Analysis of Concurrency and Coordination Runtime CCR and DSS

CCR has been developed by Microsoft and applied to several applications – especially robotics. CCR has also been explored as a runtime supporting an interesting concurrent programming model and has DSS – a lightweight service runtime – built on top of it. In this note we discuss its application to high performance computing where the messaging system MPI is the dominant paradigm as either the programming model or the runtime for a higher level programming paradigm. We conclude that one can do MPI-style programming within CCR with performance characteristics competitive with the best MPI implementations (openMPI, MPICH). We identify the loosely synchronous execution structure with independent threads executing for a few microseconds and exchanging messages – a sequence of compute-communication phases – as typical of hard technical computing problems. We design simple performance measurements of loosely synchronous execution in CCR corresponding to use of MPI ping and broadcast tests. We find latencies of around 5 microseconds and “cross-section bandwidths ” of a gigabit/second with CCR providing efficient thread execution. We compare two machines in details – the one with two dual core Opterons shows lower latencies but also lower message bandwidths than the PC with two dual core Xeons. Some results are given for a newer machine with two quad core Xeons. We discuss relationship to classic MPI messaging, dataflow, “active messages”, overlay networks and publish-subscribe communication. The implementation of MPI in terms of CCR depends on one’s goals and here we suggest it could be very interesting to generalize CCR to generate a multiparadigm runtime that’s fully supports MPI but also other messaging models that more appropriate outside technical computing. We illustrate this with an initial evaluation of the service environment DSS built on CCR; it supports around 50K two-way (requestresponse) messages per second internal to the AMD machine. This is a factor of ten faster than typical Axis-2 web service messaging. 1