Deadlock-Free Message Routing in Multiprocessor Interconnection Networks
IEEE Transactions on Computers
Data communication in hypercubes
Journal of Parallel and Distributed Computing
Cube structures for multiprocessors
Communications of the ACM
Nearest Neighbors in Random Subspaces
SSPR '98/SPR '98 Proceedings of the Joint IAPR International Workshops on Advances in Pattern Recognition
A journey into multicomputer routing algorithms
PAS '95 Proceedings of the First Aizu International Symposium on Parallel Algorithms/Architecture Synthesis
Hi-index | 14.98 |
The author describes efficient manipulations of general vectors on the synchronous binary n-cube structure. A general vector is defined as a set of elements stored in consecutive processors with arbitrary length and starting point, and one element per processor. New routing methods for manipulating general vectors are presented. The author focuses on six major vector manipulating functions: merge, split, rotation, reverse, compression, and expansion. They are frequently used to extract and structure data parallelism in image processing and parallel solutions of linear systems. It is observed that varying the dimension order is a key to collision-free vector manipulations. A formal network model is developed for determining when link collisions occur. With the aid of this network model dimension orders yielding collision-free routine for the six manipulating functions are identified. Collision-free routing allows data communication to complete in the optimal time-single network cycle. The dimension orders are easy to encode and decode, and they are feasible for physical implementation.