[Beowulf] Intel splits the network

[Greg Lindahl]

On Mon, Apr 22, 2013 at 10:35:34AM -0400, atchley tds.net wrote:

> I had forgotten about Fulcrum. I was under the impression that Fulcrum made
> the chips and sold them to switch vendors.

They did. Now at Intel, they've showed off a new architecture (based
on shared memory, an unusual implementation choice for a big switch)
and indicated that among other things, they're going to build a lower
port-count switch to go on Open Compute racks.

> With InfiniBand semantics (lossless and strong ordering), you can't define
> multiple paths

A given IB flow has to have a single path, but you can send multiple
flows between node A and node B via multiple paths in a standard-
conforming fashion if you have multiple LIDs. Also, QLogic has a
clever multipath feature in their FC switches which could be invisibly
done in an IB switch -- I stopped "needing to know" early enough that
I don't know if that feature made it into the IB switch silicon that
Intel now owns.

If I wanted to be grossly overgeneral, SDN is taking all the awesome
things that IB can do on a local network and applying it to ethernet
for both WAN and LANs. It's pretty cool, and maybe I should have
listened 6 years ago when Feldy said I would looooove what Google had
in mind for me to do!

For Ethernet, where allegedly the best you can do is spanning-tree but
vendor silicon provided the hidden ability to do routing the same way
as IB or Myrinet, this is a revolution. And for WANs, it provides all
the features that various network scheduling systems were trying to
provide.

-- greg