On the performance and scalability of an HPC enhanced Multi Agent System based evacuation simulator

Leonel Aguilar Melgar, Lalith Maddegedara, Tsuyoshi Ichimura, Muneo Hori

Abstract: This paper presents some of the techniques, algorithms and designs used to enable mass evacuation simulations to take advantage of high performance computing infrastructure. A brief overview of a tsunami mass evacuation simulator capable of simulating urban areas of hundreds of km2 in sub-meter detail is provided. Enhancements to the serial algorithms for path finding reducing the path finding time in 94% and a cache friendly visual boundary extraction algorithm cutting the overall simulation time in 50% are presented. Furthermore the hybrid parallel (distributed memory (MPI) + shared memory (OpenMP)) framework is described. A dynamic load balancing technique reducing the idling time from 50% of the execution time to 3% is presented. Finally measures of the thread parallel strong scalability up to 16 threads of 82.69% and distributed process strong scalability up to 2048 processes of 75.93% are presented.