x86_rdtscp2_perf.c - compare execution time x86 RDTSCP instruction vs. |sched_getcpu()|
Posted by Anonymous on Sat 26th Nov 2022 15:49
raw | new post
view followups (newest first): x86_rdtscp2_perf.c - compare execution time x86 RDTSCP instruction vs. |sched_getcpu()| by Anonymous
modification of post by Anonymous (view diff)

rmainz@derfwpc5131:~/work/x86_rdtscp/try3_perf$ make tests
gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32
gcc -std=gnu17 -m64 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_64
ksh93 -c 'set -o xtrace ; \
        cat x86_rdtscp2_perf.c ; \
        cat Makefile ; \
        time ./x86_rdtscp2_perf_32 ; \
        time ./x86_rdtscp2_perf_64'
+ cat x86_rdtscp2_perf.c
/*
 * x86_rdtscp2_perf.c - compare execution time x86 RDTSCP instruction
 * vs. |sched_getcpu()|
 *
 * Compile with:
 * $ gcc -std=gnu17 -m32 -Wall x86_rdtscp2_perf.c -lpthread
 *
 * Written by Roland Mainz <roland.mainz@nrubsig.org>
 *
 */

#define _XOPEN_SOURCE 700
#define _GNU_SOURCE 1

#include <stdio.h>
#include <stdlib.h>
#include <errno.h>

#include <pthread.h>
#include <sched.h>

#include <x86intrin.h>


int main(int ac, char *av[])
{
        (void)puts("#start.");

        unsigned int A = 0;

#define USE_FIXED_CPU_NUMBER 1
#if USE_FIXED_CPU_NUMBER
        int             cpu;
        cpu_set_t       cpuset;
        pthread_t       thread;

        thread = pthread_self();
        cpu = sched_getcpu();

        CPU_ZERO(&cpuset);
        CPU_SET(cpu, &cpuset);

        if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset) == 0) {
                (void)printf("# pthread_setaffinity_np(), "
                        "thread fixed to cpu %d\n",
                        cpu);
        }
        else {
                /*
                 * If you specify a CPU number beyond the maximum number
                 * of CPUs, then |pthread_setaffinity_np()| can have a
                 * |errno == 0|
                 */
                perror("pthread_setaffinity_np");
        }
#endif /* USE_FIXED_CPU_NUMBER */

        struct timespec start, stop;
        long long accum;
        long long dummy = 0LL; /* use dummy variable to defeat optimiser */

        if (clock_gettime(CLOCK_REALTIME, &start) == -1) {
                perror("clock gettime");
                exit(EXIT_FAILURE);
        }

        for (long l = 0 ; l < 100000000 ; l++) {
                dummy += _rdtscp(&A);
                dummy += A;
        }

        if (clock_gettime(CLOCK_REALTIME, &stop) == -1 ) {
                perror("clock gettime");
                exit(EXIT_FAILURE);
        }

#define NSECPERSEC (1000000000LL)
        accum = (stop.tv_sec - start.tv_sec)*NSECPERSEC +
                (stop.tv_nsec - start.tv_nsec);
        (void)printf("# time needed for x86 rdtscp instruction:\t%20.20f\n",
                (double)(((long double)accum)/((long double)NSECPERSEC)));


        if (clock_gettime(CLOCK_REALTIME, &start) == -1) {
                perror( "clock gettime" );
                exit(EXIT_FAILURE);
        }

        for (long l = 0 ; l < 100000000 ; l++) {
                dummy += sched_getcpu();
        }


        if (clock_gettime(CLOCK_REALTIME, &stop) == -1) {
                perror("clock gettime");
                exit(EXIT_FAILURE);
        }

        accum = (stop.tv_sec - start.tv_sec)*NSECPERSEC +
                (stop.tv_nsec - start.tv_nsec);
        (void)printf("# time needed for |sched_getcpu()|:      \t%20.20f\n",
                (double)(((long double)accum)/((long double)NSECPERSEC)));

        (void)dummy;

        (void)puts("#done.");

        return EXIT_SUCCESS;
}
+ cat Makefile
#
# Makefile for x86_rdtscp2_perf.c - compare execution time x86 RDTSCP
# instruction vs. |sched_getcpu()|
#

all: \
        x86_rdtscp2_perf_32 \
        x86_rdtscp2_perf_64

# explicitly use -O0 to disable peephole optimisations
x86_rdtscp2_perf_32: x86_rdtscp2_perf.c
        gcc -std=gnu17 -m32 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_32

x86_rdtscp2_perf_64: x86_rdtscp2_perf.c
        gcc -std=gnu17 -m64 -Wall -g -O0 x86_rdtscp2_perf.c -lpthread -o x86_rdtscp2_perf_64


tests: \
        x86_rdtscp2_perf_32 \
        x86_rdtscp2_perf_64
        ksh93 -c 'set -o xtrace ; \
                cat x86_rdtscp2_perf.c ; \
                cat Makefile ; \
                time ./x86_rdtscp2_perf_32 ; \
                time ./x86_rdtscp2_perf_64'

clean:
        rm -f \
                x86_rdtscp2_perf_32 \
                x86_rdtscp2_perf_64
# EOF.
+ ./x86_rdtscp2_perf_32
#start.
# pthread_setaffinity_np(), thread fixed to cpu 1
# time needed for x86 rdtscp instruction:       1.32610883200000007065
# time needed for |sched_getcpu()|:             33.35152474899999930358
#done.

real    0m34.68s
user    0m22.13s
sys     0m12.53s
+ ./x86_rdtscp2_perf_64
#start.
# pthread_setaffinity_np(), thread fixed to cpu 0
# time needed for x86 rdtscp instruction:       1.23842487100000009370
# time needed for |sched_getcpu()|:             1.95299446999999992691
#done.

real    0m3.19s
user    0m3.19s
sys     0m0.00s