Many- objective performance optimization in heterogeneous computing clusters

Abstract

In computing clusters, there are di erent performance metrics, which often appear to be con-

icting while being attempted to be optimized. For having such con icting cases along with experiencing existence of heterogeneous environment, it is often di cult for the cluster administrators to select the right number and right combination of machines. As a remedy to this situation, in this thesis, we develop a technique through which cluster administrators can select the right set of machines to enhance cluster performance. In our solution, we integrate both cooling energy consumption and empirical performance characterization of clusters. To the best of our knowledge, existing studies do not integrate these two simultaneously in solving many-objective optimization problem for clusters. We exploit a many-objective optimization approach based on NSGA-III algorithm to solve our cluster problem. Our technique attempts to simultaneously optimize many objectives such as computation time, computation energy, cooling energy, and utilization. Subsequently, we demonstrate through both real experimentation and simulation that our technique mostly performs better than optimization approaches existing in the literature. In this study, we integrate cooling energy while evaluating cluster performance. Cooling energy consumption is one of the most signi cant parts of total energy consumed by clusters and similar distributed systems. However, little e ort has been spent so far to integrate the cooling energy in simulators that are used for simulating the distributed systems. Therefore, we also perform integration of cooling energy consumption in a widely-known simulator of distributed systems namely SimGrid.