Database Architectures for Modern Hardware

Summary

Over the last years, the social and commercial relevance of efficient data management has led to the development of database systems as foundation of almost all complex software systems. Hence there is a wide acceptance of architectural patterns for database systems which are based on assumptions on classic hardware setups. However, the currently used database concepts and systems are not well prepared to support emerging application domains such as eSciences, Internet of Things or Digital Humanities. From a user's perspective, flexible domain-specific query languages or at least access interfaces are required, novel data models for these application domains have to be integrated, and consistency guarantees which reduce flexibility and performance should be adaptable according to the requirements. Finally, volume, variety, veracity as well as velocity of data caused by ubiquitous sensors have to be mastered by massive scalability and online processing by providing traditional qualities of database systems like consistency, isolation and descriptive query languages. At the same time, current and future hardware trends provide new opportunities such as:

• many-core CPUs: Next-generation CPUs will provide hundreds of compute cores already in the commodity range. In order to allow high degrees of parallelism some architectures already provide hardware support for the necessary synchronization, e.g. transactional memory. However, it is not clear yet how to fully utilize these degrees of parallelism and synchronization mechanism for database processing.
• co-processors like GPU and FPGA: Special-purpose computing units such as GPUs and FPGAs allow for parallelism at much higher degrees accelerating compute-intensive tasks significantly. Moreover, heterogeneous hardware designs such as coupled CPU-FPGA and CPU-GPU architectures represent a trend of close integration between classic hardware and emerging hardware. However, such designs require new architectural concepts for data management.
• novel storage technologies like NVRAM and SSD: Even modern in-memory database system solutions rely mostly on block-based media (e.g. SSD and HDD) for ensuring persistence of data. Emerging memory technologies such as non-volatile memory (NVRAM) promise byte-addressable persistence with latencies close to DRAM. Currently, the usage of this technology is discussed for instant failure recovery of databases, but the role of NVRAM in future data management system architectures is still open.
• high-speed networks: Both in scale-up and scale-out scenarios efficient interconnects play a crucial role. Today, high-speed networks based on 10 Gbit/s Ethernet or InfiniBand support already Remote DMA, i.e. direct access to memory of a remote node. However, this requires to deal with distributed systems properties (unreliability, locality) and it is still unclear how database systems can utilize this mechanism.

In order to open up the exemplarily mentioned application domains together with exploiting the potential of future hardware generations it becomes necessary now to fundamentally rethink current database architectures.

One of the main challenges of this rethinking is that it requires expertise from different research disciplines: hardware design, computer architectures, networking, operating systems, distributed systems, software engineering, and database systems.

Thus, the goal of this Dagstuhl Seminar was to bring together researchers and practitioners from these areas representing both the software and hardware sides and therefore different disciplines to foster cross-cutting architectural discussions. In this way, the seminar extended the series of previous Dagstuhl seminars on database systems aspects, such as "Robust Query Processing" (10381, 12321, 17222) as well as ``Databases on Future Hardware'' (17101).

The seminar was organized into six working groups where the participants discussed opportunities and challenges in order to exploit different features of modern hardware and operating system primitives for data processing:

• Database accelerators: Based on an analysis of use cases for database accelerators from the level of individual operators and algorithms up to the level of complex database tasks, the group discussed ways of exploiting and evaluating accelerator technologies as well as future research directions with respect to hardware acceleration in databases.
• Memory hierarchies: The group discussed design recipes for database nodes with non-trival memory hierarchies containing not only disk and RAM but also non-volatile memory. Within such a hierarchy different caching strategies are employed: exclusive caching for functionally equivalent levels and inclusive caching for levels with different functionality.
• Remote direct memory access: The group discussed ways of exploiting RDMA in data-intensive applications. Particularly, an interface providing a set of useful abstractions for network-aware data-intensive processing called DPI was proposed. Similar to MPI, DPI is designed as an interface that can have multiple implementations for different networking technologies to enable the exploitation of RDMA and in-network processing.
• Heterogeneous database architectures: This topic was addressed by two working groups. Both groups discussed a database software architecture that is capable of making use of multiple hardware devices (GPU, TPU, FPGA, ASICs), in addition to the CPU for handling database workloads. The principle goal was an architecture that would never be worse than a state-of-the-art CPU-centered database architecture, but would get significant benefit on those workloads were the heterogeneous devices can exploit their strengths. The first group developed a morsel-driven architecture, where pipelines are broken up into sub-pipelines and adaptive execution strategies are exploited. The second group discussed operating system support and primitives for heterogeneous architectures.
• Machine learning in database systems: The goal of this working group was to investigate the application of machine learning methods for estimating operator selectivities as part of query optimization. Such an approach could overcome the inaccuracies of traditional cost estimation techniques especially for queries comprised of complex predicates and multiple joins.

The progress and outcome of the individual working groups was presented in a daily plenary session, details of the results are given below.

References

1. Gustavo Alonso, Michaela Blott, Jens Teubner: Databases on Future Hardware (Dagstuhl Seminar 17101). Dagstuhl Reports 7(3):1–18 (2017)
2. Renata Borovica-Gajic, Goetz Graefe, Allison Lee: Robust Performance in Database Query Processing (Dagstuhl Seminar 17222). Dagstuhl Reports 7(5):169–180 (2017)
3. Goetz Graefe, Wey Guy, Harumi A. Kuno, Glenn N. Paulley: Robust Query Processing (Dagstuhl Seminar 12321). Dagstuhl Reports 2(8):1–15 (2012)
4. Goetz Graefe, Arnd Christian König, Harumi Anne Kuno, Volker Markl, Kai-Uwe Sattler: Robust Query Processing. Dagstuhl Seminar Proceedings 10381, Schloss Dagstuhl – Leibniz- Zentrum für Informatik, Germany 2010