Publications

Detailed Information

낸드 플래시 기반 저장장치의 성능 및 신뢰성 향상을 위한 계층 교차 최적화 기법 : Cross-Layer Optimization Techniques for Improving Performance and Reliability of NAND Flash-Based Storage Systems

Cited 0 time in Web of Science Cited 0 time in Scopus
Authors

하건수

Advisor
김지홍
Major
공과대학 전기·컴퓨터공학부
Issue Date
2015-08
Publisher
서울대학교 대학원
Keywords
NAND Flash MemoryFlash-Based Storage DevicesStorage Performance OptimizationOperating SystemEmbedded SystemStorage Reliability Management
Description
학위논문 (박사)-- 서울대학교 대학원 : 전기·컴퓨터공학부, 2015. 8. 김지홍.
Abstract
As the cost-per-bit of NAND flash memory is quickly improved by advanced process technologies and multi-leveling techniques, NAND flash-based storage systems are widely employed from mobile embedded systems to high-end enterprise server systems. Although the advanced process and device techniques have greatly improved the cost-per-bit of NAND flash memory, they have also significantly degraded the performance and reliability of NAND flash memory as key side effects of the advanced techniques. In order for NAND flash-based storage systems to be more broadly used in various computing environments, it is critical to overcome the performance and reliability problems of recent high-density NAND flash memory in a satisfactory fashion. In this dissertation, we argue that cross-layer optimization techniques, which vertically integrate various optimization factors from different design abstraction levels, can play key roles in improving performance and reliability of high-density NAND flash memory.
First, we propose read-disturb management techniques which reduce the expensive read-disturb management overheads while maintaining reliability of NAND flash memory. An FTL using the read-disturb management module, called redFTL, alleviates highly skewed read accesses to a small part of NAND flash memory into more balanced read accesses to a large number of blocks, thus reducing data migrations needed for avoiding read-disturb errors. As an extended version of redFTL, we propose an integrated read-disturb management technique, called redFTL+, which fundamentally solves read-disturb problems by exploiting a tradeoff between the read disturbance and write speed. By modifying NAND chips to support multiple read modes with different read voltages and write speeds, redFTL+ intelligently allocates frequently-read data to read-resistant blocks. Since the read disturbance is also proportional to the read time, redFTL+ takes advantage of the difference in the read time among different NAND pages by reallocating read-intensive data to read-resistant pages.
Second, we propose data separation techniques which reduce garbage collection overhead. We propose a program context-aware data separation technique, called PDS, which can reduce the garbage collection overhead by exploiting program context hints. By using a program context, which serves as a proper granularity of maintaining data update behavior, PDS helps an FTL gather data with similar update times to the same blocks. As an improved version of PDS, we propose an integrated data separation technique, called IDS, which uses both update history of NAND device and program context hints for predicting data update behaviors. By classifying data based
on the cross-layer information, an FTL using IDS can make more dead or near-dead blocks over PDS, thus reducing the garbage collection overhead. In order to evaluate the effectiveness of the proposed techniques, we performed a series of evaluations using both a simulator and an emulator with I/O traces which were collected from various systems. Our experimental results show that cross-layer optimization techniques are more effective over our single-layer optimization techniques. RedFTL+ decreases the read-disturb management overhead on average by 24% over redFTL. The IDS-based FTL decreases the garbage collection overhead on aver-age by 18% over the PDS-based FTL. The evaluation results demonstrate that our cross-layer optimization techniques improve an overall performance of NAND-based storage systems over previous single-layered optimization techniques by reducing overheads from read-disturb management and garbage collection while maintaining the reliability of the storage systems.
Language
Korean
URI
https://hdl.handle.net/10371/119123
Files in This Item:
Appears in Collections:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share