Study of point-to-point communication latency for MPI implementations in cloud

This paper proposes a Heuristic Latency Model that characterizes MPI point-to-point communications in both cloud and non-virtualized infrastructures and analyzes the influence on the latency of VM allocation policies in cloud for MPI point-to-point communications. PingPong and PingPing communication patterns were analyzed over commodity hardware in two types of infrastructures: a cloud based on Apache CloudStack and a non-virtualized cluster. Two opposite allocation policies were considered: concentrated, in which the Virtual Machines are allocated in the same host, and scattered, in which the Virtual Machines are spread using different hosts. The MPI communications were analyzed for message sizes ranging from 1 byte to 4096 KiB. The results show that the cloud layer introduces a considerable overhead in the latency. However, if the cloud is employed to execute MPI applications, the concentrated allocation policy is between 30% to 90% better than the scattered policy for messages up to 4 KiB.