On Thu, Apr 17, 2003 at 08:29:51PM +0200, Manuel Bouyer wrote:
> On Wed, Apr 16, 2003 at 01:55:52PM -0600, Greg Oster wrote:
> > > 
> > > Yes, software raid is cheaper to acquire.  But it's slower
> > 
> > Do you have pointers to any specifics on this?  I'd be quite interested in 
> > seeing benchmarks for "modern" hardware... (preferably using the same disks 
> > on the same hardware)
> 
> This is not true, at last for external boxes connected to a SCSI port.
> For example, in my DS20 I can get 50MB/s (though the filesystem, more from
> the raw device) out of a raid-1 built from 20 disks spread over multiple
> SCSI UW channels. An external box would be limited by the SCSI bandwidth 
> (something like 35MB/s in this case). Yes, the external raid box could

That's not really so.  Except for extremely cheap, low-end boxes, it's
been _years_ since I've seen a single-attach external RAID unit.  Even
older low-end controllers like the CMD 5400 had four parallel SCSI buses
on them, two of which were usually used for the host connection and two
for disks.

There's also nothing at all preventing you from hanging multiple hardware
RAID boxes off multiple SCSI adapters.  And unfortunately, PCI RAID
cards and external RAID controllers typically have exactly the same
bottleneck: stupid firmware that wastes the controller's most precious
resource, internal memory bandwidth, resulting in transfer rates and
transaction-per-second rates a small fraction of what they could be.
Since your typical modern x86, UltraSPARC, ppc, or Alpha box has an
order of magnitude more available bandwidth and CPU resource than the
typical RAID card, even though RAIDframe is *also* not very efficient,
it can often do better.

The exception is short writes.  A RAID card with nonvolatile cache
should be able to soak those up, no matter where on the disk they're
going, basically as fast as it can turn transactions around.  RAIDframe
cannot -- though if we had support for one of those PCI NVRAM cards in
NetBSD, it probably could...

Thor