Parallel Computer Architecture

532

CHAPTER

7 Scalable Multiprocessors

tions where most time is spent and extracting the usage characteristics of the

application to determine its requirements. This section briefly compares perfor-

mance on our case study machines for two of the NAS parallel benchmarks in the

NPB2 suite (NAS Parallel Benchmarks 1998).

The LU benchmark solves a finite difference discretization of the 3D compressible

Navier-Stokes equations used for modeling fluid dynamics through a block-lower-

triangular, block-upper-triangular approximate factorization of the difference

scheme. The LU factored form is cast as a relaxation and is solved by symmetric suc-

cessiv ...