MPI Communication Performance On the Two-Layered Crossbar PRIMEPOWER

摘要

Fujitsu recently released a series of high-performance servers named PRIMEPOWER, as highly available and scalable servers that are targeted for Internet applications. Several models are available and are classified based on the number of processors on the system. Basically PRIMEPOWER is the Symmetric Multi-Processing System (SMP) that is based on the new SPARC64 GP processors with up to 128 CPUs per system. The SMP model is structured with two-layered crossbar switch network that theoretically provides a maximum snoop performance of 57.6 gigabytes per second, This paper describes the hardware structure of the PRIMEPOWER systems together with the preliminary measured MPI inter-processor communication performance. The communication patterns measured included the point-to-point, one-to-all, all-to-one, all-to-all, and broadcast. The experimental communication elapsed times were then used to systematically model the communication cost based on the traditional LogGP model. Experiments were performed on two different PRIMEPOWER models to see the effects of level-2 crossbar structure on performance of processor communication.