MPI communication on MPPA Many-core NoC: design, modeling and performance issues

Power dissipation and energy consumption has become a major issue for high performance computing and embedded systems. Keeping up with the performance trend of the last decades cannot be achieved anymore by stepping up the clock speed of processors. The usual strategy is nowadays to use lower frequency and to increase the number of cores. On such recent systems, data communication and memory bandwidth can become the main barrier, since there are more and more processing units to coordinate. In this paper, we introduce an MPI design and its implementation on the MPPA-256 (Multi Purpose Processor Array) processor from Kalray Inc., one of the first worldwide actors in the many-core architecture field. A model was developed to evaluate the communication performance and bottlenecks on MPPA. Our achieved result of 1.2 GB/s, e.g. 75% of peak throughput, for on-chip communication shows that the MPPA is a promising architecture for next-generation HPC systems, with its high performance-to-power ratio and highbandwidth network-on-chip. However, the lack of a globally addressable memory on this distributed-memory architecture still requires the developer to take care of cache coherence and to pay attention to the limited local memory space of each compute element. Keywords—Many-core, NUMA, Distributed memory, Networkon-Chip, MPI, Performance modeling, Linpack, HPL, MPPA.