Molecular dynamics on hybrid memory machines

Abstract

The performances of new clusters of SMP machines is strongly dependent on the available bandwidth on the cluster network architecture. This happens also for applications where, in distributed memory hardware, the bottleneck was connected to the overall latency time of interprocessors communication more than to the available bandwidth. In this paper we will report the results of runs of our molecular dynamics codes for a memory channel cluster of 4 dec 4100 SNIP processors. The two molecular dynamics code are written in fortran with PVM and MPI message passing calls. The request in communication capabilities are different for the two codes, the first is used to study polarizable fluids and contains any to any communications, the second is used for the simulation of glasses at low temperatures and require only first neirbourghs processor communications. We will present the obtained speedups for different bandwidth of the memory channel hub hardware and software configuration.