SimMSG: Simulating Transportation of MPI Messages in High Performance Computing Systems.

The simulation of interconnection network and MPI message transportation behavior is an important part of the simulation of High-performance computing systems. Current research on the simulation of message transportation in HPC systems mainly focuses on the simulation of interconnection networks of the target HPC system. However, with the development of HPC techniques, the procedure of message transportation in HPC systems has become more complicated. Firstly, the low-latency transportation techniques, e.g. RDMA or RoCE, involve not only the interconnection network but also essential operations on messages inside send/receive nodes. Secondly, collective communications generate a vast number of messages and impact performance significantly. It is necessary to accurately simulate the whole procedure for all messages. To address the above problems, this paper presents SimMSG, a message-trace-driven simulator that simulates the whole transportation procedure of MPI messages in HPC systems. SimMSG simulates message transportation by incorporating simulation of MPI communication routines, in-node underlying processing of messages and message transportation in the interconnection network. We evaluate SimMSG with a real HPC system. SimMSG simulates the real-captured message-traces from the system and can re-produce the whole message transportation behavior of the parallel program in the target system. Experimental results demonstrate the accuracy and efficiency of SimMSG.