我希望通过sched_affinity在多个CPU上设置关联,如下所示。
void
pin(pid_t t, int cpu)
{
cpu_set_t cpuset;
CPU_ZERO(&cpuset);
CPU_SET(cpu, &cpuset);
sched_setaffinity(t, sizeof(cpu_set_t), &cpuset);
}我的环境是32个内核,其中有4个CPU,单个CPU有8个核心。
我希望线程0~7运行在同一个cpu上,线程8~ 15运行在同一个cpu上,依此类推。
我想知道如何在CPU_SET中设置变量cpu。
这被设置为线程id,如果核心号码被天真地分配,即cpu0有第0核,第1核心和第2核心,.,cpu1有第8核心,第9核心,.。
一方面,cpu被设置为循环规则,如果核心数被分配为循环规则,即cpu0具有第0核、第4核、第8核、.和cpu1有第1核、第5核、.
应该设置可变cpu、朴素规则还是循环规则?
发布于 2016-05-31 12:43:00
在Linux (和其他操作系统)下,程序员可以设置CPU关联,即内核可以将此进程调度到的允许CPU。在叉()上,进程继承父CPU亲缘关系。这是非常有用的,如果一个人想限制CPU出于任何原因。
例如,有人可能会限制
一般来说,将进程/线程限制在某些核心或套接字上,以避免它们被操作系统调度--最大限度地利用L1/L2高速缓存(固定在核心上时)或L3/LLC高速缓存(固定到套接字时)的好处。
关于您关于“线程分发”的问题:处理器开发引入了对称多线程(SMT)或超线程(被Intel称为),它引入了每个物理核心的2个逻辑核(例如Intel Xeon),甚至4个逻辑核(例如Intel Knights、IBM Power)。这些逻辑核与上面cpuset中的"CPU“一样被很好地表示。此外,一些处理器强制使用NUMA域,其中从一个核到它自己的内存的访问速度较快,而在另一个NUMA域中访问另一个核心内存的速度较慢.
所以,正如上面的一些评论所暗示的:这取决于!线程之间是否相互通信(通过共享内存),那么它们应该保持在同一个缓存中。您的线程是否使用相同的功能单元(例如FPU),然后在相同的物理核(有两个逻辑核,即超线程)上调度两个可能会损害性能。
若要四处游玩,请查收以下代码:
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <unistd.h>
#include <sys/param.h>
#include <sys/sysctl.h>
#include <pthread.h>
// The following is Linux-specific
#include <syscall.h> // For syscall to gettid()
#include <sched.h> // sched_[gs]etaffinity require _GNU_SOURCE
#define ERROR(t, e) do { \
const int __error = (e); \
fprintf (stderr, "ERROR: %s error:%d [%s] errno:%d [%s]\n", \
(t), __error, strerror(__error), errno, strerror(errno)); \
exit(__error); \
} while(0)
#ifndef MAX
#define MAX(a,b) ((a) > (b) ? (a) : (b))
#endif
#ifndef MIN
#define MIN(a,b) ((a) < (b) ? (a) : (b))
#endif
/* Local function definitions */
void print_schedaffinity(const char * text, const cpu_set_t cpuset, const int max_cpus);
void * thread_func(void * arg);
/* Local type definitions */
struct thread_data {
pthread_t thread;
int max_cpu;
int thread_num;
void * thread_work;
};
/* The highest value for CPU to be specified in cpuset in a call to
* sched_setaffinity -- otherwise, we get returned -1 and errno==EINVAL
*/
static int max_cpu_available = 0;
/* Local function declarations */
void print_schedaffinity(const char * text, const cpu_set_t cpuset, const int max_cpus) {
const int max = MIN(8*sizeof(cpu_set_t), max_cpus);
int i;
printf("PRINT CPU AFFINITY %s:\n", text);
printf("cpus:\t");
for (i = 0; i < max; i++) {
printf (" %3d", i);
if (i % 8 == 7)
printf(" | ");
}
printf("\nmask:\t");
for (i = 0; i < max; i++) {
if (CPU_ISSET(i, &cpuset))
printf (" X");
else
printf (" ");
if (i % 8 == 7)
printf(" | ");
}
printf("\n");
}
void * thread_func(void * arg) {
struct thread_data * thread_data = (struct thread_data *)arg;
const size_t sizeof_cpuset = sizeof(cpu_set_t);
char print_buffer[64];
cpu_set_t cpuset;
long tid;
int rc;
CPU_ZERO(&cpuset);
CPU_SET(thread_data->thread_num % max_cpu_available, &cpuset);
/* We set the affinity of the CALLING thread, aka 0 */
tid = syscall(SYS_gettid);
printf("PID:%ld tid:%ld thread_num:%d\n",
getpid(), tid, thread_data->thread_num);
rc = sched_setaffinity(0, sizeof_cpuset, &cpuset);
if (0 != rc)
ERROR("sched_setaffinity", rc);
/* Dooo SCHTUF now */
/* Somewhat sort the output... */
sleep (thread_data->thread_num);
snprintf (print_buffer, sizeof(print_buffer),
"in thread %d after sched_setaffinity", thread_data->thread_num);
print_schedaffinity(print_buffer, cpuset, 8);
return NULL;
}
int main (int argc, char * argv[])
{
const int NUM = 8;
const pid_t pid = getpid();
const size_t size_cpu_set = sizeof(cpu_set_t);
cpu_set_t cpuset;
int rc;
int i;
/* Get, and print the original CPU affinity setting (scheduling is not limited, i.e. all cores may run this PID) */
CPU_ZERO (&cpuset);
rc = sched_getaffinity(pid, size_cpu_set, &cpuset);
if (0 != rc)
ERROR("sched_getaffinity", rc);
print_schedaffinity("in main", cpuset, 8);
/* Search for the last / highest cpu being set -- claim, that this is the max cpu to be set, cough */
for (i = 0; i < 8 * size_cpu_set; i++) {
if (!CPU_ISSET(i, &cpuset)) {
max_cpu_available = i;
break;
}
}
/* Limit the process to the first core, only */
CPU_ZERO (&cpuset);
CPU_SET (0, &cpuset);
rc = sched_setaffinity (pid, size_cpu_set, &cpuset);
if (0 != rc)
ERROR("sched_setaffinity", rc);
print_schedaffinity("in main after sched_setaffinity", cpuset, 8);
/* Let's start NUM threads and have them limit their scheduling */
sleep(1);
struct thread_data * thread_data = (struct thread_data*)malloc(sizeof(struct thread_data) * NUM);
for (i = 0; i < NUM; i++) {
thread_data[i].thread_num = i;
pthread_create (&thread_data[i].thread, NULL, thread_func, &thread_data[i]);
}
/* And wait for them to finish... */
for (i = 0; i < NUM; i++) {
pthread_join (thread_data[i].thread, NULL);
}
return 0;
}编辑:应该澄清苹果,因为OSX10.5(豹)提供了亲和力,如在https://developer.apple.com/library/mac/releasenotes/Performance/RN-AffinityAPI/
https://stackoverflow.com/questions/37523936
复制相似问题