On-The-Fly Data Distribution to Accelerate Query Processing in Heterogeneous Memory Systems

Research output: Contribution to book/conference proceedings/anthology/reportConference contributionContributedpeer-review

Contributors

Abstract

To meet response time and throughput demands, data processing architects continuously adapt query processing systems to novel hardware features. For instance, data processing systems already shifted from disk-oriented to main memory-oriented architectures to efficiently exploit the ever-increasing capacities of main memory. A prominent example for such new developments are emerging memory technologies such as very large caches, high-bandwidth memory (HBM), non-uniform memory access (NUMA) or remote-memory designs like CXL. These memories complement regular DRAM by trading off between properties such as capacity, read/write throughput or access latency. However, these degrees of freedom and their inherent complexity make it difficult for database systems to profit from the new hardware. Taking HBM – as integrated in the 2023 Intel “Sapphire Rapids” Xeon Max processors – as a most recent example for emerging memory technologies, we present results from microbenchmarks that demonstrate different worker-thread saturation patterns for DRAM and HBM in this paper. Then, we subsequently derive characteristics and a cost model to make dynamic on-the-fly data-distribution/movement decisions. Based on this model, we show that for data processing queries with a specific data-reuse pattern, dynamic data (re-)distribution from DRAM to HBM can decrease end-to-end query run times.

Details

Original languageEnglish
Title of host publicationAdvances in Databases and Information Systems
PublisherSpringer Nature
Pages170-183
Number of pages14
ISBN (electronic)978-3-031-70626-4
ISBN (print)978-3-031-70628-8
Publication statusPublished - 1 Sept 2024
Peer-reviewedYes

Publication series

SeriesLecture Notes in Computer Science
Volume14918
ISSN0302-9743

External IDs

ORCID /0000-0002-1427-9343/work/167216831
ORCID /0000-0001-8107-2775/work/167216949

Keywords