(Georgia Institute of Technology, 2005)
Perumalla, Kalyan S.; Dave, Jagrut Durdant; Fujimoto, Richard M.; Karimabadi, Homa; Driscoll, Jonathan; Omelchenko, Yuri
New discrete-event formulations of physics simulation models are emerging that can outperform models based on traditional time-stepped techniques. Detailed simulation of the Earth s magnetosphere, for example, requires execution of sub-models that are at widely differing timescales. In contrast to time-stepped simulation which requires tightly coupled updates to entire system state at regular time intervals, the new discrete event simulation (DES) approaches help evolve the states of sub-models on relatively independent timescales. However, parallel execution of DES-based models raises challenges with respect to their scalability and performance. One of the key challenges is to improve the computation granularity to offset synchronization and communication overheads within and across processors. Our previous work was limited in scalability and runtime performance due to the parallelization challenges. Here we report on optimizations we performed on DES-based plasma simulation models to significantly improve their parallel performance. The mapping of model to simulation processes is optimized via aggregation techniques, and the parallel runtime engine is optimized for communication and memory efficiency. The net result of the enhancements is the capability to simulate hybrid particle-in-cell (PIC) model configurations containing over 2 billion particles using 512 processors on supercomputing platforms.