Impact of message filtering on HPC agent-based simulations

Claudio Márquez, Eduardo César, Joan Sorribes

Research output: Contribution to journalArticleResearchpeer-review

2 Citations (Scopus)

Abstract

Commonly, High Performance Computing (HPC) agent-based modelling (ABM) simulation platforms axe implemented using a Single Process Multiple Data (SPMD) programming paradigm in order to warranty scalability requirements. Consequently, along with load imbalance, communication bottleneck is the main problem affecting the efficiency of realistic ABM simulations. These ABM simulations involve a large number of agents which are continually communicating with other agents in each simulation step. Therefore, a proper communication management can contribute to reduce the communication overhead. For this reason, we propose a modification of the ABM simulation framework named Flexible Large-scale Agent Modelling Environment (FLAME) to provide filtering routines for allowing sending message groups to specified recipient processes in a 3D spatial ABM, and thereby prevent excessive all-to-all communications in the simulation platform. In order to identify the location of the sender and the recipient agents across the parallel processes, we implement a simple grid-based structure for spatial information requests. Finally, we present comparative results using three partitioning methods: FLAME default geometric, FLAME default round-robin, and Zoltan Hypergraph Partitioning showing the advantages of our approach.

Original languageAmerican English
Pages (from-to)65-72
Number of pages8
JournalModelling and Simulation 2014 - European Simulation and Modelling Conference, ESM 2014
Publication statusPublished - 2014

Keywords

  • Agent-based simulation
  • FLAME
  • Graph partitioning
  • HPC
  • Load balancing
  • Message filtering
  • SPMD
  • Tuning
  • Zoltan

Fingerprint Dive into the research topics of 'Impact of message filtering on HPC agent-based simulations'. Together they form a unique fingerprint.

Cite this