Performance prediction of paging workloads using lightweight tracing

Abstract

A trace of a workload's system calls can be obtained with minimal interference, and can be used to drive repeatable experiments to evaluate system configuration alternatives. Replaying system call traces alone sometimes leads to inaccurate predictions because paging, and access to memory-mapped files, are not modelled. The paper extends tracing to handle such workloads. At trace capture time, the application's page-level virtual memory access is monitored. The size of the page access trace, and capture overheads, are reduced by excluding recently-accessed pages. This leads to a slight loss of accuracy. Using a suite of memory-intensive applications, we evaluate the capture overhead and measure the predictive accuracy of the approach.