BASIC ROUTINES FOR THE RANK-2K UPDATE - 2D TORUS VS RECONFIGURABLE NETWORK

Our aim is to provide the Rank-2k update on different parallel machine s. In this paper, we compare the performance obtained on a fixed 2D to rus topology and on a reconfigurable system. This results in the devel opment of two basic communication subroutines, namely scattering and m atrix-transposition. And two basic computation subroutines, namely mat rix product and Rank-2k update (both belongs to the level 3 BLAS). The preceding distributed-memory machines generation used fixed networks such as grid, multidimensional tori or hypercubes. Today, vendors prop ose machines with networks that can be reconfigured during program exe cution are available. A large number of possibilities are therefore av ailable to the programmer, who can adapt his configuration during runt ime to suit both best algorithm and data distribution. This dynamical reconfiguration obviously introduces an overhead through the setting o f the network switch(es). This overhead must be taken into account in the cost of the whole computation. Using complexity analysis and exper iments on a machine issued from the SuperNode Esprit project, we compa re for each subroutine studied, one method using a 2D torus topology a nd another one using a dynamically reconfigurable network. For each so lution, performance evaluations and experiments exhibit interesting sp eed-ups for the algorithms on a reconfigurable network.