在 Linux(和其他操作系统)下,程序员可以设置 CPU 亲和性,即内核可以将此进程调度到的允许 CPU。在 fork() 之后,进程继承父 CPU 亲和性。如果有人出于某种原因想要限制 CPU,这会非常方便。
例如,可能会限制
- 某些用户的进程只分配给一个 CPU,而将其余的 CPUS 留给其他用户(参见 man 7 cpuset)。
- CPU 的进程“更接近”某物,例如限制与套接字上的内核进行通信的进程,该套接字直接连接到网卡(NIC 或 HCA)。
一般来说,将进程/线程限制在某些内核或套接字上可能是有益的,以免它们被操作系统调度出去——最大限度地发挥 L1/L2 缓存(固定到内核时)或 L3 的好处/LLC 缓存(固定到套接字时)。
关于您关于“线程分布”的问题:处理器开发引入了对称多线程 (SMT) 或超线程(英特尔称为),它引入了 2 个逻辑内核(例如 Intel Xeon)甚至 4 个逻辑内核(例如 Intel Knights Landing、IBM Power ) 每个物理核心。这些逻辑核心在上面的 cpuset 中也被表示为“CPU”。此外,一些处理器强加了 NUMA 域,从一个内核到它“自己的”内存的内存访问速度很快,而在另一个 NUMA 域中访问另一个内核内存的速度较慢......
所以,正如上面的一些评论所暗示的:这取决于!您的线程是否相互通信(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),然后在同一个物理内核(具有 2 个逻辑内核,即超线程)上调度两个可能会损害性能。
要玩转,请找到以下代码:
#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Apple,因为OSX 10.5(Leopard)提供Affinity,如https://developer.apple.com/library/mac/releasenotes/Performance/RN-AffinityAPI/