RE: disk arms (was RE: Tips for user ASP)

Steve,

Stuff below is intuitive, but makes sense to me.

When a drive starts exceeding its 40 per cent threshold, the likelihood
that a write will need to be re-queued increases dramatically.  I would
expect under these circumstances that cache would start filling up.  If
the condition persisted, I would expect cache to become full.  I assume
that a full cache would lead to programs waiting for I/O completion
before proceeding to the next instruction, just as if there were no
cache.

I would expect cache to be a remedy for brief spurts of intense disk
activity.  I would expect that it would not be a remedy for chronically
over-loaded disk.  It might take a while to fill up the buffer, but once
full I would expect no benefit.

I don't see any more than a slight reduction in required disk arms based
upon the presence of a large cache.  The activity threshold of 40 per
cent is still there, and the writes to disk still need to occur.  Brief
overloads would have no impact, as you state.  But consistently
overloaded disk arms should still occur at the same point.

If your point is that there should be less difference in throughput
between RAID-0, RAID-1, and RAID-5 on a moderately loaded system because
of cache, I would agree.  RAID-5 will still require a recalculation of
the parity bits, which involves looking at every drive in the set.  I
would expect mirroring and unprotected storage to be closer in
response-time than they were before write caching.

Regards,
Andy Nolen-Parkhouse

> On Behalf Of Steve Richter
> Subject: RE: disk arms (was RE: Tips for user ASP)
>
> Doesn't the presence of a disk cache, esp a large cache, invalidate
all
> these considerations of arms and raid protection ?  A cache is non-
> volatile
> memory. When writing to a cached mirrored pair I assume the system is
> writing to paired cache. So, unless the cache is full, there will be
no
> delay in performing the mirrored write.
>
> Steve Richter