J. Miksatko and D. Andresen (USA)
scheduling, neural networks, J2EE, load balancing
We propose a novel adaptive scheduling policy built on the prediction of Web service CPU requirements based on their run-time arguments. The prediction offers a relatively pre cise load indicator and the advantage of obtaining load in formation at any time without querying the nodes. A his tory of service execution times with respect to their inputs is collected during a typical run of the system, which is then statistically analyzed and a predictor based on neural networks is created. A fully automated learning system for predictor generation is proposed. No additional knowledge about service inputs and its behavior is required. A simulation framework based on Java RMI was de veloped to evaluate proposed algorithms and compare them to existing location policies. A specialized set of tasks was used to investigate the predictor generation and tar get learning-specific issues, including tasks from the Jav aGrande benchmark suite. The evaluation has shown very promising results. The proposed automated learning sys tem confirmed the accuracy of CPU time prediction, and the load balancing system as a whole has shown a signif icant improvement in the mean response time (up to 32% and 44% for round-robin and random scheduling policy, re spectively). Moreover, the standard deviation of the mean response time decreased substantially (over 70% in some cases).
Important Links:
Go Back
