Abnormal CPU utilization for stopwatch
danlkv opened this issue · 2 comments
Hello
I noticed that when I use stopwatch / clock mode, termdown loads CPU almost to 100%. This is not how a simple stopwatch should behave. Timer mode is fine.
Below see some profiling data obtained using perf tool for stopwatch, timer and bash sleep function.
Note the task-clock and CPU clock frequency. The difference is also noticable for cooler fan noise.
⟩ perf stat termdown -z
Performance counter stats for 'termdown -z':
14274.19 msec task-clock:u # 0.855 CPUs utilized
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
3777 page-faults:u # 0.265 K/sec
42938581915 cycles:u # 3.008 GHz
67333349803 instructions:u # 1.57 insn per cycle
10616874674 branches:u # 743.781 M/sec
83548787 branch-misses:u # 0.79% of all branches
16.686076853 seconds time elapsed
13.758134000 seconds user
0.515512000 seconds sys
⟩ perf stat sleep 5
Performance counter stats for 'sleep 5':
3.34 msec task-clock:u # 0.001 CPUs utilized
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
58 page-faults:u # 0.017 M/sec
344081 cycles:u # 0.103 GHz
240960 instructions:u # 0.70 insn per cycle
53836 branches:u # 16.137 M/sec
4084 branch-misses:u # 7.59% of all branches
5.009497461 seconds time elapsed
0.000000000 seconds user
0.005746000 seconds sys
⟩ perf stat termdown 5
Performance counter stats for 'termdown 5':
450.24 msec task-clock:u # 0.084 CPUs utilized
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
3775 page-faults:u # 0.008 M/sec
823480722 cycles:u # 1.829 GHz
1074623603 instructions:u # 1.30 insn per cycle
220645709 branches:u # 490.064 M/sec
5177317 branch-misses:u # 2.35% of all branches
5.358574635 seconds time elapsed
0.343564000 seconds user
0.112419000 seconds sys
I managed to fix it by putting sleep(.2) in while loop in stopwatch. I'm not sure this is the right way to mitigate the problem, though.
Here're the stats for the modified termdown installed from source:
⟩ perf stat termdown -z
Performance counter stats for 'termdown -z':
685.45 msec task-clock:u # 0.074 CPUs utilized
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
6648 page-faults:u # 0.010 M/sec
896813441 cycles:u # 1.308 GHz
1270088395 instructions:u # 1.42 insn per cycle
253694359 branches:u # 370.114 M/sec
5930235 branch-misses:u # 2.34% of all branches
9.240416002 seconds time elapsed
0.561054000 seconds user
0.136016000 seconds sys
My system:
⟩ termdown --version
1.17.0
⟩ python -V
Python 3.8.3
⟩ uname -a
Linux archlinux 5.6.15-arch1-1 #1 SMP PREEMPT Wed, 27 May 2020 23:42:26 +0000 x86_64 GNU/Linux
I think using sleep in the while loop is a reasonable solution, it seems it's used already for a similar purpose in input_thread_body. If you don't mind, I might write a pull request later today.
Apparently this only happens with -z and goes away if I just remove
if time:
sleep_target = sleep_target.replace(microsecond=0)Will need to investigate more. We probably end up sleeping for zero seconds for a number of iterations each second, depending on when within a second we started the program, hence the wildly varying amount of CPU utilization:
> time termdown -zq 10
10.010 10 total
real 0m10.143s
user 0m1.772s
sys 0m0.041s
> time termdown -zq 10
10.011 10 total
real 0m10.132s
user 0m8.829s
sys 0m0.068s
> time termdown -zq 10
10.011 10 total
real 0m10.136s
user 0m6.246s
sys 0m0.072s