使用 OProfile for Linux on POWER 识别性能瓶颈

  
int fast_multiply(x,  y) 
{
        return x * y;
}
int slow_multiply(x, y) 
{
        int i, j, z;
        for (i = 0, z = 0; i < x; i++) 
                z = z + y;
        return z;
}
int main()
{
        int i,j;
        int x,y;
        for (i = 0; i < 200; i ++) {
                for (j = 0; j " 30 ; j++) {
                        x = fast_multiply(i, j);
                        y = slow_multiply(i, j);
                }
        }
        return 0;
}

 gcc  -g multiply.c -o multiply   

# opcontrol --vmlinux=/boot/vmlinux-2.6.5-7.139-pseries64
# opcontrol --reset
# opcontrol --setup --event=CYCLES:1000
# opcontrol --start
Using 2.6+ OProfile kernel interface.
Reading module info.
Using log file /var/lib/oprofile/oprofiled.log
Daemon started.
Profiler running.
# ./multiply
# opcontrol --dump
# opcontrol --stop
Stopping profiling.
# opcontrol -h
Stopping profiling.
Killing daemon.

# opannotate --source ./multiply
/* 
 * Command line: opannotate --source ./multiply 
 * 
 * Interpretation of command line:
 * Output annotated source file with samples
 * Output all files
 * 
 * CPU: ppc64 POWER5, speed 1656.38 MHz (estimated)
 * Counted CYCLES events (Processor cycles) with a unit mask of
0x00 (No unit mask) count 1000
 */
/* 
 * Total samples for file : "/usr/local/src/badcode/multiply.c"
 * 
 *   6244 100.000
 */
               :int fast_multiply(x, y) 
    36  0.5766 :{ /* fast_multiply total:     79  1.2652 */
    26  0.4164 :        return x * y;
    17  0.2723 :}
               :
               :int slow_multiply(x, y) 
    50  0.8008 :{ /* slow_multiply total:   6065 97.1332 */
               :        int i, j, z;
  2305 36.9154 :        for (i = 0, z = 0; i " x; i++) 
  3684 59.0006 :                z = z + y;
    11  0.1762 :        return z;
    15  0.2402 :}
               :
               :int main()
               :{ /* main total:    100  1.6015 */
               :        int i,j;
               :        int x,y;
               :
     1  0.0160 :        for (i = 0; i " 200; i ++) {
     6  0.0961 :                for (j = 0; j " 30 ; j++) {
    75  1.2012 :                        x = fast_multiply(i, j);
    18  0.2883 :                        y = slow_multiply(i, j);
               :                }
               :        }
               :        return 0;
               :}
               

36  0.5766 :{ /* fast_multiply total:     79  1.2652 */

50  0.8008 :{ /* slow_multiply total:   6065 97.1332 */

struct shared_data_struct {
   unsigned int data1;
   unsigned int data1;
}

#include <stdlib.h>
#include <sched.h>
struct shared_data_struct {
        unsigned int data1;
        unsigned int data2;
};
struct shared_data_struct shared_data;
static int inc_second(struct shared_data_struct *);
int main(){
        int i, j, pid;
        void *child_stack;
        /* allocate memory for other process to execute in */
        if((child_stack = (void *) malloc(4096)) == NULL) {
                perror("Cannot allocate stack for child");
                exit(1);
        }
        /* clone process and run in the same memory space */
        if ((pid = clone((void *)&inc_second, child_stack,
 	            CLONE_VM, &shared_data)) < 0) {
                perror("clone called failed.");
                exit(1);
        }
        /* increment first member of shared struct */
        for (j = 0; j < 2000; j++) {
                for (i = 0; i < 100000; i++) {
                        shared_data.data1++;
                }
        }
        return 0;
}
int inc_second(struct shared_data_struct *sd)
{
        int i,j;
        /* increment second member of shared struct */
        for (j = 1; j < 2000; j++) {
                for (i = 1; i < 100000; i++) {
                        sd->data2++;
                }
        }
}

gcc -o cache-miss cache-miss.c

# opcontrol --vmlinux=/boot/vmlinux-2.6.5-7.139-pseries64
# opcontrol --reset
# opcontrol --setup –event=PM_LSU_LMQ_LHR_MERGE_GP9:1000
# opcontrol --start
Using 2.6+ OProfile kernel interface.
Reading module info.
Using log file /var/lib/oprofile/oprofiled.log
Daemon started.
Profiler running.
# ./cache-miss
# opcontrol --dump
# opcontrol -h
Stopping profiling.
Killing daemon.
# opreport -l ./cache-miss 
CPU: ppc64 POWER5, speed 1656.38 MHz (estimated)
Counted PM_LSU_LMQ_LHR_MERGE_GP9 events (Dcache miss occurred for
the same real cache line as earlier req, merged into LMQ) with a
unit mask of 0x00 (No unit mask) count 1000
samples  %        symbol name
47897    58.7470  main
33634    41.2530  inc_second

struct shared_data_struct {
   unsigned int data1;
   char pad[124];
   unsigned int data1;

# opcontrol --vmlinux=/boot/vmlinux-2.6.5-7.139-pseries64
# opcontrol --reset
# opcontrol --setup –event=PM_LSU_LMQ_LHR_MERGE_GP9:1000
# opcontrol --start
Using 2.6+ OProfile kernel interface.
Reading module info.
Using log file /var/lib/oprofile/oprofiled.log
Daemon started.
Profiler running.
# ./cache-miss
# opcontrol --dump
# opcontrol -h
Stopping profiling.
Killing daemon.
# opreport -l ./cache-miss 
error: no sample files found: profile specification too strict ?

# opcontrol --vmlinux=/boot/vmlinux-2.6.5-7.139-pseries64
# opcontrol --reset
# opcontrol --setup –event=CYCLES:1000
# opcontrol --start
Using 2.6+ OProfile kernel interface.
Reading module info.
Using log file /var/lib/oprofile/oprofiled.log
Daemon started.
Profiler running.
# ./cache-miss
# opcontrol --dump
# opcontrol -h
Stopping profiling.
Killing daemon.
# opreport -l ./cache-miss 
CPU: ppc64 POWER5, speed 1656.38 MHz (estimated)
Counted CYCLES events (Processor cycles) with a unit mask of 0x00
(No unit mask) count 1000
samples  %        symbol name
121166   53.3853  inc_second
105799   46.6147  main
 

# opcontrol --vmlinux=/boot/vmlinux-2.6.5-7.139-pseries64
# opcontrol --reset
# opcontrol --setup –event=CYCLES:1000
# opcontrol --start
Using 2.6+ OProfile kernel interface.
Reading module info.
Using log file /var/lib/oprofile/oprofiled.log
Daemon started.
Profiler running.
# ./cache-miss
# opcontrol --dump
# opcontrol -h
Stopping profiling.
Killing daemon.
# opreport -l ./cache-miss 
CPU: ppc64 POWER5, speed 1656.38 MHz (estimated)
Counted CYCLES events (Processor cycles) with a unit mask of 0x00
(No unit mask) count 1000
samples  %        symbol name
104916   58.3872  inc_second
74774    41.6128  main

#include <stdlib.h>
#include <sched.h>
struct shared_data_struct {
        unsigned int data1;
        unsigned int data2;
};
struct shared_data_struct shared_data;
static int inc_second(struct shared_data_struct *);
int main(){
        int i, j, pid;
        cpu_set_t cmask;
        unsigned long len = sizeof(cmask);
        pid_t p = 0;
        void *child_stack;
        __CPU_ZERO(&cmask);
        __CPU_SET(0, &cmask);
        /* allocate memory for other process to execute in */
        if((child_stack = (void *) malloc(4096)) == NULL) {
                perror("Cannot allocate stack for child");
                exit(1);
        }
        /* clone process and run in the same memory space */
        if ((pid = clone((void *)&inc_second, child_stack,
CLONE_VM, &shared_data)) < 0) {
                perror("clone called failed");
                exit(1);
        }
        if (!sched_setaffinity(0, len, &cmask)) {
                printf("Could not set cpu affinity for current
process.\n");
                exit(1);
        }
        if (!sched_setaffinity(pid, len, &cmask)) {
                printf("Could not set cpu affinity for cloned
process.\n");
                exit(1);
        }
        /* increment first member of shared struct */
        for (j = 0; j < 2000; j++) {
                for (i = 0; i < 100000; i++) {
                        shared_data.data1++;
                }
        }
        return 0;
}
int inc_second(struct shared_data_struct *sd)
{
        int i,j;
        /* increment second member of shared struct */
        for (j = 1; j < 2000; j++) {
                for (i = 1; i < 100000; i++) {
                        sd->data2++;
                }
        }
}

John Engel 在 IBM 工作，他是 IBM eServer Solutions Enablement 组织的一名 Linux 技术顾问。他居住在 Rochester, MN。他的主要职责是帮助解决方案开发人员将其应用程序迁移到 Linux on POWER。除了在 IBM 工作以外，他还在进行多方面的 Linux 软件开发。