Can Scatter Communication Beneet from Multidestination Message Passing? 1

Multiple research directions in the past have emphasized the beneets of multidesti-nation message passing to support eecient non-personalized communication like broadcast and multicast. In this paper, we investigate whether personalized communications like scatter can beneet by using multidestination message-passing support, if available in a system. In order to use such support, we propose a new approach where personalized data for a set of destinations can be grouped together and sent as a single multidestination broadcast/multicast worm. This approach is shown to use less number of communication steps while achieving the lower bound for scatter communication. Using this approach, we propose two schemes, sequential multidestination tree-based (SMT) and hierarchical multidestination tree-based (HMT), to implement one-to-all scatter in k-ary n-cube wormhole systems. Performance of these two schemes are compared with two commonly used unicast-based schemes, sequential tree-based (ST) and binomial tree-based (BT). Analytical and simulation studies are carried out to determine the scatter latency under the four schemes for k-ary n-cube systems with varying communication start-up times (t s), propagation times (t p), and message lengths (l, size of data being sent to each destination). The study indicates that for l ts ktp , the scatter can be best implemented by the HMT scheme. Similarly, the SMT scheme is shown to outperform where ts ktp < l ts tp. For l > ts tp , both ST and SMT perform better than other algorithms. Simulation results on an 8 8 system with t s =10.0 microsec show that the HMT and SMT schemes reduce scatter latency up to a factor of 30 and 8 compared to the ST scheme for small message lengths. These results connrm that scatter communication can in fact take advantage of multidestination message passing support.