Re: Storage array advice anyone?

  • From: Matthew Zito <mzito@xxxxxxxxxxx>
  • To: jonathan@xxxxxxxxxxxxxxxxxx
  • Date: Fri, 17 Dec 2004 09:04:59 -0500

Well, I'll agree with your later posting that a lot of vendors have 
jumped on the virtualization bandwagon and I'll agree that a lot of the 
virtualization products offer little or no value.  For example, the sun 
6320 is described as a virtualizing array, when the real benefit seems 
to be that you can span management across several smaller arrays (but 
not span storage devices!).

When I'm talking about virtualization, I am talking about using large 
numbers of spindles to satisfy I/O, and in better examples, software on 
the array to rebalance I/O.  In the example you give, its true that one 
read could span 8 drives with a huge queue of pending requests, but it 
depends on the size of the I/O and the stripe size configured on the 
array, like anything else - it seems a little unfair to pick an 
arbitrary example and use that as a definitive reason why a technology 
is bad.

We can look at the other case - I could have a bunch of RAID-10 volumes 
and any given read could span two drives.  When I then have 64 I/Os 
that are pending against that lone RAID-10 volume, wouldn't I be better 
off having those I/Os against 50 disks instead of 14?

And as far as sharing disks, even EMC's RAID-1 implementation splits a 
disk into a set of volumes and mirrors them to another drive.  There's 
always the possibility of spindle contention with any RAID group, 
including a RAID-10 volume.  How well an array copes with that is a 
factor of workload, cache, and the elegance/functionality of the array 
OS.  Like everything else in the world, quality makes a difference.  
That's why I suggested that you vet your vendors heavily.


Matthew Zito
GridApp Systems
Email: mzito@xxxxxxxxxxx
Cell: 646-220-3551
Phone: 212-358-8211 x 359

On Dec 17, 2004, at 4:23 AM, Jonathan Lewis wrote:

> Sorry to come in so late on this one - I've
> had a busy three months, and only just got
> back to reading the mail.
> Personally I find the whole 'virtualization' thing a
> complete con-trick.
> Sure, I now have a LUN which is really 50
> different spindles - so what good is that if I
> send a 'single read request' and that activates
> eight of them.   It only takes 8 requests like
> that and there are 64 reads queued up somewhere,
> and who knows where they might be ?  Ask
> Cary Millsap about queueing and unstable
> response times.  (Then ask Stephen Barr what
> the minimum and maximum response times were
> for his Parallel Query problem).
> And another thought - I've got a LUN which
> has 50 different spindles. Using reasonably
> modern discs, that's probably around 4TB of
> spindles.  How many other databases are going
> to hitting those spindles ?  That's what I asked
> the DBA's at a site recently when there 56GB
> database was on a 4TB SAN.  Their S/A was
> insisting that the SAN has no performance issues -
> the database had recorded its first 3-second
> 'db file sequential read' time just fifteen minutes
> after I reset the wait times.
> Regards
> Jonathan Lewis
> The Co-operative Oracle Users' FAQ
> Optimising Oracle Seminar - schedule updated Sept 19th
> ----- Original Message -----
> From: oracle-l-bounce@xxxxxxxxxxxxx 
> [mailto:oracle-l-bounce@xxxxxxxxxxxxx]
> On Behalf Of Matthew Zito
> Sent: Thursday, December 16, 2004 12:54 AM
> To: Oracle-L
> Subject: Re: Storage array advice anyone?
> -Virtualization/abstraction of storage objects - when the LUN you are
> sending I/Os to is comprised of chunks from 50 different spindles from
> 10 different RAID-5 groups, the performance is excellent.
> --


Other related posts: