This repository is a sub-project of ICU4X project and contains a harness and implementations of small benchmarks for measuring performance, memory and size differences between various implementations of ICU.
The main goal of this project is to establish a reasonable framework for comparing those measurements across multiple implementations and validate a claim that ICU4X is not slower, doesn't use more memory and doesn't take more space than a comparable alternatives.
At the moment the harness is focused on measuring performance between:
- ICU4X 0.1
- ICU4C 67 (using C++ apps)
- ICU4C 67 (via
rust_icu) - ICU4C 67 (via FFI)
- A set of earlier Rust implementations of components such as
intl_pluralrules,unic_locale, andunic_datetime
The aim is to make it easy to compare different backends for similar basic operations and minimize noise in data.
The tests can be launched via stand-alone binaries, or in case of all Rust apps via criterion benchmark.
Specs:
-
Dell Tower 5820
- Intel(R) Xeon(R) W-2155 CPU @ 3.30GHz x20
- 32 GB RAM
-
C++: GCC 10
-
Rust: 1.47
-
ICU4C: 67
-
ICU4X: icu4x 0.1
Sample: 956 locale strings provided to MozLocale constructor during fresh-profile startup of Firefox Nightly on Oct 21st 2019.
| Test | ICU4X | ICU4C (rust_icu) | ICU4X (FFI) | ICU4C (C++) | Unic |
|---|---|---|---|---|---|
| Locale | |||||
| Construct an instance from a string | 28,531 ns | 821,947 ns | 541,333 ns | 1,235,646 ns | 41,578 ns |
Filter all locales against en-US |
3,035 ns | 5,539 ns | 10,013 ns | 48,195 ns | 3,067 ns |
| Serializing all locales to a string | 68,072 ns | 77,826 ns | 66,682 ns | 2,982,821 ns | 75,760 ns |
| Canonicalize all strings | 88,416 ns | 942,376 ns | 217,938 ns | 2,010,337 ns | 115,324 ns |
| Measuring memory allocation of all instances | 30,592 b | 22,944 b | 91,776 b | 229,376 b | 30,592 b |
| Plural Rules | |||||
| Select | 92 ns | 289 ns | 287 ns | 123,822 ns | 5 ns |
| Date Time | |||||
| Format | 1,764 ns | 8,251 ns |