MapReduce is a programming model that enablesefficient massive data processing in large-scale computing environmentssuch as supercomputers and clouds. Such large-scalecomputers employ GPUs to enjoy its good peak performanceand high memory bandwidth. Since the performace of each job isdepending on running application characteristics and underlyingcomputing environments, scheduling MapReduce tasks onto CPUcores and GPU devices for efficient execution is difficult. Toaddress this problem, we have proposed a hybrid schedulingtechnique for GPU-based computer clusters, which minimizesthe execution time of a submitted job using dynamic profiles ofMap tasks running on CPU cores and GPU devices. We haveimplemented a prototype of our proposed scheduling techniqueby extending MapReduce framework, Hadoop. We have conductedsome experiments for this prototype by using a K-meansapplication as a benchmark on a supercomputer. The resultsshow that the proposed technique achieves 1.93 times faster thanthe Hadoop original scheduling algorithm at 64 nodes (1024 CPUcores and 128 GPU devices). The results also indicate that theperformance of map tasks, including both CPU and GPU tasks,is significantly affected by the overhead of map task invocationin the Hadoop framework.
各種検索
サポート
ヘルプ