How the Linux Kernel Observes Page Accesses and Infers the Working Set

As AI, cloud services, and large-scale applications continue to drive up memory demand, memory usage is no longer merely a capacity concern—it also directly affects system cost and runtime performance. From personal computers and mobile devices to everyday software such as web browsers and Electron applications, excessive memory consumption has become a common concern for both users and developers.

When analyzing a program’s memory usage on Linux, we often begin with metrics such as RSS and PSS. These metrics describe how much of a process’s memory is currently resident in physical memory, but they provide limited insight into which data still belongs to the program’s actual working set. Meanwhile, page hotness has no single, universal definition within the Linux kernel. Instead, the kernel approximates it using signals such as page accesses, aging, reclaim, and refaults.

Through reproducible experiments and kernel tracing, this talk examines the key implementations of Linux Idle Page Tracking, mm/workingset.c, and its MGLRU refault path. We will explore how the kernel observes page accesses, identifies refaults, and turns these signals into working-set inference and reclaim feedback.