CN102855171A - Consumer price index (CPI) real-time monitoring method based on linux system - Google Patents

Abstract

The invention provides a consumer price index (CPI) real-time monitoring method based on a linux system. The method comprises the following steps of: calling a sys-perf-event-open system provided by utilizing linux, reading the performance management unit (PMU) contents of a central processing unit (CPU), acquiring the real-time CPU period and execution instruction number, and calculating the real-time CPI of the system. According to the method, the defects of the previous system performance monitoring system in micro-architecture monitoring aspects are overcome, each CPU in the system can be accurately monitored in real time, and the CPI data is calculated. The method can be used for timely and accurately judging the real performance exerted by the system and can reflect the advantages of an application program from the side in the field of high-performance servers.

Description

A kind of method of the CPI Real-Time Monitoring based on the linux system

Technical field

The present invention relates to the CPU framework, C language, and the technical field such as linux programming, specifically a kind of method of the CPI Real-Time Monitoring based on the linux system.

Background technology

Along with the development that high-performance server is used, increasing people begin to note that the performance of micro-architecture aspect is on using the impact of performance.Wherein CPI is an important indicator of micro-architecture aspect, but nearly all monitoring tool all can't this content of Real-Time Monitoring at present.

Summary of the invention

The sys_perf_event_open system call that the objective of the invention is to utilize the linux kernel to provide realizes the CPI method of real-time based on the linux system.

The objective of the invention is to realize in the following manner, based on the linux platform, cpu cycle by PMU among every CPU in the reading system and carry out the related content such as instruction number, real-time calculate entire system CPI value and export, performing step is as follows:

1) packing sys_perf_event_open system call:

static?int?perf_event_open(struct?perf_event_attr?*hw_event_uptr,

pid_t?pid,?int?cpu,?int?group_fd,?unsigned?long?flags)?{

return?syscall(__NR_perf_event_open,?hw_event_uptr,?pid,?cpu,?group_fd,flags);

}

2) statistical system CPU number:

int?cpu_num?=?sysconf(_SC_NPROCESSORS_ONLN)；

3) generate the perf_event array:

static?int?event_num=2;

static?int?event[2]={\

PERF_COUNT_HW_CPU_CYCLES,

PERF_COUNT_HW_INSTRUCTIONS\

};

4) generate a group for each CPU:

for(i=0;i<cpu_num;i++){

_perf_event_open(i);

ioctl(pperf_fd[i][0],?PERF_EVENT_IOC_ENABLE,?0);

}

Wherein _ and the perf_event_open function of perf_event_open function for packing, be specially:

static?void?_perf_event_open(int?cpuid)?{

int?i;

struct?perf_event_attr?pe;

memset(&pe,0,sizeof(struct?perf_event_attr));

pe.type?=?PERF_TYPE_HARDWARE;

pe.size?=?sizeof(struct?perf_event_attr);

pe.config?=?event[0];

pe.exclude_kernel=1;

pe.exclude_hv=1;

pe.read_format=PERF_FORMAT_GROUP|PERF_FORMAT_ID;

pe.disabled?=?1;

pperf_fd[cpuid][0]?=?perf_event_open(&pe,?-1,?cpuid,?-1,?0);

if?(pperf_fd[cpuid][0]?>=?0)

for(i=1;i<event_num;i++)

{

pe.config?=?event[i];

pe.disabled?=?0;

pperf_fd[cpuid][i]=?perf_event_open(&pe,?-1,?cpuid,pperf_fd[cpuid][0],?0);

}

return;

}

5) read PMU numerical value and export whole CPI:

for(i=0;i<sys_info.cpu_num;i++){

read(pperf_fd[i][0],?&rf,?sizeof(rf));

pperf_clean(i);

t_c+=?rf.cr[0].value;

t_i+=?rf.cr[1].value;

printf("%f?",(float)rf.cr[0].value/(float)rf.cr[1].value);

}

printf("%f\n",t_c/t_i);

Wherein pperf_clean is the PMU counter that empties corresponding CPU.Specific as follows:

static?inline?void?pperf_clean(cpuid)

{

int?i;

for?(i=0;i<event_num;i++)

{

ioctl(pperf_fd[cpuid][i],?PERF_EVENT_IOC_RESET,?0);

}

}

6) empty the value of t_c and t_i, wait for one second, and continue to read its value.

Major part to this native system is finished substantially.

The invention has the beneficial effects as follows: the content that traditional system performance monitoring system can not monitoring system micro-architecture aspect, but along with the adding that has comprised sys_perf_event_open function in PMU and the linux kernel in the CPU makes it become possibility.The present invention has overcome the in the past deficiency of system performance monitoring system aspect the micro-architecture monitoring, can accomplish in real time accurately each CPU in the supervisory system, and calculate its CPI data.In the high-performance server field, this method can be used for the actual performance of judgement system performance promptly and accurately and the quality that can reflect from the side a application program.

Description of drawings

Fig. 1 is system flowchart.

Embodiment

Explain below with reference to Figure of description method of the present invention being done.

Based on the linux platform, cpu cycle by PMU among every CPU in the reading system and carry out the related content such as instruction number, real-time calculate entire system CPI value and export, performing step is as follows:

1) packing sys_perf_event_open system call:

static?int?perf_event_open(struct?perf_event_attr?*hw_event_uptr,

pid_t?pid,?int?cpu,?int?group_fd,?unsigned?long?flags)?{

return?syscall(__NR_perf_event_open,?hw_event_uptr,?pid,?cpu,?group_fd,flags);

}

2) statistical system CPU number:

int?cpu_num?=?sysconf(_SC_NPROCESSORS_ONLN)；

3) generate the perf_event array:

static?int?event_num=2;

static?int?event[2]={\

PERF_COUNT_HW_CPU_CYCLES,

PERF_COUNT_HW_INSTRUCTIONS\

};

4) generate a group for each CPU:

for(i=0;i<cpu_num;i++){

_perf_event_open(i);

ioctl(pperf_fd[i][0],?PERF_EVENT_IOC_ENABLE,?0);

}

Wherein _ and the perf_event_open function of perf_event_open function for packing, be specially:

static?void?_perf_event_open(int?cpuid)?{

int?i;

struct?perf_event_attr?pe;

memset(&pe,0,sizeof(struct?perf_event_attr));

pe.type?=?PERF_TYPE_HARDWARE;

pe.size?=?sizeof(struct?perf_event_attr);

pe.config?=?event[0];

pe.exclude_kernel=1;

pe.exclude_hv=1;

pe.read_format=PERF_FORMAT_GROUP|PERF_FORMAT_ID;

pe.disabled?=?1;

pperf_fd[cpuid][0]?=?perf_event_open(&pe,?-1,?cpuid,?-1,?0);

if?(pperf_fd[cpuid][0]?>=?0)

for(i=1;i<event_num;i++)

{

pe.config?=?event[i];

pe.disabled?=?0;

pperf_fd[cpuid][i]=?perf_event_open(&pe,?-1,?cpuid,pperf_fd[cpuid][0],?0);

}

return;

}

5) read PMU numerical value and export whole CPI:

for(i=0;i<sys_info.cpu_num;i++){

read(pperf_fd[i][0],?&rf,?sizeof(rf));

pperf_clean(i);

t_c+=?rf.cr[0].value;

t_i+=?rf.cr[1].value;

printf("%f?",(float)rf.cr[0].value/(float)rf.cr[1].value);

}

printf("%f\n",t_c/t_i);

Wherein pperf_clean is the PMU counter that empties corresponding CPU.Specific as follows:

static?inline?void?pperf_clean(cpuid)

{

int?i;

for?(i=0;i<event_num;i++)

{

ioctl(pperf_fd[cpuid][i],?PERF_EVENT_IOC_RESET,?0);

}

}

6) empty the value of t_c and t_i, wait for one second, and continue to read its value.

Except the described technical characterictic of instructions, be the known technology of those skilled in the art.

Claims (1)

1. method based on the CPI Real-Time Monitoring of linux system is characterized in that the cpu cycle by PMU among every CPU in the reading system and carries out the related content such as instruction number, real-time calculate entire system CPI value and exports, and the specific implementation step is as follows:

Packing sys_perf_event_open system call:

static?int?perf_event_open(struct?perf_event_attr?*hw_event_uptr,

pid_t?pid,?int?cpu,?int?group_fd,?unsigned?long?flags)?{

return?syscall(__NR_perf_event_open,?hw_event_uptr,?pid,?cpu,?group_fd,flags);

}

Statistical system CPU number:

int?cpu_num?=?sysconf(_SC_NPROCESSORS_ONLN)；

Generate the perf_event array:

static?int?event_num=2;

static?int?event[2]={\

PERF_COUNT_HW_CPU_CYCLES,

PERF_COUNT_HW_INSTRUCTIONS\

};

4) generate a group for each CPU:

for(i=0;i<cpu_num;i++){

_perf_event_open(i);

ioctl(pperf_fd[i][0],?PERF_EVENT_IOC_ENABLE,?0);

}

Wherein _ and the perf_event_open function of perf_event_open function for packing, be specially:

static?void?_perf_event_open(int?cpuid)?{

int?i;

struct?perf_event_attr?pe;

memset(&pe,0,sizeof(struct?perf_event_attr));

pe.type?=?PERF_TYPE_HARDWARE;

pe.size?=?sizeof(struct?perf_event_attr);

pe.config?=?event[0];

pe.exclude_kernel=1;

pe.exclude_hv=1;

pe.read_format=PERF_FORMAT_GROUP|PERF_FORMAT_ID;

pe.disabled?=?1;

pperf_fd[cpuid][0]?=?perf_event_open(&pe,?-1,?cpuid,?-1,?0);

if?(pperf_fd[cpuid][0]?>=?0)

for(i=1;i<event_num;i++)

{

pe.config?=?event[i];

pe.disabled?=?0;

pperf_fd[cpuid][i]=?perf_event_open(&pe,?-1,?cpuid,pperf_fd[cpuid][0],?0);

}

return;

}

Read PMU numerical value and export whole CPI:

for(i=0;i<sys_info.cpu_num;i++){

read(pperf_fd[i][0],?&rf,?sizeof(rf));

pperf_clean(i);

t_c+=?rf.cr[0].value;

t_i+=?rf.cr[1].value;

printf("%f?",(float)rf.cr[0].value/(float)rf.cr[1].value);

}

printf("%f\n",t_c/t_i);

Wherein pperf_clean is the PMU counter that empties corresponding CPU;

Specific as follows:

static?inline?void?pperf_clean(cpuid)

{

int?i;

for?(i=0;i<event_num;i++)

{

ioctl(pperf_fd[cpuid][i],?PERF_EVENT_IOC_RESET,?0);

}

}

6) empty the value of t_c and t_i, wait for one second, and continue to read its value;

Major part to this native system is finished substantially.

Images