No abstract

The source code, executables and documents can be downloaded from http://abcgrid.cbi.pku.edu.cn

Many-Task Computing (MTC) is a widely used computing paradigm for large-scale task-parallel processing. One of the key issues in MTC is to schedule a large number of independent tasks onto heterogeneous resources. Traditional task-level scheduling heuristics, like Min-Min, Sufferage and MaxStd, cannot readily be applied in this scenario. As most of MTC tasks are usually fine-grained, the resource management overhead would be prominent and the multi-core nodes might become hard to be fully utilized. In this paper we propose an application-level scheduling with task bundling approach that utilizes the knowledge of both applications and tasks to overcome these difficulties. Furthermore we adapt the traditional task-level heuristics to our model for MTC scheduling. Experimental results show that these application-level scheduling approaches, when equipped with task bundling, can deliver good performance for Many-Task Computing in terms of both Makespan and Flowtime.

Abstract.ChinaGrid aims at building a public service system for Chinese education and research. ChinaGrid Support Platform (CGSP) is a grid middleware developed for the construction of the ChinaGrid. Function modules of CGSP for system running are Domain Manager, Information Center, Job Manager, Data Manager, Service Container and Security Manager. Developing tools for gird constructor and application developers consist of Service Packaging Tool, Job Defining Tool, Portal Constructor and Programming API. CGSP architecture is introduced first. Then, CGSP function modules and developing tools are described. At last, job executing flow in CGSP is also put forward in the paper.

No abstract