Dependency-Driven Interconnection Network Simulation Using MPI Traces

As the scale and complexity of high performance computing (HPC) systems grow, interconnection networks have become a key subsystem. A comprehensive network simulation needs actual workloads to represent important characteristics of HPC systems. In this paper, we propose a flexible framework, NetApp, which generates actual message passing interface (MPI) workloads for network simulation. We analyze MPI communication algorithms and detailed HPC system characteristics, based on which we build the Chain-Expansion and Ideal-System models to enable targeted simulators to replay MPI messages and track their dependencies. We also extract the communication behavior property of the parallel application to construct the NetApp prototype to generate causal workloads consistent with program semantics. Unlike existing frameworks, our design does not rely on specific simulation platforms for support. The NetApp framework offers several novel features. It takes into account the MPI algorithm and system properties, preserves message dependencies, generates workloads with application-specific characteristics, and enables third-party simulators to handle MPI-based network traffic. Our key simulation demonstrated that the NetApp framework could generate effective workloads with dependency and display the variation trend of network and system performance.