Static Task Scheduling for Heterogeneous Distributed Computing Systems with Memory Constraints

• Main Authors: Chang-ChihLiu, 劉昶志
• Other Authors: Da-Wei Chang
• Format: Others
• Language: en_US
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/11115762411488195551

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 98 === Effective task scheduling of parallel applications represented by directed acyclic graph (DAG) is critical for obtaining high performance in heterogeneous distributed computing systems (HDCSs). As the problem of finding optimal scheduling algorithm has been shown to be NP-complete in general cases, many heuristic algorithms for scheduling on HDCSs have been proposed recently. However, none of them consider the case where processing elements (PEs) have memory constraints which prevent operating system from being installed on PEs to provide memory management. Tasks have to be stored on a specific physical memory address and loaded on demand, which means that task schedule in such computing systems requires the consideration of the loading time of tasks associated with their code size besides the heterogeneity of PEs and the inter-PE communication overhead. For identifying different code size of tasks, the nodes of DAG are colored with different color types according to their functionalities, called colored-DAG. This thesis presents a new static list-based scheduling algorithm, called the Heterogeneous Loading Time Aware (HLTA) algorithm, which has three distinctive features. First, the algorithm partitions the priority list by layer heuristic of colored-DAG. Second, the algorithm uses novel greedy mechanism to reorder and schedule the identical color of nodes to the same PE as could as possible. Finally, a braking mechanism is used to refine the greedy reordering when scheduling a high color ratio colored-DAG. The comparison study, based on randomly generated graphs, shows that HLTA algorithm significantly surpasses the Heterogeneous Earliest-Finish-Time (HEFT) algorithm in terms of quality of schedules and average schedule length ratio improvement.