SiCortex fabric backplane

[C Bergström]

Hi

I'm not sure how much of the details about SiCortex's fabric backplane
are public, but I want to develop something similar for a bunch of
processors.

Is there anyone still around who worked on this?

I'm also fully open to non-IB alternatives.

Thanks

[Lawrence Stewart]

I think pretty much the full details of the SiCortex fabric are public. I (modestly) know
quite a lot about it.

The topology was a Kautz graph, see http://en.wikipedia.org/wiki/Kautz_graph
or http://planetmath.org/kautzgraph

For a technical paper, see http://www.dia.eui.upm.es/asignatu/arq_ava/articulos/anetworkfabricsicortex.pdf

About a year ago, I looked into whether it would be possible to wire up standard IB switches
with a Kautz topology, and ... I couldn't figure it out. If the switches are purely adherent to the IB spec, then no, because you would need to choose VC and output link based on both the
arrival link and the destination, and the IB spec doesn't do that.

It might be that particular switches are more flexible.

I'm happy to chat more, here or via direct email or whatever...

-Larry Stewart

> --
> You received this message because you are subscribed to the Google Groups "SiCortex Users" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to sicortex-user...@googlegroups.com.
> To post to this group, send email to sicorte...@googlegroups.com.
> Visit this group at http://groups.google.com/group/sicortex-users.
> For more options, visit https://groups.google.com/d/optout.

[Jud]

Larry does know quite a lot about it, and I'm probably the other one to ask, having dug up the topology in the beginning.

Your note makes it appear that you think it uses InfiniBand, which is not the case.

The Kautz graph is attractive if the design is constrained to combine the switch with the processor chip, but if you can afford two chip designs, you may actually be better off separating the switch from the processor, because it means the processor can dedicate more pins to memory bandwidth, and the switch can use more pins to achieve a higher radix.  At that point, something like the Dragonfly topology becomes more attractive.

    Jud

[C Bergström]

Of course I know it doesn't use IB. I want to design something similar
without IB. I wanted to make the 2nd requirement clear since many
clusters depend on it.