Energy-Optimization Scheduling of Task Dependent Graph on DVS-Enabled Cluster System

Abstract

At present, power management in High-Performance Computing (HPC) environment is becoming a hot topic owning to its high operation cost, low reliability and environmental impact. In this paper, we investigate energy minimization scheduling algorithm of data dependent tasks in DVS-Enabled cluster system. Considering the data-intensive characteristics, the proposed EOTD (Energy Optimization scheduling for Task Dependent graph) algorithm adopts task clustering to reduce data transmission time and communication energy consumption. In order to decrease dynamic power of processing elements, it uses one of the power-saving techniques in system level—Dynamic Voltage Scaling while not violating the deadline users specify. Moreover, on the premise that application execution is predictive and exclusive for processing elements, we employ Dynamic Power Management and Binary Search technique to reduce the static power of idle processing elements and last find the optimal number of processing elements. EOTD algorithm not only optimizes the energy consumption of task dependent graph, but also satisfies the QoS requirements service level agreement gives. Compared with VM and LJ-VM algorithm, experimental results demonstrate that EOTD algorithm can achieve larger energy optimization in less optimizing time.