Group Management Schemes for Implementing MPI collective Communication over IP-Multicast

Recent advances in multicasting present new opportunities for improving communication performance for clusters of workstations. Realizing collective communication over multicast primitives can achieve higher performance than over unicast primitives. However, implementing collective communication using multicast primitives presents new issues and challenges. Group management, which may result in large overheads, is one of the important issues. In this paper, we propose three group management schemes that can be used to implement MPI collective communication routines, static group management, dynamic group management, and compiler–assisted group management. We implement the three schemes in a prototype MPI library and evaluate the performance of the schemes. We further compare our multicast based implementation with an efficient unicast based MPI implementation, LAM/MPI. The results show that multicasting can greatly improve the performance of collective communication in many situations. In particular, combining multicasting with compiler– assisted group management can deliver very high performance.