History-based adaptive load sharing heuristics

The conventional load representation method expressing a load as queue length has a weakness in that it does not reflect execution characteristics of processes. To solve the inaccuracy of this conventional method of load representation, we propose the weighted load representation method which assigns graded weights to processes based on the execution characteristics of processes. The proposed method uses history information obtained during process scheduling. In order to derive the upper bound and the lower bound of the load in a node, the finite population queuing model is used. We devise heuristics for adaptive load sharing using the proposed representation method and derived upper and lower bounds. Simulation results show that GRRs of weight functions log i and i are improved 2.37-3.14 times compared with no load sharing and 14-24% as many as GRRs of the conventional load representation method.<>