Graph-based automatic dynamic load balancing for HPC agent-based simulations

Claudio Márquez*, Eduardo César, Joan Sorribes

*Corresponding author for this work

Research output: Contribution to journalArticleResearchpeer-review

3 Citations (Scopus)

Abstract

The main problem of Agent-Based Modelling (ABM) simulations in High Performance Computing (HPC) is load imbalance due to a non-uniform distribution of the agents that may generate uneven computation and increase communication overhead, inhibiting the efficiency of the available computing resources. Moreover, the agents’ behaviours can considerably modify the workload at each simulation step thereby affecting the workload progression of the simulation. In order to mitigate such problems, automatic mechanisms for dynamically adjusting the computation and/or communication workload are needed. For this reason, we introduce an Automatic Dynamic Load Balancing (ADLB) strategy to reduce imbalance problems as the simulation proceeds. The ADLB tunes the global simulation workload migrating groups of agents among the processes according to their computation workload and their message connectivity map modelled using a Hypergraph. This Hypergraph is partitioned using the Zoltan Parallel HyperGraph partitioner method (PHG). In addition, to prevent excessive all-to-all communications, the ADLB uses filtering routines to send message groups to specified recipient processes in a simple 3D grid-based structure. Our method has been tested with a biological ABM using the framework Flexible Large-scale Agent Modelling Environment (Flame), obtaining a significant impact on the application performance.

Original languageAmerican English
Pages (from-to)405-416
Number of pages12
JournalLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
DOIs
Publication statusPublished - 2015

Keywords

  • Agent-based simulation
  • Flame
  • Graph partitioning
  • HPC
  • Load balancing
  • Message filtering
  • Performance tuning
  • SPMD

Fingerprint Dive into the research topics of 'Graph-based automatic dynamic load balancing for HPC agent-based simulations'. Together they form a unique fingerprint.

Cite this