Improving the Load Balance of Parallel Adaptive Mesh Refined Simulations

In order to quickly take advantage of petascale machines, which are in the near future, we must improve the performance of current codes. In the past, general-purpose adaptive codes, such as Uintah, have been shown to scale to thousands of processors. However, in order to take advantage of petascale machines these codes will have to scale to hundreds of thousands of processors. One important aspect of parallel codes is the load balance. Load imbalance can greatly hinder performance and scalability at large numbers of processors. In this paper we describe two improvements to the existing load balancer within Uintah. The first improvement involves using timings gathered at run-time in order to predict the cost of work with a high degree of accuracy allowing the work to be more effectively load balanced. The second improvement proposes a method to distributed work according to a space-filling curve more effectively by attempting to minimize the maximum amount of work assigned to any one processor. The effect of this minimization technique has been compared to load balancers within Zoltan and has shown that significant improvements are possible. Improving the Load Balance of Parallel Adaptive Mesh Refined Simulations Justin Luitjens, Qingyu Meng, Martin Berzins, Tom Henderson