Scalability of Interconnects (was Myrinet scalability)

Dan Kidger daniel.kidger at quadrics.com
Sun Jun 16 07:10:04 PDT 2002


Greg,
----- Original Message -----
From: "Greg Lindahl" <lindahl at keyresearch.com>

> On Sun, Jun 16, 2002 at 10:53:45AM +0200, Ole W. Saastad wrote:
> > with this talk about scalability and switches I would like to
> > point out that the SCI interconnect uses no switch.

> It's very important to not look like you're using any excuse to
> advertise your product, or people will get very angry with you.

I agree.

There are many different topologies available for parallel computers, each
has its strengths and weaknesess.
A switchless 2d/3d torus is not fundementaly bad - after all the highly
successful Cray T3E is an example.
Some systems use rather sparce networks such as the SGI Origins hypercube
topology. This makes cabling much easier, but only works well if your
interconnect is much 'better' than the nodes you attach it to.
Full crossbars are very good, but become exponentially expensive for large
systems.
Several vendors now offer systems based on a tree of smaller crossbars.
These scale well in terms of cost, but there will always be a point where
there is a  'step' in the cost-curve when you cross certain cluster-size
thresholds.

The real point is to use the correct interconnect for your cluster baed on
its size, node-speed and not least the end-user's application.


Yours,
Daniel.

--------------------------------------------------------------
Dr. Dan Kidger, Quadrics Ltd.      daniel.kidger at quadrics.com
One Bridewell St., Bristol, BS1 2AA, UK         0117 915 5505
----------------------- www.quadrics.com --------------------





More information about the Beowulf mailing list