HPC performance analysis of a distributed information enterprise simulation

Simulations of distributed information enterprises using the DIEMS framework were performed on HPC clusters and SMP machines. The simulation results were analyzed with respect to the computation time involved in each process, the number of times processes were executed, the number of simulation rollbacks invoked during each simulation, simulation overhead, and the total wall clock time used to perform the simulation. The analysis identified several performance limitations and bottlenecks. One critical limitation addressed and eliminated was simultaneously mixing a periodic process model with an event driven model causing rollbacks. The second major factor limiting performance on cluster based systems was the cross-node communication. An optimization technique that exploits the knowledge of the publication and subscription paradigm of the information architecture being simulated was developed. This paper describes the simulation analysis, the modifications to the simulation models, the development of an optimization technique, and the impact of the code improvements on simulation performance.