Load-Balancing for Parallel Derivative-Free Optimization Methods on Shared Computing Networks

Abstract

Over the past decade, the area of high performance computing has grown dramatically in Canada, in part through the growth of shared computing networks such as WestGrid, ACEnet, and SHARCNET. An important property of these types of multi-user networks is the role of the job-scheduler, which places priorities on and orders incoming requests for computing resources. When designing parallel algorithms for use on such networks, one can improve the completion time by taking advantage of prior knowledge about how the job scheduler defines its priority placement. In this talk, we present implementation details on a parallel derivative-free optimization package, designed for use on multi-user computing networks.

Details