Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Is Binary Search Really All You Need? Supercharging Lightweight Database Indexing on GPUs

Justus Henneberg, Felix Schuhknecht

TL;DR

This work questions the dominance of complex GPU indexes by revisiting binary search on a sorted dense array as a strong baseline for read-only workloads. It introduces five lightweight indexing schemes (AO+BS, EO+BS, AO+PB+BS, EO+KS, AO+PB+KS) and a suite of low-level GPU optimizations, showing how to push these minimal structures toward state-of-the-art performance. Through extensive experiments on an RTX 4090, the authors demonstrate that carefully tuned EO+KS and AO+PB+KS configurations can outperform or closely match B+-tree and hash-table baselines, while maintaining far smaller memory footprints and faster rebuilds. The results imply that minimalist, GPU-tailored indexing can be a practical alternative for read-only or read-mostly workloads, particularly when memory is constrained or range lookups are important.

Abstract

Performing binary search on a sorted dense array is a widely used baseline when benchmarking sophisticated index structures, as it is simple to implement and exhibits a low construction time. However, the popular opinion is that such a simple approach cannot compete with highly-optimized GPU index structures in terms of lookup performance, and hence, should not actually be considered in practice. Interestingly, in our recent works on GPU indexing, we observed a surprisingly good performance of binary search in a variety of situations. Since binary search requires nothing but a sorted array to operate on, which makes it very attractive in the presence of scarce GPU memory, the question arises whether binary search and related variants of it can be made truly competitive and actually replace state-of-the-art index structures, such as a GPU-resident B-Tree and two different hash tables, in read-only scenarios. To find out, as a starting point, we consider five variants of lightweight GPU indexing schemes that offer a minimal or close to minimal memory footprint and analyze how far they are still behind the sophisticated index structures. Step by step, we then "supercharge" them with a set of carefully designed low-level optimizations to incrementally reveal their true potential and the best overall scheme and configuration for answering point lookups and range lookups. Our experimental evaluation reveals that the best optimized lightweight indexes are not only competitive to the sophisticated baselines, but actually manage to outperform them partially while offering a significantly lower memory footprint.

Is Binary Search Really All You Need? Supercharging Lightweight Database Indexing on GPUs

TL;DR

This work questions the dominance of complex GPU indexes by revisiting binary search on a sorted dense array as a strong baseline for read-only workloads. It introduces five lightweight indexing schemes (AO+BS, EO+BS, AO+PB+BS, EO+KS, AO+PB+KS) and a suite of low-level GPU optimizations, showing how to push these minimal structures toward state-of-the-art performance. Through extensive experiments on an RTX 4090, the authors demonstrate that carefully tuned EO+KS and AO+PB+KS configurations can outperform or closely match B+-tree and hash-table baselines, while maintaining far smaller memory footprints and faster rebuilds. The results imply that minimalist, GPU-tailored indexing can be a practical alternative for read-only or read-mostly workloads, particularly when memory is constrained or range lookups are important.

Abstract

Performing binary search on a sorted dense array is a widely used baseline when benchmarking sophisticated index structures, as it is simple to implement and exhibits a low construction time. However, the popular opinion is that such a simple approach cannot compete with highly-optimized GPU index structures in terms of lookup performance, and hence, should not actually be considered in practice. Interestingly, in our recent works on GPU indexing, we observed a surprisingly good performance of binary search in a variety of situations. Since binary search requires nothing but a sorted array to operate on, which makes it very attractive in the presence of scarce GPU memory, the question arises whether binary search and related variants of it can be made truly competitive and actually replace state-of-the-art index structures, such as a GPU-resident B-Tree and two different hash tables, in read-only scenarios. To find out, as a starting point, we consider five variants of lightweight GPU indexing schemes that offer a minimal or close to minimal memory footprint and analyze how far they are still behind the sophisticated index structures. Step by step, we then "supercharge" them with a set of carefully designed low-level optimizations to incrementally reveal their true potential and the best overall scheme and configuration for answering point lookups and range lookups. Our experimental evaluation reveals that the best optimized lightweight indexes are not only competitive to the sophisticated baselines, but actually manage to outperform them partially while offering a significantly lower memory footprint.

Paper Structure

This paper contains 30 sections, 4 equations, 22 figures.

Table of Contents

  1. Introduction
  2. Hardware Architecture and Background
  3. Lightweight Indexing Methods
  4. Ascending Key Order + Binary Search (AO+BS)
  5. Eytzinger Order + Binary Search (EO+BS)
  6. Ascending Key Order + Pivot Buffer + Binary Search (AO+PB+BS)
  7. Generalizing Binary Search to $K$-ary Search
  8. Building the Structures
  9. Optimizations
  10. Limiting Concurrency
  11. Cache Pinning Strategies
  12. Steps Pinning
  13. Full Pinning
  14. Reordering Strategies
  15. Lookup Reordering
  16. ...and 15 more sections

Figures (22)

  • Figure 1: Situations in which binary search (AO+BS) outperforms its traditional competitors.
  • Figure 2: Architecture of the NVIDIA RTX 4090.
  • Figure 3: Search paths of AO+BS.
  • Figure 4: Performing a point lookup for key 9 using EO+BS (bottom) and its equivalent in an ascending array (top).
  • Figure 5: Performing the range lookup [9,12] using EO+BS (bottom) and its equivalent in ascending order (top). Shaded orange indices mark the qualifying entries, while red indices mark entries that do not qualify, but still have to be scanned.
  • ...and 17 more figures