Idris2/benchmark/readme.md
2021-01-12 13:33:57 +00:00

33 lines
1.9 KiB
Markdown

# UNIX Benchmark suite for Idris2
Dependencies:
- A working Idris2 installation
- GNU time command (at usr/bin/time)
## Motivation
We were aiming to optimise the reference counting backend for Idris2, but it quickly became apparent that this would be difficult without a consistent performance metric. This is the product of that issue, a simple benchmarking shell script. We aim for this to be usable across all of the code generators, though so far it has only been tested on the chez and refc (still in development) backends.
## Structure
The tests provided should serve as an example of the structure - here's the general idea.
Each test should be contained within a root folder TEST inside the `benchmarks` directory. A benchmark consists of a `TEST.idr` file, along with a `TEST.in` *and* a `TEST_fast.in`.
The tests are structured such that a value for the limiting factor on the duration is read from stdin from the idris code. Suitable values should be given in the `TEST.in` and `TEST_fast.in` files, which cause the test to take roughly 5 seconds and 1 second respectively (on the chez backend). This might be paths to input files of differing size/complexity, or it may simply be a number of repeats. There's a timeout currently set at 15 seconds which will end the test with signal 2. The aim is that after optimisation each backend should be able to complete each test within that period.
If you've got any Idris2 code that you think could be adapted to make a good benchmark, please do contribute!
## Using the script
Usage: bench.sh [OPTION]...
-b build each test first
-c specify the backend to use during compilation (requires -b flag)
must be one of chez, refc, racket, gambit
(default is chez)
-f perform a faster version of the benchmark
-o specify name of output csv file
(default is `bench-results_$BACKEND[_fast]`)
-h shows this message
e.g. `./bench.sh -fbc refc -o myrefcoutput`