 16:49, an unknown sender wrote:
> >Hmm, I actually think that Software RAID on Linux kicks some serious
> >ass.  The only other implementation that I've used heavily is on NT, and
> >that one blows goats.  Anyway, the Linux software RAID implementation is
> >faster than Dave chasing a pretty girl.  It easily outstripes any and
> >all of the PCI based RAID solutions that I've been able to get people to
> >run benchmarks on.  Using RAID 5 on modern CPUs uses code tuned to fit
> >into L1 cache, making the checksumming FAR faster than you can hope to
> >deliver from disks.  RAID 1, 0, and 10 don't show up as being any faster
> >or slower than the PCI based RAID cards (the differences are within the
> >margin of error for my testing).  Booting from RAID 1 works and is
> >pretty well tested.  Software RAID also offers you the flexibility of
> >working with slices (err, partitions, whatever) instead of whole disks
> >(I've never seen a hardware-based solution that worked this way).
> >Anyway, I think software RAID is a clear winner for -small- RAID arrays,
> >meaning no more than a couple dozen disks, and maybe a TB of moderately
> >use.
> >	Greg
> 
> What were your observations on CPU utilisation for the systems you were
> benchmarking? Surely you don't get all that software-RAID goodness for free?

On the IDE drives, and a P-III 1GHz CPU, I think they were generally
around 80% -when running the benchmarks-.  The benchmark that I still
have handy was done with a file size of 800MB, and took about 15 seconds
to either read or write.  I've inherited a couple of machines running
Software RAID 5 across 4 SCSI disks, with dual P-II 400 procs, and the
highest I've seen the kernel thread for RAID 5 work is about 15%.  So,
certainly not free, but pretty cheap, all things considered.
	Greg

-- 
Gregory Leblanc <gleblanc at linuxweasel.com>