Reducing contention in multiple multicasts
Zhiyu Zhou, Zhimin Tang · 2002
Multicast operation plays a significant role in parallel computing. Most multicast algorithms that have been proposed for single multicast deliver poor performance when applied to multiple multicasts, because the destination sets of different multicasts often overlap, which leads to node contention in multiple multicasts. The paper presents two algorithms, SPTBM (Source Partitioned Tree-Based Multidestination) and SHTBM (Source Half-divided Tree-Based Multidestination), using source based information to partition the network and reduce contention in multiple multicasts. Although these two algorithms are based on TBM (Tree-Based Multidestination) designed for single multicast, simulation results show SPTBM is superior to TBM, and SHTBM outperforms TBM and SPTBM with the increase of system size. These algorithms demonstrate potential to implement fast and scalable collective communication operations.