A Reliable Data Management Framework based on Access Patterns for Flash Storage Systems
- 주제(키워드) Flash Storage Systems , Data Management Framework , Diverse Access Patterns
- 발행기관 아주대학교
- 발행년도 2010
- 학위수여년월 2010. 2
- 학위명 박사
- 학과 및 전공 정보통신전문대학원 정보통신공학과
- 파일정보 pdf
- 실제URI http://www.dcollection.net/handler/ajou/000000010305
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Flash memory is hugely been using for small handheld to large enterprise applications due to its attractive features. However, flash hardware characteristics like erasebefore-write and limited-erase-cycles are becoming big hurdle for researchers to provide reliable and performance oriented system softwares. The effective way to mitigate the impacts of flash drawbacks is to manage the data by its access patterns. But, as a side affect, such approach imposes the demand of high main memory space and lengthy time for initialization. This thesis proposes the work towards reliable data management for NAND flash memory based storage systems. Proposed framework classifies data intellectually according to their access frequencies and proves the efficiency and effectiveness for data and memory management on all levels of system operations. Meticulous analytical discussions and comprehensive experimental results demonstrate the highly improved system performance achieved by considering the diverse natures of data.
more목차
List of Figures xiii
List of Tables xv
Abstract xvi
1 ？ Introduction 1
1.1 Brief Background 1
1.1.1 Flash Memory Types 2
1.1.2 Flash Memory Constraints 3
1.1.3 Flash Translation Layer 5
1.1.3.1 Non-In-Place Memory Mapping 5
1.1.3.2 Free Space Maintaining 5
1.1.3.3 Leveling Erasures on Blocks 6
1.1.3.4 Maintaining System Reliability 7
1.2 Motivation 8
1.3 Overview of Contributions 10
1.4 Organization of Thesis 11
2 ？ Proposed Data Management Framework 12
2.1 Memory Blocks Structure 12
2.2 Data Organization Framework 13
2.2.1 Data Blocks Allocation 13
2.2.2 Write operation 15
x
2.2.3 Log Blocks Allocation 16
2.2.4 Read Operation 16
2.3 Free Space Revival 17
2.3.1 Hot Log Blocks Cleaning 18
2.3.2 Cold Log Blocks Cleaning 19
2.3.3 Hot/Cold Data Blocks Cleaning 19
2.4 Block Allocation Strategy 22
2.5 Device Life Enhancement 22
2.5.1 Hybrid Wear-Leveling Policy 23
2.6 Data Reliability Sustainability 26
2.6.1 Mounting Technique 26
2.6.1.1 Initial RAM footprints 27
2.6.1.2 RAM footprints reduction on run time
28
2.6.2 Crash Recovery Technique 29
2.6.2.1 User-Data Crash Recovery 29
2.6.2.2 Metadata Crash Recovery 31
2.6.2.3 Erase Operation Crash Recovery 36
2.7 Summary 40
3 ？ Performance Evaluation 41
3.1 Simulation Environment 41
3.2 Experimental Results 43
3.2.1 Memory Bandwidth Utilization 43
3.2.2 Reduced Erase Operations 45
xi
3.2.3 Even Erase Operations 46
3.2.4 Main Memory Consumption 48
3.2.5 Time Requirements 51
3.3 Summary 53
4 ？ A Case Study: PIYAS 54
4.1 Introduction 54
4.2 System Architecture of Sensor Node 59
4.3 PIYAS: Proposed Memory Management Scheme 61
4.3.1 Data Organization Framework 61
4.3.1.1 Data Buffers Management 62
4.3.1.2 Memory Blocks Organization 65
4.3.2 Mapping Structures Management 68
4.3.3 Query Processing Framework 71
4.3.3.1 Query on Raw Data Blocks 75
4.3.3.2 Query on Aggregate Data Blocks 76
4.3.4 Garbage Collection 77
4.3.5 Wear Leveling 80
4.4 Performance Evaluation 81
4.4.1 Simulation Methodology 81
4.4.2 Experimental Results 83
4.4.2.1 Memory Bandwidth Utilization 84
4.4.2.2 Main Memory Consumption 85
4.4.2.3 Throughput Performance 86
xii
4.4.2.4 Fast Initialization 89
4.4.2.5 Resources Preservation 89
4.4.2.6 Throughput Optimization 90
4.5 Summary 92
5 ？ Conclusion and Future Research Goals 94
5.1 Future Plans 97
5.1.1 Large Enterprise DBMSs and Flash Memory 97
5.1.2 Sensitive applications and Flash Memory 97
5.2 Summary 98
References 99
