You are right in that Linux will hog all free memory for buffers and cache, and this is exactly what a host OpenVZ server tries to do. Underutilized OpenVZ servers will use more buffers/cache than an oversubscribed server because an underutilized server will have more free memory.

Virtual environments do not have buffers and cache of their own (with the exception of network buffers). Instead, they use the host server’s global buffers and cache. With an underutilized OpenVZ server, this can lead to a single active VE using a disproportionately large amount of the host server’s buffers and cache.

If the virtual environments are set up so that they all point to the same system libraries, they will be sharable throughout and will not consume additional memory for each additional VE. In other words, you’d only see one copy of the system library in memory, not one copy for each VE. For resource accounting, each VE would be charged for only a portion of the in-memory library size.

We’re not sharing system libraries across VEs in our setup, so I’m not sure what kind of memory savings this would result in. I can see how this feature would be useful for hosting providers looking to squeeze as many VEs as possible out of a machine.

Furthermore, what about shared libraries? If they are the same across different virtual instances, shouldn’t their memory consumption be shareable throughout those, too? Is something done along optimizing memory usage in that direction, or isn’t it, meaning that the explained overcommiting of memory is only a drop in the ocean?