from COMPUTER to COMPUTing centER: the history of linux memory management

The presentation is structured into three main segments:

1. The History of RMAP We begin by tracing the origins of RMAP from the pte_chain (per-page reverse mapping) in Linux v2.5, examining its initial scalability challenges. The talk will explore the design background of anon_vma, introduced by Andrea Arcangeli in 2004 to address VM scalability, and discuss how anon_vma_chain mitigates scalability bottlenecks in heavy fork workloads. Finally, we will cover the adoption of interval trees to optimize VMA lookups and RMAP walk scalability within extremely large address spaces.
2. The Role of RMAP in Modern Page Replacement The second segment highlights the integration of RMAP within contemporary memory management strategies, particularly Multi-Gen LRU (MGLRU). We will discuss how MGLRU utilizes generation-based aging to significantly reduce redundant scanning of page tables and reverse mappings during the page reclaim process, emphasizing its critical importance in optimizing modern datacenter workloads.
3. Experimental Quantification of RMAP Overhead In the final section, we present an empirical analysis of RMAP overhead. By leveraging bpftrace to hook into rmap traversal path, we measure and compare the latency distributions when Transparent HugePage (THP) is enabled versus disabled. Furthermore, we will analyze how specific kernel paths(such as huge page split, folio reclaim, and deferred split queues) impact the overall cost of reverse mapping operations in real-world scenarios.