Extracting accounting information from a container

[yaqwsx]

Hello,
I would like to be able to extract accounting information from a running or stopped rootless container (e.g., memory usage, CPU time/cycles used). The podman inspect does not give any of this information; podman stats gives only information about memory (in human-readable units) but no information about consumed CPU cycles. Is there any other podman command that can give me this information?

Since I wasn't able to get the accounting information from podman commands, I tried to inspect the group mounted in the container under /sys/fs/cgroup. When I examine these files from the container, they give me the information I want. However, I am struggling with an effective way of reading them outside container. The podman cp reports files under /sys as missing. I could in theory use podman exec <cId> cat /sys/fs/cgroup/cpu.stats but that affects the statistics (probably negligible, but still). What is a nice and clean way to read out the accounting information from groups?

[mheon]

The podman top command seems like it would do what you want - both memory and CPU cycles are available as format descriptors.

[mheon]

The container subgroup is an implementation detail specific to some OCI runtimes. It shouldn't matter, though - the parent cgroup should have cumulative statistics that include all sub-cgroups (at least on cgroups v2, where cgroups were redesigned to be sane).

[yaqwsx]

Perfect, thanks for the confirmation!

[rhatdan]

This might be something we would think about extending podman stats to gather and present to the user.

[rhatdan]

podman stats --help
Display a live stream of container resource usage statistics

Description:
Display percentage of CPU, memory, network I/O, block I/O and PIDs for one or more containers.

[yaqwsx]

Thank you for pointing me in the right direction. I think the procedure mentioned above will work; I have just stumbled upon this (probably a bug): #10173 The workaround would be to execute podman from a process in the target cgroup, however, that would affect the collected statistics.

@rhatdan I think that would be useful for my use case; however, I am not sure if there is a widespread demand for such a feature.