Implementação e avaliação de desempenho de dois algoritmos de balanceamento de carga para clusters híbridos

Abstract

This work presents an implementation and performance analysis of two load balancing algorithms for applications based on data parallelism, when these applications are executed in hybrid distributed memory environments. In this work, hybrid environment is defined as a computational environment composed by a set of processors and accelerators, in other words, processing elements that have different processing capabilities, such as CPUs, APUs, GPUs or other types of accelerators. The purpose of the load balancing algorithm is to equalize the computation time, i.e. the load balancing algorithm has to send to the processing elements a workload proportional to their processing power, so they can finish their executions at about the same time. In order to test and validate the load-balancing algorithms, a computational model of the Human Immune System (HIS) was used to describe the space-temporal response of some cells and molecules of the HIS in the presence of an antigen, represented in this work by the Lipopolysaccharide (LPS). Two versions of the load balancer have been developed. The first one, the static load balancer, maintains the load allocation in the devices until the end of the execution. The second one, the dynamic load balancer, may change the load allocation on devices during the execution. After the tests performed with the two versions of the load balancer, it can be concluded that, for the evaluated application, the dynamic version was more efficient than the static one.