Hybrid Memory Systems를 위한 머신러닝 기반 페이지 스케줄러 - 석사학위 논문

Aug 3, 2026

Upload by Logic

Growing demand for machine learning and big data workloads makes memory performance a key determinant of application efficiency. This thesis studies hybrid memory systems composed of multiple memory types and shows that data placement across them strongly affects runtime. It proposes SMA, an RNN-based page scheduler that learns page access patterns and pre-prepares pages likely to be needed in fast memory. Using the insight that a subset of pages dominates performance, it manages important pages with a single RNN while handling remaining pages via a history-based method, reducing training time and memory usage versus state-of-the-art approaches. It also demonstrates that one RNN can generalize to unseen applications with comparable accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages31

Reading Time1 h 18 min

AccessFree to download

I Introduction

II Background

2.1 Hybrid Memory System

2.2 Recurrent Neural Networks

III Motivation

IV Design

4.1 Inference Plane

4.2 Training Plane

V Methodology

5.1 Applications

5.2 Simulated Hybrid Memory System

5.3 LSTM Model Implementation

5.4 Dateset for LSTM Model

VI Evaluation

6.1 Model Accuracy

6.2 Training Time

6.3 Memory Usage

6.4 Discussion

VII Related Work

VIII Future Work

IX Conclusion

References

Acknowledgements

저작자표시-비영리-변경금지 2.0 대한민국 이용자는 아래의 조건을 따르는 경우에 한하여 자유롭게 l 이 저작물을 복제, 배포 , 전송 , 전시 , 공연 및 방송할 수 있습니다 . 다음과 같은 조건을 따라야 합니다: 저작자표시 . 귀하는 원저작자를 표시하여야 합니다 . 비영리 . 귀하는 이 저작물을 영리 목적으로 이용할 수 없습니다 . 변경금지 . 귀하는 이 저작물을 개작 , 변형 또는 가공할 수 없습니다 . l 귀하는, 이 저작물의 재이용이나 배포의 경우, 이 저작물에 적용된 이용허락조건을 명확하게 나타내어야 합니다 . l 저작권자로부터 별도의 허가를 받으면 이 러한 조건들은 적용되지 않습니다 . 저작권법에 따른 이용자의 권리는 위의 내용에 의하여 영향을 받지 않습니다 . 이것은 이용허락규약( Legal Code) 을 이해하기 쉽게 요약한 것입니다 . Disclaimer Master’s Thesis A Page Scheduler using Machine Learning for Hybrid Memory Systems Minje Kim Department of Computer Science and Engineering Ulsan National Institute of Science and Technology A Page Scheduler using Machine Learning for Hybrid Memory Systems Minje Kim Department of Computer Science and Engineering Ulsan National Institute of Science and Technology Abstract As the demand for machine learning and big data workloads grows, the memory becomes important in application performance. The main memory is extended by using hybrid memory systems that include different types of memory components. Data placement across multiple memory components has a significant impact on application performance. We propose SMA, an RNN-based page scheduler to learn page access patterns and ensure that pages to be accessed by applications in the future are prepared in a fast memory in advance. This paper utilizes the existing observation that there is a set of pages that are important to application performance. A single RNN model manages all important pages for efficient page management, and the remaining pages are managed using a history-based method. This work reduces training time and memory usage compared to the existing state-of-the-art machine learningbased page scheduler while providing higher accuracy. It also shows that a single RNN model can learn general page accesses patterns by achieving similar accuracy with the existing page scheduler for applications that were not included in the training dataset of the model. Contents I Introduction ......................................... 1 II Background ......................................... 3 2.1 Hybrid Memory System .............................. 3 2.2 Recurrent Neural Networks ............................ 3 III Motivation .......................................... 5 IV Design ............................................ 7 4.1 Inference Plane ................................... 7 4.2 Training Plane ................................... 9 V Methodology ........................................ 10 5.1 Applications .................................... 10 5.2 Simulated Hybrid Memory System ........................ 10 5.3 LSTM Model Implementation ........................... 10 5.4 Dateset for LSTM Model .............................. 10 VI Evaluation .......................................... 11 6.1 Model Accuracy .................................. 11 6.2 Training Time .................................... 13 6.3 Memory Usage ................................... 13 6.4 Discussion ..................................... 14 VII Related Work ........................................ 15 VIII Future Work ......................................... 16 IX Conclusion ......................................... 17 References ............................................. 18 Acknowledgements ........................................ 20 List of Figures 1 Overview of our hybrid page scheduler SMA that combines the intelligent page management and existing history page scheduler......................... 7 2 DRAM hit rate according to the number of misplaced pages managed by the Oracle page scheduler........................................... 8 3 Mean absolute error (MAE) for 100 important pages of SMA and Kleio for the first combination of selecting three unseen applications. Unseen applications are backprop, cpd, and pennant. Applications with orange bars mean unseen applications

FAQs

What problem does SMA address in hybrid memory systems?

SMA addresses how to place and prepare pages across multiple memory types so that future page accesses are available in fast memory, improving application performance.

How does SMA use machine learning for page scheduling?

SMA uses an RNN-based scheduler to learn page access patterns for important pages, while managing the remaining pages using a history-based method.

What are the main benefits compared with existing state-of-the-art approaches?

The approach reduces training time and memory usage while achieving higher accuracy, and it can generalize to unseen applications with similar accuracy.

Aug 3, 2026

Upload by Logic