Generating an efficient broadcast sequence using reflected Gray codes

One-to-all broadcast is frequently implemented on multicomputers using a binomial tree communication pattern. The source processor, at the root of the broadcast tree, completes the broadcast before the processors at the leaves of the tree. This imbalance can degrade performance when a data-parallel algorithm requires a series of broadcasts from different source processors. In this paper, we show that using a reflected Gray code to order broadcasts minimizes the communication time.