Memory pages allocated by application, pinned to NUMA node, are being slowly migrated to another NUMA node with swapping involved

Issue

Application was restricted to use memory only from a specific NUMA node. libnuma was used to pin the application (see How to pin a C application to a specific NUMA node using libnuma?). Also, there was an attempt for cpuset to achieve the same isolation (see How to use cgroup cpusets with systemd in RHEL7?).

However, if the application requests for more memory than the NUMA node has, thus triggering swapping, then pages are being migrated to another NUMA node. The migration happens slowly, by 2 MiB chunks.