Dynamic workload optimisation on NUMA and heterogeneous architectures

Abstract

This thesis faces the challenges of dynamic workload optimisation and workload balancing in two different problems: in conventional systems using heterogeneous (CPU and GPU) parallelism, and in NUMA systems. On one hand, a library named IHP is proposed. Dynamically, CPU and GPU performance are evaluated so computational workload is divided accordingly. Results show that execution times can be improved between 3% and 55% depending on the code and the performance of the computing units. On the other hand, a tool for migrating threads and memory pages in NUMA systems has been developed. This tool incorporates several algorithms that, considering performance measurements, decide whether migrations are required. Experiments show that performance can be improved by up to 47%, particularly in multi-tasking scenarios.