First, I would point out that the concentration of discs on fewer channels
is almost always one of the consequences of implementing RAID in the real
world, so its effects are a valid issue. Second, the web sites you pointed
out compared RAID-5 favorably with other RAID levels for reads (mostly due
to faster transfers, which is typically a small part of total access time),
but not with JBOD. Since RAID-5 can be assumed to be somewhat slower than
JBOD for reads, and lots slower for writes, and since most systems have a
mixture of both types of access, saying there is a performance penalty for
RAID-5 is reasonable and accurate. Third, the question was made with
respect to HP's AutoRAID product, which based on my first-hand experience,
suffers significant performance degradation when RAID-5 is used. The
degradation is greater for writes than for reads, but is evident for both.
-----Original Message-----
From: Steve Dirickson [mailto:[log in to unmask]]
Sent: Wednesday, December 08, 1999 11:45 AM
To: [log in to unmask]
Subject: Re: RAID5 Disc's
> Despite theoretical advantages, in practice RAID 5 nearly
> always results in
> worse performance. Even with a smart frame like an EMC this is
> true. We've done extensive benchmarking on customer's
> systems and maintain
> that, while RAID 5 has significant cost advantages over other high
> availability options, a performance penalty will have to be paid.
I certainly can't challenge your benchmark data without seeing it.
One error that people make in such benchmarks is a failure to maintain
exactly the same I/O capability. I.e., if you want to see if a volume set
consisting of 'n' drives would benefit from RAID, you have to replace *each*
of the 'n' drives with a RAID array. Obviously, the RAIDized system will
have a different (probably larger) capacity, but it presents the same I/O
capability to the 3000. Replacing all the individual drives with a single
RAID array produces a completely different I/O opportunity, and is an
invalid comparison. Since the 3000 already does a sort of non-parity
striping (a.k.a. RAID0) in software, taking away any of those I/O targets is
likely to reduce the performance. In the extreme case of replacing, say, the
4 2GB drives of an 8GB volume set with a single RAID5 array composed of 5
2GB drives (which maintains the same capacity), you have, effectively, taken
away a 4-spindle (software) RAID0 array and replaced it with a 5-spindle
RAID5 array. I would not be at all surprised to see performance take a hit
after such a change, but it is completely invalid for the purpose of
comparing RAID/non-RAID performance.
Steve Dirickson WestWin Consulting
[log in to unmask] (360) 598-6111
|