Becoming a better host through origami: A mesh can be more than rows and columns

Endowing a communication network with the ability to realize arbitrary communication patterns is an expensive proposition, both in hardware and in system software. One might instead ask whether a system can be built that performs well for a given application program. In this paper we look at the question of when a set of communication patterns is suitable for fast realization on a given network. In particular we look at which patterns are realizable quickly on a mesh. Contrary to common wisdom, transpose is efficiently realizable on a mesh. However, some other important patterns, such as perfect shuffle, are not