[Beowulf] Intel splits the network

[atchley tds.net]

On Mon, Apr 22, 2013 at 7:21 AM, Joe Landman <
landman at scalableinformatics.com> wrote:

> On 04/22/2013 07:16 AM, atchley tds.net wrote:
> > The Intel page linked from above says that it is standards based. I
> > assume OpenFlow, but I don't see it listed anywhere. Ahh, looking at one
> > of their white papers, it mentions it could be based on an API like
> > OpenFlow.
> >
> > Intel is moving into switches? Does Intel already make switches?
>
> Intel owns Fulcrum (ASIC in many 10GbE switches), Aries interconnect,
> Infiniband (from QLogic/PathScale).  Switches were, to a degree,
> inevitable for them.  A natural progression.
>

I had forgotten about Fulcrum. I was under the impression that Fulcrum made
the chips and sold them to switch vendors. OpenFlow takes effort from the
switch vendor, not necessarily the silicon vendor. If they get in to the
business of selling complete switches, then this makes sense.

Does this does apply to Aries or IB-style switches? Aries, like Gemini,
uses hardware routing (butterfly versus 3D torus). What would it mean to
have a software defined routing on a hardware routed fabric? The routers
would need to be (more) programmable.

With InfiniBand semantics (lossless and strong ordering), you can't define
multiple paths or reroute so what is the advantage of software defined
networking? Possibly in management, but I assumed that vendors already
provided tools to manage switches en masse within a single fabric.

DOE is spending some research money on software defined networks (SDN)
using OpenFlow on ESnet. The driver is to provide special capabilities over
the wide-area between labs (e.g. multi-pathing for higher throughput) that
may not apply within the cluster/HPC system.

Scott
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20130422/e66bfce1/attachment.html>