Implementation of parallel numerical routines using broadcast communication schemes
In this paper, we present the theoretical analysis and the implementation of fundamental communication routines for broadcasting data on hypercube distributed-memory computers. We introduce desynchronized strategies, where computations and communications are mixed. We take into account some communication limits of Transputer-based computers to study specific communications routines. We finally present a parallelisation of the Jordan Diagonalization based on these broadcast routines.
Unable to display preview. Download preview PDF.
- [AEOS]C. Aykanat, F. Ercal, F. Ozguner & P. Sadayappan: "Iterative Algorithms for Solution of Large Sparse Systems of Linear Equations on Hypercubes", IEEE Trans. on Computers, vol.37, no12, 1988.Google Scholar
- [BTV]J-Y. Blanc, D. Trystram & G. Villard: "Desynchronized Communication Schemes on Distributed-Memory Achitectures" DMCC5 Charleston SC, 1990Google Scholar
- [Frai]P.D. Fraignaud: "Performance Analysis of broadcasting in Hypercubes", in Hypercube and Distributed Computers, Rennes, Elsevier Science Publisher, 1989.Google Scholar
- [RoTV]Y. Robert, B. Tourancheau & G. Villard: "Data Allocation Strategies for the Gauss and Jordan Algorithms on a Ring of Processors", Information Proc. Letters 31, pp 21–29, 1989.Google Scholar
- [StWa]Q. F. Stout & B. Wager: "Intensive Hypercube Communication I: Prearranged Communication in Link-Bound Machines", Technical Report CRL-TR-9 87, University of Michigan, 1987.Google Scholar