Genetic algorithm based task reordering to improve the performance of batch scheduled massively parallel scientific applications

摘要

The growth in size of networked high performance computers along with novel accelerator-based nodernarchitectures has further emphasized the importance of communication efficiency in high performancerncomputing. The world’s largest high performance computers are usually operated as shared userrnfacilities due to the costs of acquisition and operation. Applications are scheduled for execution in arnshared environment and are placed on nodes that are not necessarily contiguous on the interconnect.rnFurthermore, the placement of tasks on the nodes allocated by the scheduler is sub-optimal, leadingrnto performance loss and variability. Here, we investigate the impact of task placement on the performancernof two massively parallel application codes on the Titan supercomputer, a turbulent combustionrnflow solver (S3D) and a molecular dynamics code (LAMMPS). Benchmark studies show a significantrndeviation from ideal weak scaling and variability in performance. The inter-task communicationrndistance was determined to be one of the significant contributors to the performance degradation andrnvariability. A genetic algorithm-based parallel optimization technique was used to optimize the taskrnordering. This technique provides an improved placement of the tasks on the nodes, taking intornaccount the application’s communication topology and the system interconnect topology. Applicationrnbenchmarks after task reordering through genetic algorithm show a significant improvement in performancernand reduction in variability, thereby enabling the applications to achieve better time to solutionrnand scalability on Titan during production.