This Small Business Innovation Research Phase I project will investigate techniques for dynamic load balancing in parallel finite element analyses. The techniques will be particularly suitable for distributed analyses on heterogeneous networks. In the area of high performance computing, significant efforts are underway to effectively use the idle computing power of workstation networks. One of the key issues in designing algorithms for such systems is that of load balancing. The proposed project offers to devise techniques for dealing with this problem. The prevalent balancing methods aim to achieve a balanced state at discrete instants in time by rebalancing the whole system. This project will explore methods to maintain a continually (near) balanced state by exploiting balancing opportunities as and when they occur. Imbalances that occur in distributed analyses due to the dynamically varying capabilities of participating processors and nonuniform adaptations of meshes (as a result of using solution adaptive techniques of analysis) will be dealt with. A two-level approach is suggested to efficiently correct these imbalances with minimal communication overhead while maintaining the data locality that is so important for the efficiency of distributed implementations of finite element type discretizations.
