$(nproc) X 1/2 MiB
The basic units of OpenShift Container Platform applications are called containers. Linux container technologies are lightweight mechanisms for isolating running processes so that they are limited to interacting with only their designated resources.
Many application instances can be running in containers on a single host without visibility into each others' processes, files, network, and so on. Typically, each container provides a single service (often called a "micro-service"), such as a web server or a database, though containers can be used for arbitrary workloads.
The Linux kernel has been incorporating capabilities for container technologies for years. OpenShift Container Platform and Kubernetes add the ability to orchestrate containers across multi-host installations.
Due to Red Hat Enterprise Linux (RHEL) behavior, a container on a node with high CPU usage might seem to consume more memory than expected. The higher memory consumption could be caused by the
kmem_cache in the RHEL kernel. The RHEL kernel creates a
kmem_cache for each cgroup. For added performance, the
kmem_cache contains a
cpu_cache, and a node cache for any NUMA nodes. These caches all consume kernel memory.
The amount of memory stored in those caches is proportional to the number of CPUs that the system uses. As a result, a higher number of CPUs results in a greater amount of kernel memory being held in these caches. Higher amounts of kernel memory in these caches can cause OpenShift Container Platform containers to exceed the configured memory limits, resulting in the container being killed.
To avoid losing containers due to kernel memory issues, ensure that the containers request sufficient memory. You can use the following formula to estimate the amount of memory consumed by the
nproc is the number of process reported by the
nproc command. The lower limit of container requests should be this value plus the container memory requirements:
$(nproc) X 1/2 MiB