Bang, Tiemo (2022)
Adaptive Architectures for Robust Database Management Systems.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00021383
Ph.D. Thesis, Primary publication, Publisher's Version
Text
(Dissertation)
dissertation_Tiemo_Bang_2022.pdf Copyright Information: CC BY 4.0 International - Creative Commons, Attribution. Download (4MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Adaptive Architectures for Robust Database Management Systems | ||||
Language: | English | ||||
Referees: | Binnig, Prof. Dr. Carsten ; Hellerstein, Prof. Joseph | ||||
Date: | 2022 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xvi, 218 Seiten | ||||
Date of oral examination: | 26 July 2022 | ||||
DOI: | 10.26083/tuprints-00021383 | ||||
Abstract: | Ever since, the workload and hardware conditions for Database Management Systems (DBMSs) are expanding through new use cases and hardware. Starting from the first transactional DBMSs supporting the moon landing, today’s DBMSs process billions of sales transactions for online commerce on a single day and process many further workloads like reporting or fraud detection. Similarly, for the early DBMSs only few hardware platforms were available, while today’s DBMSs face a host of diverse hardware platforms. Indeed, today a single DBMS is exposed to changing and load-fluctuating workloads, e.g., depending on the popularity of sales items, and is operated on changing hardware. However, serving changing workloads and supporting diverse hardware platforms is non-trivial, as for the best performance DBMS designs must be specialized. For flexibly specializing DBMS components to changing workload or hardware conditions, adaptation approaches have been proposed, e.g., adaptive query execution. Whereas the DBMS architecture that deploys the components is manually specialized and statically implemented at design-time. While in essence all DBMS architectures determine which components are executed together on what resource partitions, today there exist only few static architectures that largely predetermine this at design-time for specific conditions, e.g., for multi-processor hardware the NUMA-aware architecture dictates a resource partitioning per processor for all components. Besides high re-implementation effort for adjusting these static architectures, this approach also inadequately simplifies architectures with coarse-grained specialization for many components at once, neglecting the distinct workload and hardware effects on individual DBMS components and their contained functions. Hence, these static DBMS architectures severely degrade DBMS performance, when unfit. This dissertation pursues the adaptation of DBMS architectures. For high and robust performance under changing workload and hardware, the static specialization at design-time is progressed to the flexible and precise adaptation of the architecture when deploying the DBMS or even at runtime. The approach is an initial evaluation of DBMSs with static architectures. Then, general concepts for the adaptation of DBMS architectures are proposed, based on which adaptive architectures for the classes of single-server and multi-server DBMSs are realized. The overall idea for the adaptation of DBMS architectures is to flexibly compose fine-grained building blocks of the DBMS to a best-fit architecture, i.e., adapting at the granularity of distinct functions of DBMS components without requiring any re-implementation. Besides the effortless adjustment of the architecture, this dissertation proposes concepts with an emphasis on fine-granular and separate adaptation for distinct DBMS functions, such that optimizers can derive architectures best-fit for the specific conditions and functions at hand. By constructing a navigable optimization space for architectures of single- and multi-server DBMSs, the proposed concepts not only enable the flexible mimicking of any existing architecture, but importantly enable the creation of entirely new architectures. The key findings are that both the realized adaptive single-server and the adaptive multi-server architecture prove effective and efficient, for adapting to the conditions considered in this dissertation. Under changing transactional and mixed workloads, the proposed adaptive architectures generally perform at least on par with the individually best state-of-the-art architecture. Indeed, when adopting novel better-fit architectures, all existing architectures are outperformed, e.g., with resource assigned at a granularity unlike any of today’s single-server architectures or when separately specializing for distinct queries of mixed workloads rather than compromising as today’s multi-server architectures. That is, the proposed flexible and precise adaptation demonstrates higher and more robust performance. While our findings exhibit novel better-fit architectures only for a subset of possible workload and hardware conditions, this dissertation overall indicates high potential for adapting architectures with the proposed concepts. As the proposed concepts make a vast optimization space generally navigable, optimizers will be able to adapt DBMS architectures flexibly and more precisely to many workloads and hardware. Instead of fragile static architectures, the proposed adaptive architectures thus provide the necessary foundation for DBMSs to achieve high and robust performance under changing workload and hardware. |
||||
Alternative Abstract: |
|
||||
Uncontrolled Keywords: | DBMS, architectures, scale-up, scale-out, adaptive, evaluation | ||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-213831 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science | ||||
Divisions: | 20 Department of Computer Science > Data Management (2022 umbenannt in Data and AI Systems) | ||||
TU-Projects: | SAP|SOWNO.DE-2018-31|CON-00705: HSCD-0068 | ||||
Date Deposited: | 03 Aug 2022 12:31 | ||||
Last Modified: | 16 Dec 2022 16:17 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/21383 | ||||
PPN: | 499051114 | ||||
Export: |
View Item |