Question

我在NUMA计算机上运行的程序中存在openMP性能问题。

该平台是一台8核机器。（NUMA有2个节点，每个节点4个核心。）

我的程序剂量中有一个功能A： 1.开始时CPU掩码为0xFF（8个核心） 2.在中间，它使用掩码0xF0调用后台的子程序，并将当前pid设置为0x0F。该函数与子程序并行执行某些操作。 3.后台子程序完成后，它会加入并执行CPU_ZERO以重置绑定。

没有固定亲和力的所有openMP代码在函数A之后变慢。然后我在功能A之后将其重置为0xFF并且它可以工作。经济放缓已经消失。这是否意味着openMP的掩码0x0与0xFF不同？

使问题清楚。我在下面写了一个示例代码

原来你不能设置CPU。是否意味着取消绑定CPU意味着将掩码设置为所有cpus？

#define _GNU_SOURCE             /* See feature_test_macros(7) */
#include <sched.h>
#include <sys/sysinfo.h>
#include <stdio.h>


int main() {
    cpu_set_t cpu_set;
    int i, ret;
    const int num_cores = get_nprocs_conf();
    CPU_ZERO( &cpu_set );
    ret = sched_getaffinity( 0, sizeof(cpu_set), &cpu_set );


    printf(" default cpu_set: ");
    for( i=0; i< num_cores; i++ ){
        if( CPU_ISSET( i, &cpu_set )) printf( "%d ", i);
    }
    printf("\n");

    CPU_ZERO( &cpu_set );

    CPU_SET( 0, &cpu_set );
    CPU_SET( 1, &cpu_set );
    CPU_SET( 2, &cpu_set );
    CPU_SET( 3, &cpu_set );

    ret = sched_setaffinity( 0, sizeof(cpu_set), &cpu_set  );
    printf(" set cpu_set: 0 1 2 3 %s\n", ret==0? "success":"failed" );


    printf(" new cpu_set: ");
    ret = sched_getaffinity( 0, sizeof(cpu_set), &cpu_set );
    for( i=0; i< num_cores; i++ ){
        if( CPU_ISSET( i, &cpu_set )) printf( "%d ", i);
    }
    printf("\n");

    CPU_ZERO( &cpu_set );

    ret = sched_setaffinity( 0, sizeof(cpu_set), &cpu_set );
    printf(" set cpu_set: none %s\n", ret==0? "success":"failed" );

    printf(" new cpu_set: ");
    ret = sched_getaffinity( 0, sizeof(cpu_set), &cpu_set );
    for( i=0; i< num_cores; i++ ){
        if( CPU_ISSET( i, &cpu_set )) printf( "%d ", i);
    }
    printf("\n");

    CPU_ZERO( &cpu_set );
    ret = sched_getaffinity( 0, sizeof(cpu_set), &cpu_set );
    for( i=0; i< num_cores; i++ ){
        if( CPU_SET( i, &cpu_set ));
    }

    ret = sched_setaffinity( 0, sizeof(cpu_set), &cpu_set );
    printf(" set cpu_set: all %s\n", ret==0? "success":"failed" );

    printf(" new cpu_set: ");
    ret = sched_getaffinity( 0, sizeof(cpu_set), &cpu_set );
    for( i=0; i< num_cores; i++ ){
        if( CPU_ISSET( i, &cpu_set )) printf( "%d ", i);
    }
    printf("\n");




}

感谢您阅读我的问题。

openMP在sched_setaffinity之后变慢

0 个答案: