Communication overhead on the CM5: an experimental performance evaluation

The authors present experimental results for communication overhead on the scalable parallel machine CM-5. It is observed that the communication latency of the data network is 88 mu s. It was also observed that the communication cost for messages that are a multiple of 16 bytes is much smaller than for messages that are not, and therefore, for better performance, a user should pad messages to make them a multiple of 16 bytes. The authors also studied the communication overhead of three complete exchange algorithms. For small message sizes, the recursive exchange algorithm performs the best, especially for large multiprocessors. However, for large message sizes, the pairwise exchange algorithm is preferable. Finally, the authors studied two algorithms for one-to-all broadcast: the linear broadcast algorithm and the recursive broadcast algorithm. Linear broadcast does not perform well; the recursive broadcast algorithm performs well.<
>