Vectorizing Database Column Scans with Complex Predicates

Abstract

The performance of the full table scan is critical for the overall performance of column-store database systems such as the SAP HANA database. Compressing the underlying column data format is both an advantage and a challenge, because it reduces the data volume involved in a scan on one hand and introduces the need for decompression during the scan on the other hand. In previous work [26] we have shown how to accelerate the column-scan with range predicates using SIMD instructions. In this paper, we present a framework for vectorized scans with more complex predicates. One important building block is the In-List predicate, where all rows whose values are contained in a given list of values are selected. While this seems to exhibit only little data parallelism on first sight, we show that a performant vectorized implementation is possible using the new Intel AVX2 instruction set. We also improve our previous algorithms by leveraging the increased vector-width. Finally in a detailed performance evaluation, we show the benefit of these optimizations and of the new instruction set: in almost all cases our scans needs less than one CPU cycle per row including scans with In-List predicate, leading to an overall throughput of 8 billion rows per second and more on a single core.

View Slides

Extracted Key Phrases

01020201520162017
Citations per Year

Citation Velocity: 11

Averaging 11 citations per year over the last 3 years.

Learn more about how we calculate this metric in our FAQ.

Cite this paper

@inproceedings{Willhalm2013VectorizingDC, title={Vectorizing Database Column Scans with Complex Predicates}, author={Thomas Willhalm and Ismail Oukid and Ingo M{\"{u}ller and Franz F{\"a}rber}, booktitle={ADMS@VLDB}, year={2013} }