CPU more than 100%

[gsamsonov1989]

Hello, we run jprofiler for CPU and allocations and collecting results into a JFR file. After examining jfr file we can see that application + JVM CPU usage is reported as 120% but overall machine CPU is reported at 100%. This seems not ok. I'd like to know how async profiler collects this percentage stats (overall machine CPU and jvm (user + system) CPU). JVM runs in a docker container

[franz1981]

as said elsewhere, the defaul way to collect profiles is via cpu-clock events, per thread, unless jprofiler is making uses of itimer, which seems not given that you have more samples in the jfr profiling data than via cpu statistics collected outside of async-profiler...
My guess is that machine CPU usage is not accurate enough (it's an average? it access the same stats from the CPU? how?), but cpu-clock perf events should be decently accurate (although they don't include some interrupt handling, meaning you can still have fever cpu usage perception than the real one).

[gsamsonov1989]

@franz1981 i am sorry mate i don't understand your answer. Yes we use itimer and it generates profiling signal every 10ms interval. But then you lost me. What is you have more samples in the jfr profiling data than via cpu statistics collected outside of async-profiler ? What is it's an average? it access the same stats from the CPU? how? ? Question is how exactly CPU percentage usages are calcualted, for machine and for jvm (user + system)

[franz1981]

Afaik async profiler, with itimer, can and will loose samples, see #272 (comment)

And, additionally, if kernel stacks cannot be collected, it won't be able to report the kernel CPU consuming ones too, further reducing the amount of samples.

[gsamsonov1989]

And the loss of samples somehow leads to CPU usage (application + jvm) being 120% and overall machine usage being 100% ?

[apangin]

we run jprofiler

What does jprofiler have to do with this?!

After examining jfr file we can see that application + JVM CPU usage is reported as 120%

Can you show an example .jfr?

[gsamsonov1989]

@apangin I am sorry i meant async profiler, not jprofiler.

Unfortunately i can't show becuse of legal reasons.

We assumed it's because there is only 1 core reported from the JVM. There existed a bug some time when JVM in docker reported only 1 core available instead of all the cores available on the machine. We assumed async profiler collects these events from JVM and that's that.

[apangin]

Right, async-profiler previously obtained the number of available processors from the JVM.
This has been changed in #751 - now it shows whatever OS reports. So the bug is likely fixed already.