A few months ago, Sean Hefty from Intel started an effort to design a new low-level network API to replace libibverbs.

That is, it’s not libibverbs 2.0 — it’s a new API that aims to both expand the scope of what libibverbs did, and also to address many of its much-criticized shortcomings.  Sean and Paul Grun even held a BoF about the proposed “libfabric” at SC’13.

Design discussions have been progressing within the “OpenFrameworks” working group in OpenFabrics. Progress has mainly occurred in the form of emails and weekly conference calls. Two weeks ago, Sean asked me to gather feedback about what an MPI implementation would want from a new underlying network API, and present that feedback to the libfabric group.

So I asked the MPI community.

I received oodles and oodles of feedback from academics, researchers, and fellow MPI implementors.  Woo hoo!

After multiple rounds of discussions in the MPI community, I haven’t finished assembling the final set of feedback yet, but I sent around a first set of slides to the libfabric group today, and we got through about half of them on the call today. Here’s the slides I sent:

Next week, I hope to have completed the first round of requirements and have a full set of slides to present.

What’s really exciting here is that multiple members of the MPI community have asked for this to be an ongoing collaboration with the OpenFabrics community. This will all take some time to shake out, but it could turn into something really interesting…!


Jeff Squyres

The MPI Guy

UCS Platform Software