Performance and Energy Consumption Prediction of Randomly Selected Nodes in Heterogeneous Cluster

Abstract

The heterogeneous computing cluster consists of a number of nodes that are different in their computing powers and their consumed energy. Executing parallel application over heterogeneous cluster produces various tasks’ execution times according to the heterogeneity of the nodes’ computing powers. Consequently, different task execution times in a synchronous application are wasting the energy consumed of the underlying parallel platform running them. However, the execution time prediction process of parallel applications over a dynamically selected set of heterogeneous nodes is significant in the area of parallel computing. The predicted execution time almost uses in the process of optimizing both the speedup and the consumed energy of the parallel application running on the selected set of heterogeneous nodes in the cluster. Therefore, prediction helps the algorithm to select in advance the best solution without the need for trial and error methods or other costly methods. This paper presents new models to predict the execution time and consumed energy of parallel application when a random set of heterogeneous nodes are selected. The performance and energy consumption results are predicted for message passing synchronous applications. Moreover, the experimental results are evaluated over SimGrid/SMPI simulator. The results show that the proposed prediction models give a very acceptable precision when they are compared with the real measurements.