Scheduling Modifications for Improvement Performance on a Beowulf Cluster Single-Node

摘要

This paper studies the improvement of performance of a single compute-node of the Beowulf cluster Virgo 2.0. We want to implement kernel modifications to the Linux scheduler which will improve the performance of High Performance Linpack (HPL) benchmark runs. We compare the performance when spawning processes for two processing cores in a local processor up to all eight cores in two processors. The processing hardware of the compute-node was modeled using Instruction per Cycle (IPC) in order to identify advantages and limitations of the hardware. The modifications were then focused to avoid the bottlenecks of performance of each process by changes of process policy, grouping, load balancing, processing affinity and termination. The results showed improvements of performance for all processes at different dimension, block and problem sizes of HPL.