Architectural Design of In-Storage Processing Systems for High-Performance Query Processing
고성능 질의 처리를 위한 스토리지 내 처리 아키텍처 설계
- 주제(키워드) In-Storage Processing , Computational Storage , Key-Value SSD , Vector Database SSD , Multi-Tenancy , Secondary Indexing , Scheduling for SLOs
- 발행기관 서강대학교 일반대학원
- 지도교수 김영재
- 발행년도 2026
- 학위수여년월 2026. 8
- 학위명 박사
- 학과 및 전공 일반대학원 컴퓨터공학과
- 세부분야 해당없음
- 실제URI http://www.dcollection.net/handler/sogang/000000083035
- UCI I804:11029-000000083035
- 본문언어 영어
- 저작권 논문은 저작권에 의해 보호받습니다.
초록(요약문)
The explosive growth of data has intensified the mismatch between host processor-centric architectures and modern data-intensive workloads. While modern PCIe interconnects and NVMe solid-state drives （SSDs） deliver unprecedented bandwidth and capacity, the inherent inefficiency of host-driven I/O persists, as all computation occurs at the host CPU, requiring data to traverse repeatedly across the storage hierarchy before any processing can take place. In-storage processing （ISP） paradigm addresses this fundamental limitation by enabling data processing directly within storage devices, reducing costly data movement and freeing host resources. This dissertation presents architectural foundations for ISP systems that integrate diverse data management logic within SSDs. Three complementary systems are proposed, each targeting a distinct but practically significant limitation of existing ISP systems. First, PRESCAR introduces a preemptive, SLO-aware scheduling framework for vector database SSDs （VDSSDs） that enables concurrent online data ingestion and similarity vector search under tight device resource constraints. By estimating per-operation latency and resource demand, PRESCAR dynamically coordinates competing operations to maximize satisfaction of freshness, availability, and query performance objectives. Second, ISO-KVSSD establishes tenant isolation for key-value SSDs （KVSSDs） through a per-namespace LSM-tree indexing structure, achieving both data access and read performance isolation across tenants without proportional increases in on-board DRAM consumption. ISO-KVSSD achieves up to 2.9× higher read throughput under multi-tenant workloads with negligible write overhead. Third, CERBERUS redesigns secondary indexing for predicate-based analytical queries within KVSSDs by proposing a value-addressable secondary （VAS） indexing strategy. This strategy directly maps non-primary attributes of value part to data locations, eliminating conventional multi-hop traversal overhead. CERBERUS improves query processing speed by up to 9.31× and reduces significant data transfer compared to existing approaches. Collectively, these systems demonstrate that practical ISP system requires not merely hardware acceleration, but principled co-design of data structures, scheduling, and resource management within constrained device environments, providing both architectural insights and concrete guidelines for next-generation, low-cost storage systems.
more목차
1 Introduction 16
1.1 Emergence of Computational Storage Paradigm 18
1.2 Limitations of Current In-Storage Processing Systems 20
1.3 Research Objectives 23
1.4 Contributions of the Dissertation 24
1.5 Dissertation Organization 26
2 Architectural Background of Modern SSDs 27
2.1 Overall SSD Architecture 27
2.2 Internal I/O handling flow 29
2.3 Vector Database SSDs and its Execution Models 29
2.3.1 Offline Vector Insert 30
2.3.2 Offline Index （HNSW） Build 31
2.3.3 Online Vector Search 31
2.4 Prior Works for ANN-based VDSSDs 32
2.5 Key-Value SSDs based on Log-Structured Merge-Tree 32
2.5.1 LSM-tree Structure 34
2.5.2 Workflow of LSM-tree-based KVSSD 35
2.6 Prior Works for LSM-tree-based KVSSDs 37
3 Maximizing SLOs for Online Ingestion and Search Operations in Vector Database SSD Environments 38
3.1 Motivation: Inability to balance and maximize diverse SLOs 39
3.2 PRESCAR: A Framework for SLO-Aware VDSSD Design 40
3.2.1 Design Goals and Our Research Scope 40
3.2.2 Towards Online Ingestion Support in VDSSDs 41
3.2.3 Resource Contention and Challenges 46
3.2.4 Preemptive and SLO-Aware Scheduling Framework 49
3.3 Discussion 56
3.3.1 Applicability to HPC Storage Systems 56
3.3.2 SLO-specialized Configuration of PRESCAR 56
3.3.3 Generalizability of PRESCAR 57
3.3.4 Limitations 58
3.4 Evaluation 59
3.4.1 Experimental Setup 59
3.4.2 Availability vs. Freshness vs. Performance Conflict 61
3.4.3 Effectiveness of Preemption Policy 64
3.4.4 Usefulness of PRESCAR Scheduling for SLOs 65
3.4.5 Cross-validating PRESCAR Scheduling in Multiple SSDs 67
3.4.6 Overhead of Preemption and Polling 67
4 Supporting Namespace-based Tenant Isolation for Multi-Tenancy in Key-Value SSDs 69
4.1 Motivation: Lack of Tenant Isolation Mechanism 69
4.2 ISO-KVSSD: Design of Multi-tenant KVSSDs 73
4.2.1 Design Overview 73
4.2.2 Per-namespace LSM-tree Index Management 75
4.2.3 Empirical Study on Shared vs. Separated Index Components . 77
4.2.4 Index Isolation Mechanism 79
4.2.5 Namespace and Storage Space Management in ISO-KVSSD . 81
4.3 Evaluation 84
4.3.1 Experimental Setup 84
4.3.2 Per-tenant Performance Comparison 85
4.3.3 Impact of Per-namespace LSM-tree 88
4.3.4 Impact on Bloom Filter Overhead and Caching Efficiency 88
5 Enabling Predicate-based Query Processing Through Secondary Indexes in KVSSDs 91
5.1 Motivation: Lack of SSD-friendly Secondary Indexing Strategy 93
5.2 CERBERUS: Design of Secondary Indexing in KVSSDs 98
5.2.1 Design Overview 98
5.2.2 Value-addressable Secondary （VAS） Indexing 99
5.2.3 In-Storage Processing （ISP） with VAS Index 104
5.3 Evaluation 105
5.3.1 Experimental Setup 105
5.3.2 Evaluation for Effectiveness of VAS Index 107
5.3.3 Overhead Analysis for VAS Index 110
5.3.4 Evaluation for In-Storage Processing of CERBERUS 112
6 Conclusion 114
6.1 Lesson Learned from Studies 114
6.2 Concluding Remarks 115
References 118
