|

Overview

park-bench is a quick-and-dirty benchmarking tool for comparing the performance of Haskell functions. Specifically, it is designed to optimize the workflow in which a programmer makes a small change to a function and wants to measure its performance impact with as little friction as possible.

Configuration

Environment variable name	Type	Meaning	Default value
PARK_BENCH_RUNLEN	Float	The target number of seconds that each benchmark run takes	0.1

Example usage

Say I am interested in improving the performance of fib, which is a function defined in module MyMathUtilities.Fib in a local package called my-math-utilities.

Step 1: Write the function you'd like to benchmark

First, I'm going to copy the implementation of fib to a new top-level definition called fastfib, tweak its implementation, and export both from module MyMathUtilities.Fib.

If fib was private before, that's ok. We only need to expose it for as long as we are interested in benchmarking.

module MyMathUtilities.Fib (fib, fastfib, ...) where

Step 2: Write a standalone bench/Main.hs module with a main function

Next, I'm going to write a standalone Main.hs in a subdirectory called bench, which will be compiled to an executable that runs my benchmark.

module Main where

-- The module in my local package that I want to benchmark
import MyMathUtilities.Fib

-- This library
import ParkBench

main :: IO ()
main =
  benchmark
    [ function "fib" fib 20
    , function "fastfib" fastfib 20
    ]

Step 3: Define an executable component

Next, I'm going to define an executable component for my benchmark in my my-math-utilities.cabal file.

executable bench
  build-depends:
    base,
    -- The local package that I want to benchmark
    my-math-utilities,
    -- This library
    park-bench
  default-language: Haskell2010
  ghc-options: -O -rtsopts -with-rtsopts=-T
  hs-source-dirs: bench
  main-is: Main.hs

I need to compile the benchmark with -rtsopts -with-rtsopts=-T, otherwise my benchmark will not be able to get RTS statistics from GHC at runtime.

Alternatively, I could compile the benchmark with only -rtsopts, but then I'll have to provide +RTS -T to the executable later.

Step 4: Run the benchmark

If all goes well, I'll have an executable component to run.

cabal run my-math-utilities:exe:bench

stack run my-math-utilities:exe:bench

Or, if I only compiled with -rtsopts, but not -with-rtsopts=-T,

cabal run my-math-utilities:exe:bench -- +RTS -T

stack run my-math-utilities:exe:bench -- +RTS -T

Step 5: Clean up

After benchmarking, I can choose to keep the benchmark (and associated executable component) around, but I'll probably delete them instead. I've learned something, collected some sweet screenshots for my PR, and I'm ready to move on.

Caveat emptor

The statistical analysis performed by park-bench is simplistic, written by a novice, and may have bugs. Results should not necessarily be trusted; please use (or at least compare to) a different tool.