Exploring the Design Space of a GPU-Aware Database Architecture

Abstract

Abstract. The vast amount of processing power and memory bandwidth provided by modern graphics cards make them an interesting platform for data-intensive applications. Unsurprisingly, the database research community has identified GPUs as effective co-processors for data processing several years ago. In the past years, there were many approaches to make use of GPUs at different levels of a database system. In this paper, we summarize the major findings of the literature on GPU-accelerated data processing. Based on this survey, we present key properties, important trade-offs and typical challenges of GPU-aware database architectures, and identify major open research questions.

Hybrid Execution of Queries to Analytical Databases

“…Использование графических процессоров для ускорения работы базы данных не ограничивается только исполнением отдельных операторов, например, исследовались подходы совместного использования с ЦПУ. Один из подходов [5] решает проблему высокой нагрузки на шину PCIe и недостаточности размера «быстрой» памяти графического процессора с помощью побитового разделения данных -горячие части данных (над которыми производятся вычислительно-емкие и независимые операции) хранятся в локальной памяти устройства, менее активно ис-пользуемые -в системной памяти. Каждый запрос обрабатывается в 2 стадии: сначала неточная фильтрация записей, затем промежуточный результат передается на ЦПУ и объединяется с остаточными данными в основной памяти.…”

GPU-Accelerated Database Systems: Survey and Open Challenges

“…Based on these estimates, a scheduler can allocate the cheapest processing device. 17 Since these models need to be able to estimate comparable operator runtimes across different devices, we and others [13] argue that dynamic cost models, which ap-Furthermore, a query should make use of multiple processing devices to speed up execution. Hence, the physical optimizer has to optimize hybrid CPU/GPU query plans, which significantly increases the optimization space.…”

“…Furthermore, as a substantial extension to a previous version of this paper [17], we conducted an in-depth literature survey of eight GPU-accelerated database management systems to validate and refine our theoretical discussions. This complements our findings in proposing a reference architecture.…”

“…We find that GDBMSs should be in-memory column stores, should use the blockat-a-time processing model and exploit all available processing devices for query Acknowledgements We thank Tobias Lauer from Jedox AG and the anonymous reviewers of the GPUs in Databases Workshop for their helpful feedback on the workshop version of this paper [17]. We thank Jens Teubner from TU Dortmund University, Michael Saecker from ParStream GmbH, and the anonymous reviewers of the TLDKS journal for their helpful comments on the journal version of this paper.…”