Working Set Model for Multithreaded Programs appears in the 2014 Conference on Timely Results in Operating Systems (TRIOS).
Abstract: Knowledge of the working set of pages associated with an application provides an opportunity for effective allocation of resources in multicore multiprocessor systems. Various techniques for approximating the working set size using either simulations or program traces have been proposed. However, these techniques are very expensive, and therefore not practical for dynamically optimizing resources. To alleviate this problem, in this work, we develop a statistical model based on machine learning techniques for approximating the working set size of multithreaded programs running on multicore multiprocessor systems. The basic idea is to correlate the working set size of a program with its resource usage characteristics, such as resident set size and TLB miss rate. Through extensive experimentation with 20 multithreaded programs from SPEC OMP2012 and PARSEC on a SPARC T4-4 running Oracle Solaris 11, we demonstrate that the model has 96% prediction accuracy and its overhead is negligible.