A new load-balancing strategy for the solution of dynamical large-tree-search problems using a hierarchical approach

  • Authors:
  • S. Crivelli;T. Head-Gordon

  • Affiliations:
  • Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720;Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720

  • Venue:
  • IBM Journal of Research and Development
  • Year:
  • 2004

Quantified Score

Hi-index 0.00

Visualization

Abstract

We describe a new load-balancing strategy, applied here to the protein structure prediction problem, for improving the efficiency of the hierarchical approach when dealing with coarse-grained problems associated with large tree searches. Unlike other load-balancing strategies that reassign load from the heavily loaded processors to the lightly loaded or idle ones, the proposed strategy changes the virtual communication tree among the processors as the computational tree changes. The strategy incurs minimal overhead and is scalable.