目前,我通过Python的psutil
模块监控多个进程,并以百分比形式检索CPU使用率,该百分比基于execution_time/total_time
。这样做的问题是动态电压和频率调节(DVFS,或ACPI的P状态,或cpufreq等)。当前CPU频率越低,进程需要执行的时间越长,CPU使用率就越高。与此相反,我需要相对于CPU的最大性能 fixed CPU使用率。
为避免永久性地改变“当前频率”的多次重新安装,一种方法是直接使用该过程使用的CPU周期。原则上,这可以通过C中的perf_event.h
或Linux命令行上的perf
来完成。不幸的是,我找不到一个提供类似功能的Python模块(基于前面提到的)。
答案 0 :(得分:2)
感谢BlackJack的评论
如何在C中将其作为共享库实现并通过Python中的
ctypes
使用它?库调用引入的开销更少。子进程调用启动整个外部进程,并在每次需要该值时将结果作为字符串传递给管道。共享库将一次加载到当前进程中,结果将在内存中传递。
我将其实现为共享库。库 cpucycles.c 的源代码(严重基于perf_event_open
's man page的示例):
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <sys/ioctl.h>
#include <linux/perf_event.h>
#include <asm/unistd.h>
static long
perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
int cpu, int group_fd, unsigned long flags)
{
int ret;
ret = syscall(__NR_perf_event_open, hw_event, pid, cpu,
group_fd, flags);
return ret;
}
long long
cpu_cycles(unsigned int microseconds,
pid_t pid,
int cpu,
int exclude_user,
int exclude_kernel,
int exclude_hv,
int exclude_idle)
{
struct perf_event_attr pe;
long long count;
int fd;
memset(&pe, 0, sizeof(struct perf_event_attr));
pe.type = PERF_TYPE_HARDWARE;
pe.size = sizeof(struct perf_event_attr);
pe.config = PERF_COUNT_HW_CPU_CYCLES;
pe.disabled = 1;
pe.exclude_user = exclude_user;
pe.exclude_kernel = exclude_kernel;
pe.exclude_hv = exclude_hv;
pe.exclude_idle = exclude_idle;
fd = perf_event_open(&pe, pid, cpu, -1, 0);
if (fd == -1) {
return -1;
}
ioctl(fd, PERF_EVENT_IOC_RESET, 0);
ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
usleep(microseconds);
ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
read(fd, &count, sizeof(long long));
close(fd);
return count;
}
此代码通过以下两个命令编译到共享库中:
$ gcc -c -fPIC cpucycles.c -o cpucycles.o
$ gcc -shared -Wl,-soname,libcpucycles.so.1 -o libcpucycles.so.1.0.1 cpucycles.o
最后,Python可以在 cpucycles.py 中使用该库:
import ctypes
import os
cdll = ctypes.cdll.LoadLibrary(os.path.join(os.path.dirname(__file__), "libcpucycles.so.1.0.1"))
cdll.cpu_cycles.argtypes = (ctypes.c_uint, ctypes.c_int, ctypes.c_int,
ctypes.c_int, ctypes.c_int, ctypes.c_int,
ctypes.c_int)
cdll.cpu_cycles.restype = ctypes.c_longlong
def cpu_cycles(duration=1.0, pid=0, cpu=-1,
exclude_user=False, exclude_kernel=False,
exclude_hv=True, exclude_idle=True):
"""
See man page of perf_event_open for all the parameters.
:param duration: duration of counting cpu_cycles [seconds]
:type duration: float
:returns: cpu-cycle count of pid
:rtype: int
"""
count = cdll.cpu_cycles(int(duration*1000000), pid, cpu,
exclude_user, exclude_kernel,
exclude_hv, exclude_idle)
if count < 0:
raise OSError("cpu_cycles(pid={}, duration={}) from {} exited with code {}.".format(
pid, duration, cdll._name, count))
return count
答案 1 :(得分:0)
最后,我通过perf
命令行工具读取CPU周期并将其包装成Python(简化代码):
import subprocess
maximum_cpu_frequency = 3e9
cpu_percent = []
while True: # some stop criteria
try:
cpu_percent.append(int(
subprocess.check_output(["perf", "stat", "-e", "cycles",
"-p", pid, "-x", ",", "sleep", "1"],
stderr=subprocess.STDOUT).decode().split(",")[0]
)/maximum_cpu_frequency)
except ValueError:
cpu_percent.append(0.0)
不幸的是,这不是准确的,因为不精确的sleep
命令以及由于为每个样本生成新的perf
进程而有效。