The problem is not the balanced memory allocation of the said process, but the kernel evicting IO caches:

https://groups.google.com/forum/#!topic/fa.linux.kernel/4IKY...

We experienced this bug in production on hardware where the cost of accessing the other node is exactly over 20 (the limit defined in the kernel).

node distances: node 0 1 0: 10 21 1: 21 10