Energy-Aware Scheduling of Parallel Application in Hybrid Computing System  

作　　者：MA Yan GONG Bin GUO Zhihong CHEN Yufeng ZOU Lida 

机构地区：[1]State Grid Shandong Electric Power Research Institute [2]School of Computer Science and Technology, Shandong University

出　　处：《Chinese Journal of Electronics》2014年第4期688-694,共7页电子学报（英文版）

摘　　要：Energy management is emerging as an important issue for High performance computing(HPC)owning to high operational cost and low reliability. Compared with low-power architectural approach, energyaware scheduling based on Dynamic voltage scaling(DVS)and Dynamic power management(DPM) is regarded as a promising way since it is practical and low-cost. At present,most studies focus on pure DVS or non-DVS environment, while most high performance computing systems are hybrid non-DVS/DVS platforms. We propose an energyaware scheduling algorithm for parallel application to consider both DVS and non-DVS characteristics of hybrid system. We present the rule of task assignment, make analysis on DVS and DPM technique and give their mathematical formulation, which maintains makespan optimization and energy conservation. The clustering and merging algorithm, and priority computation method consider the situation of resource constraints. The extensive simulations demonstrate that the proposed algorithm has stronger ability of energy saving and time optimization than Heterogeneous earliest finish time(HEFT), Energy-efficient task duplication scheduling(EETDS) and Heterogeneous energy-aware duplication scheduling(HEADUS) algorithm no matter for synthetic workload or realistic workload.Energy management is emerging as an important issue for High performance computing （HPC） owning to high operational cost and low reliability. Compared with low-power architectural approach, energyaware scheduling based on Dynamic voltage scaling （DVS） and Dynamic power management （DPM） is regarded as a promising way since it is practical and low-cost. At present, most studies focus on pure DVS or non-DVS environment, while most high performance computing systems are hybrid non-DVS/DVS platforms. We propose an energyaware scheduling algorithm for parallel application to consider both DVS and non-DVS characteristics of hybrid system. We present the rule of task assignment, make analysis on DVS and DPM technique and give their mathematical formulation, which maintains makespan optimization and energy conservation. The clustering and merging algorithm, and priority computation method consider the situation of resource constraints. The extensive simulations demonstrate that the proposed algorithm has stronger ability of energy saving and time optimization than Heterogeneous earliest finish time （HEFT）. Energy-efficient task duplication scheduling （EETDS） and Heterogeneous energy-aware duplication scheduling （HEADUS） algorithm no matter for synthetic workload or realistic workload.

关 键 词：Energy efficiency Dependent tasks Multiprocessor systems. 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]