Person  sitting on top of container and gold circle with 3 stars, looking at open laptop.

Checking Kubernetes pod CPU and memory

Matthew Casperson

Tracking the resource usage of local processes is relatively easy in Linux with the top or htop command. But how do you track resource usage of pods spread across a Kubernetes cluster?

In this post, I show you how to view the CPU and memory usage of pods in Kubernetes.

Read our white paper: A leader's guide to Kubernetes CD at scale

Learn how to overcome technical hurdles and business challenges and successfully implement Kubernetes.

Get the white paper

The metrics server

The metrics server provides Kubernetes clusters with a lightweight and highly scalable solution for collecting CPU and memory resources. Although the metrics server isn't built into Kubernetes, most clusters either bundle it or provide an easy solution for enabling it.

After the metrics service is installed, pod resources are displayed with the command:

kubectl top pod

Node resource usage is available with the command:

kubectl top node

The following error indicates that the metrics server is not installed:

error: Metrics API not available

In this case, you can install the metrics server with the instructions on GitHub.

cgroup resource usage

If the metrics service isn't available, it's still possible to determine the memory usage of a single pod by entering an interactive session and printing the contents of cgroup interface files.

Enter an interactive session with the following command, replacing podname with the name of the pod you wish to inspect:

kubectl exec -it podname -- sh

Print the current memory usage with the command:

cat /sys/fs/cgroup/memory/memory.usage_in_bytes

Print the current cpu usage with the command:

cat /sys/fs/cgroup/cpu/cpuacct.usage

Note the value returned by cpuacct.usage is not immediately useful as it returns:

the CPU time (in nanoseconds) obtained by this group

Converting this value into a more usable measurement like CPU usage percentage requires some calculation. This post on Stack Exchange provides more details, and this Python code provides a useful practical example.

You can find more information on these files in the Linux kernel docs.

cgroup2 resource usage

If the directories /sys/fs/cgroup/memory or /sys/fs/cgroup/cpu don't exist, you're likely working on a system with cgroups v2.

On systems with cgroups v2, print the current memory usage with the command:

cat /sys/fs/cgroup/memory.current

Print the current cpu usage with the command:

cat /sys/fs/cgroup/cpu.stat

This prints a file with a value called usage_usec. As with value returned by the cgroup v1 cpuacct.usage file, this value must be converted into a CPU usage percentage to be useful.

Note the usage_usec value is measured in milliseconds, unlike the value returned by the cpuacct.usage file, which is in nanoseconds. Convert the usage_usec value to nanoseconds by multiplying it by 1000, at which point it can be used in the same calculations returned by the cpuacct.usage file.

You can find more information on these files in the Linux kernel docs.

Conclusion

The metrics server provides a convenient method for inspecting the CPU and memory resources of your Kubernetes pods and nodes. It's also possible to find these values manually by inspecting the cgroup interface files, although some manual calculations are required to determine CPU usage as a percentage.

Happy deployments!

Loading...