MoL-2024-20: Predicate Pushdown in FastLanes

MoL-2024-20: Duņamalijevs, Raufs (2024) Predicate Pushdown in FastLanes. [Report]

Text MoL-2024-20.text.pdf - Published Version Download (8MB)

Abstract

This project explores predicate evaluation for the FastLanes file format within the framework of cascaded encoding, which encodes the data in multiple layers to achieve higher compression ratios. Predicate pushdown is an optimisation technique that accelerates selective queries by applying filters directly within the data scanning process, thus reducing the volume of data entering the query execution pipeline. While predicate pushdown was widely studied for single-layer encodings, it was not examined in the context of cascaded encodings. Evaluating predicates on data with multi-layer cascades introduces new challenges, such as determining whether to fully decode the data before filtering or to apply filters on partially decoded data. Furthermore, previous works do not utilise data parallelism to its full extent, missing a critical opportunity for performance improvement of predicate evaluation. We address these gaps by developing a fast and portable predicate pushdown technique that functions across multiple architectures without requiring platform-specific implementations. We introduce a data-parallel method for predicate evaluation that stores results in a bitmap with a specific layout, enabling rapid evaluation across multiple columns. In addition, we demonstrate that certain encodings can be partially decompressed and evaluated, avoiding full decompression and increasing efficiency.

Item Type:	Report
Report Nr:	MoL-2024-20
Series Name:	Master of Logic Thesis (MoL) Series
Year:	2024
Subjects:	Computation Logic
Depositing User:	Dr Marco Vervoort
Date Deposited:	23 Dec 2024 13:37
Last Modified:	23 Dec 2024 13:37
URI:	https://eprints.illc.uva.nl/id/eprint/2348

Actions (login required)

View Item