Merlin: Improving Page Prefetching via Online Reinforcement Learning

While the operating system (OS) page swapping mechanism enables memory oversubscription, memory-intensive workloads suffer significant performance degradation when the main memory cannot accommodate their entire working sets. This is primarily caused by the performance gap between main memory and backing storage. To hide the swapping overhead, prefetching techniques predict future data accesses and proactively load necessary contents from slower backing storage into the main memory. However, existing OS prefetching algorithms rely on predefined rules to analyze page access patterns, which are not adaptive to changing workloads and fail to align prefetching policies with system status. In this paper, we introduce Merlin , an adaptive, fully automated, and application-transparent page prefetching scheme that employs reinforcement learning (RL) to make intelligent prefetch decisions. We formalize the prefetcher as an RL agent, which extracts critical features from the swap request and system environment to continuously make prefetch decisions. For each prefetch decision, Merlin receives a numerical reward or penalty based on the current system status and prefetch-hit. Merlin learns towards getting higher rewards to make more timely, accurate, and system-aware prefetch decisions. Through extensive evaluation, Merlin achieves up to \(98\% \) prefetching accuracy with less than 20KB metadata overhead, requiring no code modifications to workloads. Compared to Linux’s default prefetcher, it delivers up to 2.7 × speedup in overall execution time on microbenchmarks and up to 1.5 × on real-world workloads.

Authors

Institutions

Publication Details

Journal
ACM Transactions on Storage
Published
2026-11-18
DOI
https://doi.org/10.1145/3857346
Primary Topic
Parallel Computing and Optimization Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Merlin: Improving Page Prefetching via Online Reinforcement Learning

Yingying Liu, Chenxiong Qian, Junzhe Li, Junzhou Fang
2 links
ACM Transactions on Storage
Parallel Computing and Optimization Techniques
article

Merlin: Improving Page Prefetching via Online Reinforcement Learning

Yingying Liu, Chenxiong Qian, Junzhe Li, Junzhou Fang
article en

Abstract

While the operating system (OS) page swapping mechanism enables memory oversubscription, memory-intensive workloads suffer significant performance degradation when the main memory cannot accommodate their entire working sets. This is primarily caused by the performance gap between main memory and backing storage. To hide the swapping overhead, prefetching techniques predict future data accesses and proactively load necessary contents from slower backing storage into the main memory. However, existing OS prefetching algorithms rely on predefined rules to analyze page access patterns, which are not adaptive to changing workloads and fail to align prefetching policies with system status. In this paper, we introduce Merlin , an adaptive, fully automated, and application-transparent page prefetching scheme that employs reinforcement learning (RL) to make intelligent prefetch decisions. We formalize the prefetcher as an RL agent, which extracts critical features from the swap request and system environment to continuously make prefetch decisions. For each prefetch decision, Merlin receives a numerical reward or penalty based on the current system status and prefetch-hit. Merlin learns towards getting higher rewards to make more timely, accurate, and system-aware prefetch decisions. Through extensive evaluation, Merlin achieves up to \(98\% \) prefetching accuracy with less than 20KB metadata overhead, requiring no code modifications to workloads. Compared to Linux’s default prefetcher, it delivers up to 2.7 × speedup in overall execution time on microbenchmarks and up to 1.5 × on real-world workloads.

ACM Transactions on Storage
Zhejiang University (CN), University of Hong Kong (HK)
Parallel Computing and Optimization Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon