Parallel Computer Architecture

8.12 Exercises 677

The input data set is a 1,024 x 1,024 matrix of double-precision floating-point

numbers (i.e., nrows in

1,024),

decomposed so that each processor is respon-

sible for generating a contiguous block of rows in the transposed matrix AT

(i.e.,

a receiver-initiated transpose). Ignoring the contention problem caused

by all processors first going to processor 0, what is the major performance

problem with this code? What technique would you use to solve it? Restruc-

ture the code to alleviate all performance problems as much as possible. Write

the entire restructured loop.

8.24 Consider a hierarchical bus-based system with a centralize ...