[Openframeworkwg] Googleable names
dledford at redhat.com
Thu Jan 30 13:27:33 PST 2014
On 1/30/2014 1:15 PM, Paul Grun wrote:
> We have focused on RDMA-enabled fabrics, but have been very careful
> to build the framework in such a way that it could be run over e.g.
> conventional TCP/IP/Enet by pushing the details down into the
> provider layer. It might not perform very well, but it would work.
> Sean please correct me if that is wrong.
OK, but it would be emulating what we get with RDMA fabrics. The app
would not see Berkeley Sockets semantics.
> There may also be a semantics problem here; I talk about
> "RDMA-enabled fabrics" to generally mean e.g. IB, or RoCE or iWARP -
> a fabric that has the fundamental characteristic that it places data
> directly in user memory w/o OS intervention (i.e. stack bypass). In
> the case of IB, there are two modes for doing so using either a
> channel semantic (e.g. SEND/RCV) or a memory semantic (e.g RDMA READ
> or RDMA WRITE).
> There are others in the world who refer to RDMA, as I believe Sean
> does and the iWARP community does, as one of several specific
> transfer models, in this case RDMA refers to directly referencing
> remote memory. Note Sean's use of the expression RMA.
OK, I agree that the term RDMA is overloaded. When I was using it I was
really referring to what I'm going to start calling Process Direct I/O
(PDIO). And in that sense, this code might be able to run over Ethernet
and TCP/IP, but it would still be emulating PDIO, not exposing those
Berkeley semantics to the app (there really is no sense in doing such a
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 899 bytes
Desc: OpenPGP digital signature
More information about the ofiwg