Report on Online tracking of Working Sets, and Locality Study of Dirty and Reference Bits

Effective management of a machine’s memory resources needs comprehensive information regarding the realtime memory consumption of applications. Memory demands can be estimated accurately by tracking the working set sizes of programs. We propose a software approach for monitoring the working set size of applications, based on Denning’s mathematical model of working sets. This technique is implemented using an existing tool, BadgerTrap, and its ability to accurately track working set sizes is demonstrated using memory-intensive workloads. Large pages are increasingly being used for improving the performance of big-memory applications. However, current implementations fail to accommodate page swapping adequately, and thus have limited use in conditions of memory pressure. We observe the distribution of dirty and reference bits for several workloads to study the feasibility of tracking these at a large 2MB page granularity. These observations are made in the presence and absence of memory pressure, to observe the effects of page swapping.