Ziegler, Tobias (2023)
Towards Scalable OLTP Over Fast Networks.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00024716
Ph.D. Thesis, Primary publication, Publisher's Version
Abstract
Online Transaction Processing (OLTP) underpins real-time data processing in many mission-critical applications, from banking to e-commerce. These applications typically issue short-duration, latency-sensitive transactions that demand immediate processing. High-volume applications, such as Alibaba's e-commerce platform, achieve peak transaction rates as high as 70 million transactions per second, exceeding the capacity of a single machine. Instead, distributed OLTP database management systems (DBMS) are deployed across multiple powerful machines. Historically, such distributed OLTP DBMSs have been primarily designed to avoid network communication, a paradigm largely unchanged since the 1980s.
However, fast networks challenge the conventional belief that network communication is the main bottleneck. In particular, emerging network technologies, like Remote Direct Memory Access (RDMA), radically alter how data can be accessed over a network. RDMA's primitives allow direct access to the memory of a remote machine within an order of magnitude of local memory access. This development invalidates the notion that network communication is the primary bottleneck. Given that traditional distributed database systems have been designed with the premise that the network is slow, they cannot efficiently exploit these fast network primitives, which requires us to reconsider how we design distributed OLTP systems.
This thesis focuses on the challenges RDMA presents and its implications on the design of distributed OLTP systems. First, we examine distributed architectures to understand data access patterns and scalability in modern OLTP systems. Drawing on these insights, we advocate a distributed storage engine optimized for high-speed networks. The storage engine serves as the foundation of a database, ensuring efficient data access through three central components: indexes, synchronization primitives, and buffer management (caching). With the introduction of RDMA, the landscape of data access has undergone a significant transformation. This requires a comprehensive redesign of the storage engine components to exploit the potential of RDMA and similar high-speed network technologies. Thus, as the second contribution, we design RDMA-optimized tree-based indexes — especially applicable for disaggregated databases to access remote data efficiently. We then turn our attention to the unique challenges of RDMA. One-sided RDMA, one of the network primitives introduced by RDMA, presents a performance advantage in enabling remote memory access while bypassing the remote CPU and the operating system. This allows the remote CPU to process transactions uninterrupted, with no requirement to be on hand for network communication. However, that way, specialized one-sided RDMA synchronization primitives are required since traditional CPU-driven primitives are bypassed. We found that existing RDMA one-sided synchronization schemes are unscalable or, even worse, fail to synchronize correctly, leading to hard-to-detect data corruption. As our third contribution, we address this issue by offering guidelines to build scalable and correct one-sided RDMA synchronization primitives. Finally, recognizing that maintaining all data in memory becomes economically unattractive, we propose a distributed buffer manager design that efficiently utilizes cost-effective NVMe flash storage. By leveraging low-latency RDMA messages, our buffer manager provides a transparent memory abstraction, accessing the aggregated DRAM and NVMe storage across nodes. Central to our approach is a distributed caching protocol that dynamically caches data. With this approach, our system can outperform RDMA-enabled in-memory distributed databases while managing larger-than-memory datasets efficiently.
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Erschienen: | 2023 | ||||
Creators: | Ziegler, Tobias | ||||
Type of entry: | Primary publication | ||||
Title: | Towards Scalable OLTP Over Fast Networks | ||||
Language: | English | ||||
Referees: | Binnig, Prof. Dr. Carsten ; Blanas, Prof. PhD Spyros | ||||
Date: | 24 October 2023 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xxiv, 305 Seiten | ||||
Refereed: | 22 August 2023 | ||||
DOI: | 10.26083/tuprints-00024716 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/24716 | ||||
Abstract: | Online Transaction Processing (OLTP) underpins real-time data processing in many mission-critical applications, from banking to e-commerce. These applications typically issue short-duration, latency-sensitive transactions that demand immediate processing. High-volume applications, such as Alibaba's e-commerce platform, achieve peak transaction rates as high as 70 million transactions per second, exceeding the capacity of a single machine. Instead, distributed OLTP database management systems (DBMS) are deployed across multiple powerful machines. Historically, such distributed OLTP DBMSs have been primarily designed to avoid network communication, a paradigm largely unchanged since the 1980s. However, fast networks challenge the conventional belief that network communication is the main bottleneck. In particular, emerging network technologies, like Remote Direct Memory Access (RDMA), radically alter how data can be accessed over a network. RDMA's primitives allow direct access to the memory of a remote machine within an order of magnitude of local memory access. This development invalidates the notion that network communication is the primary bottleneck. Given that traditional distributed database systems have been designed with the premise that the network is slow, they cannot efficiently exploit these fast network primitives, which requires us to reconsider how we design distributed OLTP systems. This thesis focuses on the challenges RDMA presents and its implications on the design of distributed OLTP systems. First, we examine distributed architectures to understand data access patterns and scalability in modern OLTP systems. Drawing on these insights, we advocate a distributed storage engine optimized for high-speed networks. The storage engine serves as the foundation of a database, ensuring efficient data access through three central components: indexes, synchronization primitives, and buffer management (caching). With the introduction of RDMA, the landscape of data access has undergone a significant transformation. This requires a comprehensive redesign of the storage engine components to exploit the potential of RDMA and similar high-speed network technologies. Thus, as the second contribution, we design RDMA-optimized tree-based indexes — especially applicable for disaggregated databases to access remote data efficiently. We then turn our attention to the unique challenges of RDMA. One-sided RDMA, one of the network primitives introduced by RDMA, presents a performance advantage in enabling remote memory access while bypassing the remote CPU and the operating system. This allows the remote CPU to process transactions uninterrupted, with no requirement to be on hand for network communication. However, that way, specialized one-sided RDMA synchronization primitives are required since traditional CPU-driven primitives are bypassed. We found that existing RDMA one-sided synchronization schemes are unscalable or, even worse, fail to synchronize correctly, leading to hard-to-detect data corruption. As our third contribution, we address this issue by offering guidelines to build scalable and correct one-sided RDMA synchronization primitives. Finally, recognizing that maintaining all data in memory becomes economically unattractive, we propose a distributed buffer manager design that efficiently utilizes cost-effective NVMe flash storage. By leveraging low-latency RDMA messages, our buffer manager provides a transparent memory abstraction, accessing the aggregated DRAM and NVMe storage across nodes. Central to our approach is a distributed caching protocol that dynamically caches data. With this approach, our system can outperform RDMA-enabled in-memory distributed databases while managing larger-than-memory datasets efficiently. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-247162 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science | ||||
Divisions: | 20 Department of Computer Science 20 Department of Computer Science > Data and AI Systems |
||||
Date Deposited: | 24 Oct 2023 11:35 | ||||
Last Modified: | 26 Oct 2023 07:41 | ||||
PPN: | |||||
Referees: | Binnig, Prof. Dr. Carsten ; Blanas, Prof. PhD Spyros | ||||
Refereed / Verteidigung / mdl. Prüfung: | 22 August 2023 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Send an inquiry |
Options (only for editors)
Show editorial Details |