HP-UX Floating-Point Guide

Chapter 7 189

Performance Tuning

Cache Aliasing

Data cache aliasing is a more common performance problem than

instruction cache aliasing, but it is also easier to deal with. If you suspect

that your application is experiencing data cache performance problems

on scalar or small vectors of data, you can usually correct the situation

by rearranging the order in which your variables are allocated in

memory. In Fortran you can do this by reordering common block

assignments.

If your application uses very large arrays of data, hundreds of kilobytes

or megabytes in length, than chances are good that you will experience

data cache aliasing problems no matter how your data arrays are

allocated in memory, simply because the arrays are too big to ﬁt in the

data cache all at once. In this case, you may be able to improve

performance by using the +Odataprefetch option, either alone or in

conjunction with +Ovectorize. (See “Optimizing Your Program” on

page 171 for information about these options.)

Another way to improve performance for very large arrays is to increase

the locality of references to your data. This technique, sometimes called

tiling, requires selecting an algorithm that processes as much data as

possible while the data is resident in the cache so as to minimize the

number of times the data must be re-cached later. The routines in the

BLAS library use tiling when they operate on large matrices of data. For

example, in multiplying two large matrices, each matrix is cut into tiles,

which are processed in pairs. The size of the tiles is determined at run

time by the size of the matrices and the size of the data cache on the

system executing the application.