Rudy/psutilcpu

[bunelr]

Use Psutil to fix #653

Keep on using the wmi connection to get the system.cpu.interrupted metric, which is not available via psutil. I don't know if anyone is really using that metric, given that it's not part of the standard dashboard but I left it there to preserve existing behaviour.

Reason for the >100 peak:

• First, wmi doesn't return values that add to 100 so there was always this offset.
• Second, the if cpu_user to check that a value was there resulted in metrics not being sent when they were ==0, which lead to missing values on the graph and aggregation artifacts showing huge peak

[remh]

Can you take the opportunity to reorganize the import the same way i did ? 737f396

Thanks!

[remh]

Thanks!
Should we use psutil.cpu_times instead of psutil.cpu_times_percent, then we could compute a rate to get an averaged CPU consumption over two collections instead of the instantaneous consumption ?

[bunelr]

Do you mean to compute an average over the timespan between each run of each checks? That would probably be better indeed.

[remh]

Actually it won't work easily because the old check interface can't compute negative rates so forget about it.

Small other remark:
with psutil you are using a sampling interval of 0.5 sec, do you know what was the equivalent interval when using WMI ?

[bunelr]

Why would there be a negative rate? The cpu_times would be strictly growing and we would need to do a computation to get percents anyway, in order to provide the same metric.

Regarding the interval, the data is sampled by windows but the interval is not exposed.

[remh]

Ah yeah you're right.
So yeah you can define these metrics as counter https://github.com/DataDog/dd-agent/blob/master/checks/system/win32.py#L88-L91 instead of gauge and then use psutil.cpu_times.
Rate will be automatically computed.