On designing implementable algorithms for the linear reconfigurable mesh

Document Type

Conference Proceeding

Publication Date

12-1-2003

Abstract

Although algorithms for the reconfigurable mesh (R-Mesh) are fast, they are very difficult to implement because most algorithms employ buses that span Ω(N) processors. On such buses, the constant bus-delay assumption, that is central to all R-Mesh algorithms, does not hold. In this paper, we consider a powerful restriction of the R-Mesh, called LR-Mesh, and show that a large class of fundamental LR-Mesh algorithms can be efficiently implemented using limited delay buses. We introduce a new measure of bus delay, called "bends-cost", and describe an LR-Mesh implementation for which bends-cost is a faithful measure of the actual bus delay. Next, we show that an important class of LR-Mesh algorithms (that includes algorithms for prefix sums, multiple addition, and sorting) can be implemented efficiently using buses whose delay is at most D, a parameter of the design. In particular, if the technology used can support a delay of D = Nε for an arbitrarly small constant ε > 0, then the running times of these algorithms are within a constant of their idealized LR-Mesh counterparts.

Publication Source (Journal or Book title)

Proceedings of the International Conference on Parallel and Distributed Processing Techniques and Applications

First Page

241

Last Page

246

This document is currently not available here.

Share

COinS