A Performance Model of k-Ary n-Cube Under Communication Locality

Interconnection network is a key component of parallel computer which influences the parallel application’s efficiency significantly. k-ary n-cube network topology is widely adopted in many current large scale parallel computers such as Blue Gene/L and Cray XT4. The method of spatial communication locality is implied in many real parallel programs, but as we know the exiting analytical models just gave some rough and fragmented results on network performance under communication locality. The binary parameters, local message fraction and local domain’s radius, are proposed to measure parallel application’s spatial communication locality. Then based on queueing theory, an analytical model of the k-ary n-cube is established to evaluate communication locality’s impact on message latency and network throughput. The situation that the message’ length is shorter than the network radius is discussed, which has not been considered in the existing analytical models. The impacts of different parameters on network performance are compared carefully. At last, an improved network simulator is used to validate the analytical model, and it is demonstrated that the analytical results close to the simulation experiments well. The proposed model gives an efficient method to predict message latency and network throughput of large scale parallel applications whose communication patterns have local character.