随笔-纵向优化-测量最大IPC

目录

• 固定cpu运行频率
• max_ipc_test.sh
• LSD (Loop Stream Detector)

固定cpu运行频率

我的测试环境cpu频率管理是intel_pstate:

$ lscpu | grep -i hz
Model name:                           Intel(R) Core(TM) i5-10500 CPU @ 3.10GHz
CPU max MHz:                          4500.0000
CPU min MHz:                          800.0000

$ echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
1
$ echo 100 | sudo tee /sys/devices/system/cpu/intel_pstate/min_perf_pct
100
$ echo 100 | sudo tee /sys/devices/system/cpu/intel_pstate/max_perf_pct

观察设置结果：
$ sudo cpupower -c 0 frequency-info                                                                       
...
  current policy: frequency should be within 3.10 GHz and 3.10 GHz.
...

$ grep "cpu MHz" /proc/cpuinfo
...
cpu MHz         : 3100.000
cpu MHz         : 3100.665

max_ipc_test.sh

#!/bin/bash

[[ -z "$1" ]] && exit

echo 'void main() { do {__asm__ (' > nop.c
for ((i = 1; i <= $1; i++)); do
echo '"nop\n\t"' >> nop.c;
done
echo ');} while(1);}' >> nop.c

gcc -O0 nop.c -o nop

[ -f ./nop ] && sudo perf stat -C 0 --timeout 1000 taskset -c 0 ./nop

$ bash max_ipc_test.sh 4

...                
    12,290,206,450      instructions                     #    3.97  insn per cycle            
...

void main() { do {__asm__ (
"nop\n\t"
"nop\n\t"
"nop\n\t"
"nop\n\t"
);} while(1);}

LSD (Loop Stream Detector)

nop个数增加会怎么样，考虑LSD，如果循环体太长，则检测不到循环，取指速度若达不到指令执行速度，back end需要等待，出现气泡，IPC会下降

测试最大为32:

$ bash max_ipc_test.sh 33                                                                                                                                           
...               
    11,486,619,532      instructions                     #    3.71  insn per cycle            
...

看下对应的边际变化

sudo python3 toplev.py -l 3 --core S0-C0 --no-desc --verbose bash -c 'timeout 1 taskset -c 0 /data/products/hpc001/nop 33'