Most of the existing solutions to resource management and task-scheduling in computation grid are based on a traditional client-server model, employing a central administrative server/manager.Since a computation grid provides a distributed, multi-domain computational resource, we argue that it should not have a single central authority for resource management and task scheduling. Thus, we propose a P2P based framework for management of the computing resources and scheduling of tasks that make use of the resources in the grid

In our project, a generic architecture for metascheduler on peer sites, called PGS (Peer in Grid Scheduler), and a task scheduling framework based on PGS for computation grid are developed. PGS facilitate the integration of the P2P approach to task scheduling into the grid environment. Both push and pull methods are used for allocating the tasks to peers with the support of load balancing and fault tolerance. A prototype of the proposed architecture and scheduling mechanism has been developed and experiments have been conducted with the prototype.

In the proposed P2P-based approach, there is a metascheduler at every site and jobs are submitted to the local metascheduler where the job originates. The metaschedulers interact directly with each other (through GPIS) to collect load information and to make scheduling decisions All tasks of jobs submitted or captured are stored in an execution queue of the PGS. A local Job is broken into a number of small tasks and being imported in the queue noted as different status (LO,DA,DP,CA,CP), which have different execution priorities. The task dispatch mode can be either pull or push, and PGS support fault tolerance by clonning the tasks between peers rather than move them. Figure 1. shows the Architecture of the System Figure 2 and 3 . shows the senarios of the push and pull mode of job dispatch/capture

Preliminary experiments have been performed using the prototype. The experimental result shows that the combination of pull and push techniques achieved a faster convergence in speedup than using the push strategy alone, Table 4, shows the experimental result

The advantages of the proposed approach over the existing solutions include offloading/reducing the costs incurred by the central scheduler architecture, user-centric and user-manageable scheduling policy, and heterogeneous queuing system.

home>>