We present our experience in implementing several irregular problems using a high-level actor language. The problems studied require dynamic computation of object placement and may result in load imbalance as the computation proceeds, thereby requiring dynamic load balancing. The algorithms are expressed as fine-grained computations providing maximal flexibility in adapting the computation load to arbitrary parallel architectures. Such an algorithm may be composed of different partitioning and distribution strategies (PDSs) to result in different performance characteristics. The PDSs are implemented for specific data structures or algorithms and are reusable for different parallel algorithms. We demonstrate how our methodology provides portability of algorithm specification, reusability and ease of expressibility.
展开▼
