Apparently, it is not just me:
--------
See: messages.yahoo.com
"Re: q on emc specmark numbers vs netapps
by: mrnetapp (38/M/Rochester/NY)
12/13/00 12:57 pm
Msg: 49957 of 49957
The peak values are higher - but you have to look at HOW the values were derived.
EMC used RAID0 - stripping the data across multiple drives without any parity protection(RAID levels 3,4, or 5) nor any
mirroring (RAID 1). Stripping the data across multiple drives is a known method of increasing speed, but is NEVER done in a
production environment since the chance of losing data is actually higher then storing the information on one single disk. Say the
data was stripped across 7 disks and didn't have any parity protection; now if *ANY* of the 7 drives failed, all the data is lost
vs. if the data was stored on 7 separate disks and one failed, only the data on that drive would be lost. Stripping (RAID 0) is
usually combined with mirroring (RAID 1) for something typically written as "RAID 0+1". This is almost the defacto standard
for EMC even though they offer RAID5. Why? RAID5 takes a serious toll in terms of performance and if you're an EMC
sales rep, you get to sell 2x the amount of storage needed with RAID 0+1 since everything is mirrored. An EMC rep also gets
the benefit of filling up the cabinets faster too (if you're mirroring, a given box spec'd to hold say 9TB will actually only yield
4.5TB best case -- less actually, but those details are more then I care to get into...) - so they sell more boxes!
Back to your question. EMC used RAID0 - best case for performance, but completely unrealistic for any customer who
actually cares about not losing their data. Strange move for a box touted to have "built-in" High
Availability.
I guess it's there, but it's not worth using, or something.
Another VERY interesting point is the number of file systems (mount points or network disks from the perspective of a client
accessing the data over a network..) EMC needed to achieve their results. This is important because clients/applications do not
typically perform parallel reads/writes to multiple file sytems at once - i.e. a given client/application will read and/or write to
only one file system for a given I/O operation. Therefore, it's very important to look at the number of ops per file system a
given system can muster. In EMC's case, they used *18* file systems to get the 10,422 ops/sec you mentioned. This works
out to 579 operations *maximum* per file system. No matter what happens, no single user or application will ever see more
then 579 ops/sec. Sun Micro has played this game for years and years (Auspex did too...) by showing decent "total box"
ops/sec numbers, but
the real ops/file system were awful and NetApp always CRUSHED them in head-to-head benchmarks with realworld
applications.
Why? A NetApp F840 not only gives the user 7783 ops/box, it can do this for a single file system, so those 7783 ops/sec can
be used by anyone or any application hitting the Filer.
Finally, NetApp hits those numbers (over 13x faster...) with full RAID protection vs. the nonsense EMC is pulling with RAID
0.
Look closely at EMC's other results too - very seldom if EVER will you see RAID5 being used in
a benchmark. They'll sell the features and benefits of RAID5, then turn around and say "oh, you wanted performance?" - then
move the customer to RAID0+1 with either half as much available space as they originally thought they were buying or charge
for the 2x overhead to use mirroring."
-------- |
