#search #performance #eytzinger


A data structure for efficient lower-bound lookups

9 releases

Uses old Rust 2015

0.2.7 Feb 26, 2024
0.2.6 Jan 6, 2024
0.2.5 Sep 16, 2023
0.2.4 Aug 20, 2023
0.1.0 Oct 22, 2017

#283 in Data structures

Download history 3/week @ 2023-12-31 1/week @ 2024-01-07 22/week @ 2024-01-14 13/week @ 2024-01-21 2/week @ 2024-01-28 16/week @ 2024-02-04 2/week @ 2024-02-11 22/week @ 2024-02-18 224/week @ 2024-02-25 15/week @ 2024-03-03 42/week @ 2024-03-10 1/week @ 2024-03-17 55/week @ 2024-03-31

100 downloads per month


290 lines


Crates.io Documentation Build Status

This crate provides a data structure for approximate lookups in ordered collections.

More concretely, given a set A of n values, and a query value x, this library provides an efficient mechanism for finding the smallest value in A that is greater than or equal to x. In particular, this library caters to the important case where there are many such queries to the same array, A.

This library is constructed from the best solution identified in Array Layouts for Comparison-Based Searching by Paul-Virak Khuong and Pat Morin. For more information, see the paper, their website, and the C++ implementation repository.

Current implementation

At the time of writing, this implementation uses a branch-free search over an Eytzinger-arranged array with masked prefetching based on the C++ implementation written by the authors of the aforementioned paper. This is the recommended algorithm from the paper, and what the authors suggested in https://github.com/patmorin/arraylayout/issues/3#issuecomment-338472755.

Note that prefetching is only enabled with the (non-default) nightly feature due to https://github.com/aweinstock314/prefetch/issues/1. Suggestions for workarounds welcome.


The included benchmarks can be run with

$ cargo +nightly bench --features nightly

This will benchmark both construction and search with different number of values, and differently sized values -- look for the line that aligns closest with your data. The general trend is that ordsearch is faster when n is smaller and T is larger as long as you compile with target-cpu=native and lto=thin. The performance gain seems to be best on Intel processors, and is smaller since the (relatively) recent improvement to SliceExt::binary_search performance.

Below are summarized results from an Intel(R) Core(TM) i7-1068NG7 CPU @ 2.30GHz CPU run with:

$ rustc +nightly --version
rustc 1.75.0-nightly (e0d7ed1f4 2023-10-01)
$ env CARGO_INCREMENTAL=0 RUSTFLAGS='-C target-cpu=native' cargo +nightly bench --features nightly

Future work


~12K SLoC