WARNING: PMU in use. And the CPU utilization is inaccurate
changzhi1990 opened this issue · 7 comments
I try to use perf-collect.py to collect the metrics of "stress" process.
But I meet a warning like this:
ubuntu@inspur-icx-1:~/zhi/PerfSpect$ sudo python perf-collect.py -a "stress --vm 1 -c 1 --vm-bytes 1024M"
2023-05-10 16:57:37,423 INFO: nmi_watchdog disabled
2023-05-10 16:57:37,776 WARNING: PMU in use: cpu cycles
2023-05-10 16:57:37,815 WARNING: output could be inaccurate
2023-05-10 16:57:37,815 INFO: Collecting perf stat for events in : /home/ubuntu/zhi/PerfSpect/events/icx.txt
stress: info: [832408] dispatching hogs: 1 cpu, 0 io, 1 vm, 0 hdd
^C2023-05-10 16:58:18,853 INFO: Collection stopped! Caculating TSC frequency now
2023-05-10 16:58:19,903 INFO: nmi_watchdog re-enabled
2023-05-10 16:58:19,905 INFO: perf stat dumped to perfstat.csv
ubuntu@inspur-icx-1:~/zhi/PerfSpect$ sudo python perf-postprocess.py -r perfstat.csv --html cpu.html
2023-05-10 16:57:17,006 INFO: Generated results file(s) in: /home/ubuntu/zhi/PerfSpect
2023-05-10 16:57:17,467 INFO: static HTML file written at cpu.html
2023-05-10 16:57:17,467 INFO: Done!
If I ignore these warnings, the CPU utilization is inaccurate in the cpu.html file.
Is there any relationship between the warning info and the inaccurate problem?
Thanks!
@changzhi1990 What is the expected CPU utilization?
@changzhi1990 What is the expected CPU utilization?
The expected utilization is 100%
@changzhi1990 What is the expected CPU utilization?
The expected utilization is 100%
Because I use the stress process and I can see the CPU utilization is almost 100% by htop of the specific core.
@changzhi1990 The CPU utilization presented by PerfSpect is the average utilization across all cores/threads. The stress-ng command you are running is running only 2 threads.
@changzhi1990 The CPU utilization presented by PerfSpect is the average utilization across all cores/threads. The stress-ng command you are running is running only 2 threads.
Thanks for your reply. I try to add a new feature to collcet the perf events of the specific CPU cores, like this
But the CPU utilization is almost 0% in the html file. So I think maybe I need to modify the code further.
there's two things you could do here.
- Run PerfSpect in thread mode
sudo ./perf-collect --thread. This cannot generate an HTML file yet - Run PerfSpect in PID mode
sudo ./perf-collect --pid <your-stress-ng-process-id>. I also don't think this can generate an HTML file
since these aren't supported in the HTML report yet you'll have to look in the csv for metric_CPU utilization %
closing because PerfSpect is functioning as intended. It shows system-wide cpu utilization which is very low in this case (given that only one thread is occupied at 100%)