Examining the User-space Semantic Gap from Page Reclaim

In modern system environments, a single Linux host frequently executes diverse workloads concurrently, including databases, virtual machines (VMs), containers, and batch jobs. However, during Page Reclaim, the kernel lacks application-layer semantics and must rely exclusively on low-level signals—such as the Accessed bit, Refault, and Page Type—to infer data hotness. Consequently, achieving optimal memory management decisions for specific workloads remains challenging.

This session will begin by examining the Linux reclaim path and the evolution of its source code. It will explore how the Active/Inactive LRU and Multi-Gen LRU (MGLRU) utilize heuristic algorithms to infer page hotness, alongside the limitations inherent to these designs under modern workloads. Subsequently, the session will review mechanisms such as madvise, posix_fadvise, and userfaultfd, illustrating how Linux enables User-space to transition from merely providing hints to actively participating in the memory decision-making path.

Finally, the session will introduce recent efforts by the Linux community to externalize specific memory management decisions using eBPF and BPF-based memory policy patchsets. This demonstrates a paradigm shift in memory management, moving from fixed heuristic algorithms toward a programmable decision framework, thereby allowing workload semantics to directly influence kernel memory decisions.