Energy Consumption of a Duplication based Scheduling Approach: A Case Study

Ahmed Ebaid, Reda Ammar, Sanguthevar Rajasekaran, and Rehab ElKharboutly

Keywords

task duplication, static scheduling, inter-processor communications, high performance computing

Abstract

Inefficient utilization of a homogenous cluster has a huge impact on the amount of CO2 emitted. Duplication based scheduling approaches have only focused on minimizing the schedule length, completely ignoring the additional energy consumed by each processor due to task duplication, and the energy dissipated by the processor interconnects. Our former algorithm, Recursive Critical Path Approach (RCPA*), efficiently strikes a balance among schedule length, the inter-processor communications, and the processor loads. In this paper, we evaluate the energy consumption for static schedules generated when using the (RCPA*)algorithm and compare it with the energy consumed by the LG algorithm. A well-known energy consumption model, and four different cluster interconnection technologies are used in this study. Extensive simulations using random Directed Acyclic Graphs (DAGs) show that the CPU energy consumption when using the (RCPA*) is 20% less than the energy consumed when using the LG approach. Moreover, the total energy dissipated by the network interconnects when using the (RCPA*) algorithm is 17% less than the energy dissipated using the LG algorithm for communication-intensive parallel applications. Also, the schedules generated for the parallel $Gaussian$ elimination, and the Fast Fourier Transform (FFT) task graphs have shown a reduction in energy consumption of 28% and 18%, respectively, when using the (RCPA*) algorithm.

Important Links:



Go Back