Extend Derecho to support remote memory locations

[KenBirman]

This is a fairly hard suggestion, but I'll toss it into the issue tracker because as we move towards an open development model over time, it may be a great topic for someone to pick up from outside our Cornell team.

This builds on issues #99 and #139.

Right now RDMA and LibFabrics are "tied" to the single machine of the process using the layer. But one could imagine a memory model in which a buffer might reside on some other memory unit, maybe even on some other machine across the network.

The puzzle (publishable research topic!) would be: extend Derecho to allow process A to initiate a send from a memory region on process B, with proper permissions of course, and for process C, the receiver from A, to designate that the message should be stored into memory on process D, again with proper permissions. I can think of ways we could do this in software, but it is also possible that LibFabrics may be evolving to include this sort of functionality in future releases.

[KenBirman]

Oops, I mean it builds on #99 and #122.

[KenBirman]

Closing this because it looks like too large a departure from our current computing model.