Machine learning feature selection for tuning memory page swapping

Abstract

This thesis is an exploration of the virtual memory subsystem in the modern Linux kernel. It applies machine learning to find areas where better page-out decisions can be made. Two areas of possible improvement are identified and analyzed. The first area explored arises because pages in a computation appear repeatedly in a sequence. This is an example of temporal locality. In this instance, we can predict pages that will not be recalled again from the backing store with a precision and recall of 0.82 and 0.81, respectively, with a baseline of 0.30. The second is trying to predict when the system has made bad page-out decisions, those which lived in the backing store for less than one second before being recalled into RAM. In this case, we achieved a precision of 0.82 and a recall of 0.81 with a baseline of 0.12.