X-MimeOLE: Produced By Microsoft Exchange V6.5
Received: by onstor-exch02.onstor.net 
	id <01C78469.CA930662@onstor-exch02.onstor.net>; Sat, 21 Apr 2007 16:07:10 -0700
MIME-Version: 1.0
Content-Type: text/plain;
	charset="us-ascii"
Content-Transfer-Encoding: quoted-printable
Content-class: urn:content-classes:message
Subject: FW: Thoughts on Cougar
Date: Sat, 21 Apr 2007 16:07:10 -0700
Message-ID: <BB375AF679D4A34E9CA8DFA650E2B04E035CB1DE@onstor-exch02.onstor.net>
X-MS-Has-Attach: 
X-MS-TNEF-Correlator: 
Thread-Topic: Thoughts on Cougar
Thread-Index: AceEWXAWLrgUxg6QRtGN4w3sLci/rQAEERQQ
From: "Narayan Venkat" <narayan.venkat@onstor.com>
To: "Brian Stark" <brian.stark@onstor.com>,
	"Andy Sharp" <andy.sharp@onstor.com>,
	"Tim Gardner" <tim.gardner@onstor.com>

Comments from one of our top customers.

Cheers!

Narayan Venkat
Vice President, Marketing
ONStor, Inc.
Tel: (408) 963-2404
Cell: (408) 221-4297.


-----Original Message-----
From: John Welter [mailto:john.welter@nwgeo.com]=20
Sent: Saturday, April 21, 2007 2:10 PM
To: Narayan Venkat
Subject: Thoughts on Cougar

Hi Narayan,

I tool some time to evaluate our true requirement for cache on the
gateway for our application, and also spoke to a few oil and gas
colleagues about their idea requirements.

A typical processing job of ours will require typically ~8Gb of input
data.  At a boundary between files this would expand to 16Gb but just
for a short period, and then back to 8Gb as we fully traverse into the
next set of input files.  On the write side we'd be creating many (128
or so) 500Mb output files which is 64Gb if we cached the entire output
data space.  The issue with the output is that these are tiff files that
are written out as "tiles" (512*512 pixels) of data into the
pre-initialized tiff file structure.  The write side can be easily
handled by us simply keeping the entire 500Mb output structure in RAM
and writing it at the end of the process for that chunk.

So with the above in mind 16Gb of cache would be plenty if we had some
control on that cache to specify the % allocation to each vsvr, and also
a further allocation of % read versus write on a vsvr basis.  This way
we can configure the cache to make best use of the 16Gb. =20

On the oil and gas front I also got some very interesting information.
Here in Calgary we have a company called Kelman that has about 200Tb of
Isilon storage.  I was talking to their HPC director and he said they
were evaluating Gear6 to speed their performance.  I found this
interesting as Isilon claims that with each node you gain 4.5Gb of cache
and they scale up linearly.  What I am being told is that once you hit
about 15 nodes in a cluster the cache coherency traffic becomes so high
on the IB back end that it starts to limit performance.  For this reason
they disable read and write caching on the isilon to eliminate the
coherency traffic!  Obviously for some applications this kills
performance.  So for this reason they are looking at gear6 to front the
Isilon cluster to give them a cache pool again.  This is a dirty secret
Isilon doesn't tell anyone.

I think this is great news for ONStor as it means that eventually the
truth over Isilon's "unlimited" scaling will come to light and someone
will talk about it in public.  When this happens the whole focus on
their clustered approach will diminish greatly.  At that point people
will wonder why they are paying Isilon $5500 per TB of SATA disk and go
with more open solutions.

Hope this info is helpful.

John


=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
John Welter
North West Group
1-403-295-0694 Ext 233
1-403-589-4349 Mobile
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D


