Analysis of parallel inference machines to achieve dynamic load balancing

A parallel inference machine (PIM) prototype modelled on loosely coupled clusters was simulated on a hardware simulator. Performance of the PIM prototype is limited by suspension/resumption overhead in the fine granularity region and by low utilization, due to load distribution imbalance, in the coarse granularity region. It is shown that the load dispatch strategy in which loads are dispatched to the cluster with minimum loads at an AND-fork time is effective on the loosely-coupled cluster level, resulting in 20% higher performance than in the random dispatch strategy, and that the load status modification delay should be less than half of the reduction time to limit the degradation to within 5%.<>