Cypress CSC-1200T Guía de usuario descargar pdf (Pagina 107)

Chapter 10. Miscellaneous notes 107

after = cpused();

utime = after - before;

printf("ct[10][10] is %.6f\n", ct[10][10]);

printf("CPU time in user space = %ld clock ticks\n",

utime);

exit(0);

}

See Section 6.3 on page 54 for more details on calling Fortran routines

(here SGEMM) from C.

Here is an example of compiling and executing this program:

t3e% cc matmul.c

t3e% timex mpprun -n 2 ./a.out

ct[10][10] is -345.015608

CPU time in user space = 371395739 clock ticks

CPU time in user space = 371379536 clock ticks

seconds "clocks"

real 3.979838 (1193951546)

user 2.063669 (619100700)

sys 0.254990 (76497000)

Here we executed the program identically on two processors.

10.2 Parallel performance prediction

Several diﬀerent models can be used to measure the scalability of a

parallel program. Depending on your application and preferences, you

may want to use parallelism to decrease execution time, to run bigger

models or to optimize the speedup of parallel processing.

The so-called Amdahl’s law applies to a ﬁxed model size when you are

using diﬀerent numbers of processors. This model supposes that you

can split the program in two parts, sequential and parallel. The sequen-

tial part takes W

seconds in all cases. The parallel part takes W

seconds, where the W

is the size of the parallel task and p is the num-

ber of parallel processors.

Speedup S

is deﬁned as the ratio of the time on one processor divided

by the time on p processors:

+ W

1 2 ... 102 103 104 105 106 107 108 109 110 111 112 ... 123 124

Sin comentarios

Cypress CSC-1200T Guía de usuario Pagina 107