Eecient Scatter Communication in Wormhole K-ary N-cubes with Multidestination Message Passing 1

This paper presents eecient ways to implement one-to-all scatter communication on k-ary n-cubes with multidestination message-passing. Earlier researchers have applied multidestination message-passing to support non-personalized communications. In this paper, for the rst time in the literature, we demonstrate that multidestina-tion message-passing can also be used to support eecient personalized communication. Current parallel systems use either a sequential tree-based (ST) scheme or a binomial tree-based (BT) scheme with unicast message-passing to implement scatter. First, we show that the ST scheme achieves the lower bound just for a narrow range of message lengths. Next, we show that the BT scheme has a better overall performance compared to ST but, it can never achieve the lower bound. Finally, we propose a sequential multides-tination tree-based (SMT) scheme using multidestination messages which minimizes the scatter latency as well as achieves the lower bound. This scheme allows data for multiple destinations to be grouped together and sent as a single message using multidestination multicast/broadcast worms. All three schemes are evaluated for a range of system sizes (k; n), communication start-up times (t s), propagation times (t p), and message lengths (l). For small message sizes on an 8x8 torus, it is shown that the SMT scheme can reduce the scatter latency up to a factor of 8 compared to the ST scheme and up to a factor of 1.6 compared to the BT scheme. For a range of system sizes and (t s =tp) ratios, it is shown that the SMT scheme performs better than the ST and BT schemes for a wide range of message lengths. Thus, these results demonstrate signiicant potential to be applied to design scalable collective communication libraries for current and future generation wormhole systems.