Computational grids are becoming increasingly vital in organizations with ever growing IT infrastructure and the need to increase the productivity of the computing infrastructure by ensuring optimal throughput for their computational jobs. Key to computational grids in the load balancer and job scheduler that is involved in decision making about when and which node to basically use to execute a job/task submitted to the grid. Most of the existing grids use a load function that evaluates the existing resources on the nodes, accesses the resource requirements of the task submitted and decide whether to withhold the job in the queue or schedule it on a node where the resources are available for the job. This decision making process becomes more challenging with jobs that are long duration, I/O intensive and resource requirements vary at different times during the task execution. Most current grid engines factor in the maximum requirement as stated at the time of job submission and are not good at analyzing the variation in resource requirements based on past history of the same job execution and use the information gathered in the decision making process. In this paper, we try to analyze how we can change the load balancing function to introduce more statistical analysis of history of past jobs in the scheduling decision process thereby ensuring we do not end up in trashing situations for I/O intensive jobs while at the same time utilize the available resources as efficiently as possible.