An Optimized Storage Architecture for Improving ML Platforms Provisioned with Underlying Deduplication Enabled Storage Clusters
- 주제(키워드) Machine Learning based storage architectures.
- 발행기관 아주대학교
- 지도교수 Tae-Sun Chung
- 발행년도 2021
- 학위수여년월 2021. 2
- 학위명 박사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000030533
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
The advancement and ubiquitousness of Machine Learning (ML) is unarguably the new wave driving modern day and future enterprise computing platforms. However, the incessant deluge of ML associated data, collected from millions of data sources presents data storage challenges. Continuous scaling of the storage to meet the ML storage demands results in unwarranted escalating storage demands. However, there exists a lot of duplicate data in ML/DL related workloads in which if eliminated will result in significant amortization of storage costs. The adoption of deduplication provisioned storage has been so far a storage cost cutting driver in today’s enterprise clusters. However, these large scale machine learning (ML) platforms are facing challenges when integrated with deduplication enabled storage clusters. In the quest to achieve smart and efficient storage utilization, removal of duplicate data introduce bottlenecks since deduplication alters the I/O transaction layout of the storage system. Therefore, it is critical to address such deduplication overhead for acceleration of ML/DL computation in deduplication storage. Existing state of the art ML/DL storage solutions such as Alluxio and Auto-Cache adopt non deduplication-aware caching mechanisms which lacks the much needed performance boost when adopted in deduplication enabled ML/DL clusters. In this paper, we introduce REDUP, which eliminates the performance drop caused by enabling deduplication in ML/DL storage clusters. At the core, is a REDUP Caching Manager (RDCM), composed of a 2-tier deduplication layout-aware caching mechanism. The RDCM provides an abstraction of the underlying deduplication storage layout to ML/DL applications and provisions a decoupled acceleration of object reconstruction during ML/DL read operations. Our REDUP evaluation shows negligible performance drop in ML/DL training performances as compared to a baseline cluster without deduplication. When compared to other state-of-the-art solutions, our proposed design outperforms Alluxio and Auto-Cache by 16% in worst case scenario, in terms of training speed.
more목차
1 Introduction 1
1.1 Introduction 1
1.2 Contributions of This Dissertation 6
2 Background 7
2.1 Related Works 7
2.1.1 ML Pipeline and Data Access Patterns 9
2.1.2 Data Duplicate Elimination 14
Cluster-scale Data Deduplication 15
2.2 Motivation 20
2.2.1 Dedup challenges in ML and DL platforms 20
2.2.2 Preliminary Analysis 22
3 Proposed REDUP architecture 27
3.1 REDUP Overview 27
3.1.1 Goals 27
3.1.2 System Overview 29
RCR cache pool 31
RD cache pool 32
3.2 Design and Implementation 33
3.2.1 Hierarchical Object Caching 33
3.2.2 RCR Cache 34
RFCB caching policy 34
LA-caching policy 35
RCR cache admission and replacement flow 36
3.2.3 REDUP Object 37
3.2.4 RD Cache 37
RCO HitSet 38
HA-EB caching policy 39
RD cache admission flow 40
RD cache replacement flow 42
3.2.5 Cache Throttling Manager 44
4 Performance Evaluation 46
4.1 Evaluation 46
4.1.1 Testbed Setup 48
Neural Networks and ML platforms 48
4.1.2 Performance Analysis with CNNs 49
4.1.3 Performance analysis with RNNs 52
4.1.4 Hit Ratio Analysis 54
Read Hit Ratio 55
Effect of incremental component addition in REDUP 55
4.1.5 Effect of RD SSD Temp 57
4.1.6 Cumulative distribution function of Training progress 59
5 Conclusions 60
5.1 Conclusion 60
References 62
