Performance
perf
Métricas de performance.
Las métricas se construyen en base a events:
ls /sys/bus/event_source/devices/cpu/events/
tambien podemos ver las métricas que usa perf con:
perf list --details
Performance counter stats for './austal short':
7.155,69 msec task-clock # 1,000 CPUs utilized
118 context-switches # 16,490 /sec
3 cpu-migrations # 0,419 /sec
7.414 page-faults # 1,036 K/sec
29.695.155.163 cycles # 4,150 GHz
47.379.828.377 instructions # 1,60 insn per cycle
6.172.097.825 branches # 862,545 M/sec
43.727.245 branch-misses # 0,71% of all branches
7,157403331 seconds time elapsed
7,090932000 seconds user
0,064990000 seconds sys
- task-clock : meassures how “parallel” the job has been (how many cpus were used).
- context-switches : how the scheduler treated the application. How interrupted has been. Lower is better.
- cpu-migrations : how much the shceduler moved the application between cores. Lower is better.
- page-faults : the number of times the program tried to access memory that wasn’t in physical RAM, requiring the system to fetch it. Higher page faults can slow down performance.
- cycles : number of CPU cycles spent in executing the program.
- instructions : number of instructions executed. The number of instructions per cycle is a good indicator of how efficient the program is.
- branches : tell you how many jumps and loops are performed in your code. Correctly predicted branches should not hurt your performance,
- branch-misses : branch-misses on the other hand hurt your performance very badly and lead to stall cycles.
gprof
Licensed MIT.