K. Kiran and V. Sridhar (India)
Distributed computation, Grid computing, Gridutilization, Parallelization
In a dynamic environment wherein multiple jobs are being processed on a grid, the load on the grid changes on a continuous basis. One of the objectives of the proposed approach is to enhance the utilization of the grid in such dynamic situations. While distributed applications are amenable for dynamic redistribution across grid nodes, we propose an application independent approach in which jobs are split and merged as the demand, for the grid resources, varies over a period of time. We use a clustering technique to identify the practically feasible parallelism in a job and this identified parallelism is exploited by defining an overlay network (called grid overlay network) to monitor grid status and job status; defining cache control logic for automatic translation of memory read/write to message send/receive pairs; defining a way for synchronizing multiple requests for accessing shared, global variables stored in a distributed memory; identifying the conditions under which jobs can be split or merged; automatic splitting and merging of a job across multiple nodes of the grid; and setting up shared variables in a cache block on an overlay node. The proposed approach aims to achieve the resource allocation strategy of providing what the grid has and not what a job wants thereby always maximally utilizing the grid resources.
Important Links:
Go Back
