Relational database systems provide various services and applications with an efficient means for storing, processing, and retrieving their data. The performance of these systems has a direct impact on the quality of service of the applications that rely on them. Therefore, it is crucial that database systems are able to adapt and grow in tandem with the demands of these applications, ensuring that their performance scales accordingly. In the past, Moore's law and algorithmic advancements have been sufficient to meet these demands. However, with the slowdown of Moore's law, researchers have begun exploring alternative methods, such as application-specific technologies, to satisfy the more challenging performance requirements. One such technology is field-programmable gate arrays (FPGAs), which provide ideal platforms for developing and running custom architectures for accelerating database systems.
The goal of this thesis is to develop a domain-specific architecture that can enhance the performance of in-memory database systems when executing analytical queries. Our research is guided by a combination of academic and industrial requirements that seek to strike a balance between generality and performance. The former ensures that our platform can be used to process a diverse range of workloads, while the latter makes it an attractive solution for high-performance use cases.
Throughout this thesis, we present the development of a system-on-chip for database system acceleration that meets our requirements. The resulting architecture, called CbMSMK, is capable of processing the projection, sort, aggregation, and equi-join database operators and can also run some complex TPC-H queries. CbMSMK employs a shared sort-merge pipeline for executing all these operators, which results in an efficient use of FPGA resources. This approach enables the instantiation of multiple acceleration cores on the FPGA, allowing it to serve multiple clients simultaneously. CbMSMK can process both arbitrarily deep and wide tables efficiently. The former is achieved through the use of the sort-merge algorithm which utilizes the FPGA RAM for buffering intermediate sort results. The latter is achieved through the use of KeRRaS, a novel variant of the forward radix sort algorithm introduced in this thesis. KeRRaS allows CbMSMK to process a table a few columns at a time, incrementally generating the final result through multiple iterations. Given that acceleration is a key objective of our work, CbMSMK benefits from many performance optimizations. For instance, multi-way merging is employed to reduce the number of merge passes required for the execution of the sort-merge algorithm, thus improving the performance of all our pipeline-breaking operators. Another example is our in-depth analysis of early aggregation, which led to the development of a novel cache-based algorithm that significantly enhances aggregation performance. Our experiments demonstrate that CbMSMK performs on average 5 times faster than the state-of-the-art CPU-based database management system MonetDB.:I Database Systems & FPGAs
1 INTRODUCTION
1.1 Databases & the Importance of Performance
1.2 Accelerators & FPGAs
1.3 Requirements
1.4 Outline & Summary of Contributions
2 BACKGROUND ON DATABASE SYSTEMS
2.1 Databases
2.1.1 Storage Model
2.1.2 Storage Medium
2.2 Database Operators
2.2.1 Projection
2.2.2 Filter
2.2.3 Sort
2.2.4 Aggregation
2.2.5 Join
2.2.6 Operator Classification
2.3 Database Queries
2.4 Impact of Acceleration
3 BACKGROUND ON FPGAS
3.1 FPGA
3.1.1 Logic Element
3.1.2 Block RAM (BRAM)
3.1.3 Digital Signal Processor (DSP)
3.1.4 IO Element
3.1.5 Programmable Interconnect
3.2 FPGADesignFlow
3.2.1 Specifications
3.2.2 RTL Description
3.2.3 Verification
3.2.4 Synthesis, Mapping, Placement, and Routing
3.2.5 TimingAnalysis
3.2.6 Bitstream Generation and FPGA Programming
3.3 Implementation Quality Metrics
3.4 FPGA Cards
3.5 Benefits of Using FPGAs
3.6 Challenges of Using FPGAs
4 RELATED WORK
4.1 Summary of Related Work
4.2 Platform Type
4.2.1 Accelerator Card
4.2.2 Coprocessor
4.2.3 Smart Storage
4.2.4 Network Processor
4.3 Implementation
4.3.1 Loop-based implementation
4.3.2 Sort-based Implementation
4.3.3 Hash-based Implementation
4.3.4 Mixed Implementation
4.4 A Note on Quantitative Performance Comparisons
II Cache-Based Morphing Sort-Merge with KeRRaS (CbMSMK)
5 OBJECTIVES AND ARCHITECTURE OVERVIEW
5.1 From Requirements to Objectives
5.2 Architecture Overview
5.3 Outlineof Part II
6 COMPARATIVE ANALYSIS OF OPENCL AND RTL FOR SORT-MERGE PRIMITIVES ON FPGAS
6.1 Programming FPGAs
6.2 RelatedWork
6.3 Architecture
6.3.1 Global Architecture
6.3.2 Sorter Architecture
6.3.3 Merger Architecture
6.3.4 Scalability and Resource Adaptability
6.4 Experiments
6.4.1 OpenCL Sort-Merge Implementation
6.4.2 RTLSorters
6.4.3 RTLMergers
6.4.4 Hybrid OpenCL-RTL Sort-Merge Implementation
6.5 Summary & Discussion
7 RESOURCE-EFFICIENT ACCELERATION OF PIPELINE-BREAKING DATABASE OPERATORS ON FPGAS
7.1 The Case for Resource Efficiency
7.2 Related Work
7.3 Architecture
7.3.1 Sorters
7.3.2 Sort-Network
7.3.3 X:Y Mergers
7.3.4 Merge-Network
7.3.5 Join Materialiser (JoinMat)
7.4 Experiments
7.4.1 Experimental Setup
7.4.2 Implementation Description & Tuning
7.4.3 Sort Benchmarks
7.4.4 Aggregation Benchmarks
7.4.5 Join Benchmarks
7. Summary
8 KERRAS: COLUMN-ORIENTED WIDE TABLE PROCESSING ON FPGAS
8.1 The Scope of Database System Accelerators
8.2 Related Work
8.3 Key-Reduce Radix Sort(KeRRaS)
8.3.1 Time Complexity
8.3.2 Space Complexity (Memory Utilization)
8.3.3 Discussion and Optimizations
8.4 Architecture
8.4.1 MSM
8.4.2 MSMK: Extending MSM with KeRRaS
8.4.3 Payload, Aggregation and Join Processing
8.4.4 Limitations
8.5 Experiments
8.5.1 Experimental Setup
8.5.2 Datasets
8.5.3 MSMK vs. MSM
8.5.4 Payload-Less Benchmarks
8.5.5 Payload-Based Benchmarks
8.5.6 Flexibility
8.6 Summary
9 A STUDY OF EARLY AGGREGATION IN DATABASE QUERY PROCESSING ON FPGAS
9.1 Early Aggregation
9.2 Background & Related Work
9.2.1 Sort-Based Early Aggregation
9.2.2 Cache-Based Early Aggregation
9.3 Simulations
9.3.1 Datasets
9.3.2 Metrics
9.3.3 Sort-Based Versus Cache-Based Early Aggregation
9.3.4 Comparison of Set-Associative Caches
9.3.5 Comparison of Cache Structures
9.3.6 Comparison of Replacement Policies
9.3.7 Cache Selection Methodology
9.4 Cache System Architecture
9.4.1 Window Aggregator
9.4.2 Compressor & Hasher
9.4.3 Collision Detector
9.4.4 Collision Resolver
9.4.5 Cache
9.5 Experiments
9.5.1 Experimental Setup
9.5.2 Resource Utilization and Parameter Tuning
9.5.3 Datasets
9.5.4 Benchmarks on Synthetic Data
9.5.5 Benchmarks on Real Data
9.6 Summary
10 THE FULL PICTURE
10.1 System Architecture
10.2 Benchmarks
10.3 Meeting the Objectives
III Conclusion
11 SUMMARY AND OUTLOOK ON FUTURE RESEARCH
11.1 Summary
11.2 Future Work
BIBLIOGRAPHY
LIST OF FIGURES
LIST OF TABLES
| Identifer | oai:union.ndltd.org:DRESDEN/oai:qucosa:de:qucosa:85607 |
| Date | 30 May 2023 |
| Creators | Moghaddamfar, Mehdi |
| Contributors | Lehner, Wolfgang, Kumar, Akash, Technische Universität Dresden |
| Source Sets | Hochschulschriftenserver (HSSS) der SLUB Dresden |
| Language | English |
| Detected Language | English |
| Type | info:eu-repo/semantics/publishedVersion, doc-type:doctoralThesis, info:eu-repo/semantics/doctoralThesis, doc-type:Text |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.003 seconds