Stack usage analysis for efficient wear leveling in non-volatile main memory systems

C. Hakert, M. Yayla, K.-H. Chen, G.V.D. Bruggen, J.-J. Chen, S. Buschjager, K. Morik, P.R. Genssler, L. Bauer, H. Amrouch, J. Henkel

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

Abstract

Emerging non-volatile memory (NVM) technologies, such as Phase Change Memory (PCM), have been considered as a replacement for DRAM and storage due to their low power consumption, fast access speed, and low unit cost. Even so, some NVMs have a significantly lower write endurance and hence in-memory wear leveling is an important requirement for practical applicability. Since writes to the stack often target a small and dense memory region, generic, coarse-grained wear-leveling mechanisms (e.g. virtual memory page remapping) are not sufficient. An alternative solution is to relocate the stack memory regularly, which involves copying of the stack content. As the stack content changes in size during the execution of an application, the copy overhead can be significantly mitigated by performing the relocation when the stack size is small.In this paper, we investigate two approaches to determine points in time when the stack is small. First, we analyze the possibility to fit simple machine-learning models to the stack usage function. Precise predictions of this function enable the identification of the minimum stack size during execution. In our evaluation, the tested models provide accurate estimates of the future stack usage function for a subset of common applications.As a second approach, we analyze applications a priori and determine potential optimal points to perform relocation in the instruction stream. In detail, we deploy the application in an analysis environment, which determines a rating for each executed instruction. Based on this rating, we apply a genetic algorithm to identify the best points in the instruction stream to perform the stack relocation. This approach allows to save up to 85% of the write overhead for wear-leveling in our experiments.
Original languageEnglish
Title of host publication2019 ACM/IEEE 1st Workshop on Machine Learning for CAD, MLCAD 2019
ISBN (Electronic)978-1-7281-5758-0
DOIs
Publication statusPublished - 2019
Externally publishedYes
Event1st International Workshop on Challenges in Artificial Intelligence and Machine Learning for Internet of Things, AIChallengeIoT 2019 - New York, United States
Duration: 10 Nov 201913 Nov 2019
Conference number: 1

Conference

Conference1st International Workshop on Challenges in Artificial Intelligence and Machine Learning for Internet of Things, AIChallengeIoT 2019
Abbreviated titleAIChallengeIoT
CountryUnited States
CityNew York
Period10/11/1913/11/19

Fingerprint

Dive into the research topics of 'Stack usage analysis for efficient wear leveling in non-volatile main memory systems'. Together they form a unique fingerprint.

Cite this