Cpu usage is incorrectly computed

[stefanboca]

Cpu usage as calculated here is incorrect. /proc/stat counts cpu time cumulatively from boot, so what is computed and exposed as %cpu-usage is actually something like the average cpu utilization since boot, rather than the current usage percentage.

Also, /proc/stat reports overall cpu usage times on the cpu line. It's probably (mathematically) more accurate to use that line to compute an overall cpu usage percent, rather than averaging the usage of the individual cores.

[RGBCube]

Hmm, I guess since a "current" usage would still be "usage average in the last ", we should do { cpu-usage-since = "5m" } or something similar. Do you know of a path that lets us do this nicely? And how granular can this even get?

[RGBCube]

Hmm, I guess since a "current" usage would still be "usage average in the last ", we should do { cpu-usage-since = "5m" } or something similar. Do you know of a path that lets us do this nicely? And how granular can this even get?

[stefanboca]

It should be fairly simple if the CpuStats are stored in CpuLog. If you have those, then you can pick the latest CpuStat and the CpuStat from closest to 5 minutes ago and do something like https://github.com/stefanboca/watt/blob/49f482c21765b52b81c1302500ec582d52efd8ce/watt/cpu.rs#L53-L81. This of course limits the time intervals to the polling intervals.

I believe /proc/stat is updated on every scheduler tick, which on my system is 1000Hz (zcat /proc/config.gz | grep 'CONFIG_HZ='). But in practice I've noticed that anything below ~200ms between polls is too noisy to be useful.

I'm in favor of cpu-usage-since, but I also think that a cpu-usage or similar for the usage between the last two polling iterations might be useful.

[NotAShelf]

Could you see if #53 is more or less along the lines of what you were envisioning? I'm going to be traveling today, so I might not have given as much attention as the issue deserves, but I'm mostly positive that besides mathematical accuracy (which would get fixed with the time-delta impl. later) we've fixed the miscalculation. I've also went ahead and added the cpu-usage-since expression since it's very nice for fine-graining your rules for specific use cases.