Buffered Branching Channels with Rendezvous Message Passing

G. Yaikhom (UK)


Parallel Programming, Parallel Processing, Message Pass ing, Buffer Management.


In distributed parallel programming, it is desirable to have simple, uniform and flexible APIs as message passing code segments complicate a program. We discuss ambiguities that can arise when MPI collective functions are used on overlapping communication domains. We describe how these ambiguities can result to different implementations of the same communication, with different performance char acteristics. To resolve the ambiguities we introduce branch ing channels that overloads communication functionality at run-time, also providing a highly extendable and uniform interface. Branching channels support programmer speci fied dynamic buffer management, which increases interface flexibility. The buffering also enables an efficient multi threaded implementation of the rendezvous protocol.

Important Links:

Go Back