benchmark-interface

Benchmark-Interface

Introduction

Benchmark-Interface is one Ruby benchmarking interface to rule them all. It allows you to run benchmarks written for one Ruby benchmarking system using some other Ruby benchmarking system.

For example, Benchmark-Interface lets you take a benchmark from MRI's suite and run it using benchmark-ips, or take a benchmark written using Bench9000 and run it using bmbm.

Benchmark-Interface also provides a new format for writing benchmarks, which is designed to be the simplest of all, making it easy to write lots of new benchmarks and run them any way you like.

$ gem install benchmark-interface

Benchmark-Interface doesn't install any backends (other benchmark tools) by default, so you probably also want to do:

$ gem install benchmark-ips

Examples

Running an MRI benchmark using benchmark-ips

$ benchmark ruby/benchmark/bm_vm1_length.rb

These are long benchmarks - we're increasing warmup and sample time
Warming up --------------------------------------
       bm_vm1_length     1.000  i/100ms
       bm_vm1_length     1.000  i/100ms
       bm_vm1_length     1.000  i/100ms
Calculating -------------------------------------
       bm_vm1_length      0.955  (± 0.0%) i/s -     10.000  in  10.472341s
       bm_vm1_length      0.960  (± 0.0%) i/s -     10.000  in  10.433246s
       bm_vm1_length      0.975  (± 0.0%) i/s -     10.000  in  10.260680s

Running a bench9000 benchmark using bmbm

$ benchmark bench9000/benchmarks/classic/mandelbrot.rb --bmbm

                 user     system      total        real
mandelbrot   1.760000   0.030000   1.790000 (  1.804423)

The Benchmark-Interface Interface

Benchmark-Interface provides its own new format for writing benchmarks. It's very simple.

benchmark { 14 * 14 * 14 }
benchmark { 14 ** 3 }

You can give benchmarks names, and you usually should, but they will be named with the file name and a counter if you don't.

benchmark('mul') { 14 * 14 * 14 }
benchmark('pow') { 14 ** 3 }

You just write that in the file. You don't need to require anything. We talk about all the benchmarks in a file being the 'benchmark set'.

If benchmark is for some reason overloaded, you can also use Benchmark-Interface.benchmark.

Frontends

If you already have benchmarks written for a different system you can run those with Benchmark-Interface.

Supported frontends are:

• Benchmark-Interface
• MRI's benchmarks
• The benchmark standard library (measure, bm, bmbm and so on)
• RBench
• Perfer
• benchmark-ips
• bench9000

Notes on specific frontends

Benchmark-Interface

As well as normal usage, you can also require 'benchmark-interface' at the top of your file of benchmarks, and then run the file as a normal Ruby script. This will have the same effect as benchmark file.rb (and so will run benchmark-ips as the backend).

MRI's benchmarks

To run MRI's benchmarks you need two extra gems.

$ gem install parser unparser

For MRI's benchmarks we detect the last statement, which is usually a while loop and wrap that in a Benchmark-Interface block. If we see a variable being initialised and then used in the last statement, we copy that into the block. This doesn't work in all cases, but it does in most. The --show-rewrite option shows you what we are doing. Please file a bug if we're getting it wrong for any benchmarks in the MRI repository.

Running MRI's benchmarks requires rewriting their source code, which means running the fairly complex parser and unparser gems. This is a bit much to ask of new Ruby implementations, so you can do the rewriting using another implementation and get it to cache the result for the real implementation. The file is written to mri-rewrite-cache.rb in the current directory.

$ rbenv shell 2.3.1
$ benchmark ruby/benchmark/bm_vm1_length.rb --cache
$ rbenv shell topaz-dev
$ benchmark ruby/benchmark/bm_vm1_length.rb --use-cache

benchmark-ips and Perfer

Both benchmark-ips and Perfer allow you to write benchmarked blocks which take a number of iterations. This is done to reduce the overhead of calling the block by putting the iterations loop (presumably a while loop, so not itself involving another block) inside the benchmarked code. However this isn't supported by other backends. To make this work, we look at how many iterations are needed to make the benchmark run for about a second and always run that many iterations. The number of iterations is the same for all benchmarks, and we take the smallest number, as the larger number can be unbound and make some benchmarks take a very long time.

Backends

When you have a file of benchmarks, you can run it using different Ruby benchmarking systems. By default, it will run your benchmarks with benchmark/ips, which is usually the best choice.

Supported backends are:

• Simple looping, via --simple
• bm, via --bm
• bmbm, via --bmbm
• benchmark-ips, via --bips
• bench9000, via --bench9000

Notes on specific backends

Simple looping

Simple looping does the simplest thing possible and runs your benchmark in a loop printing the iterations per second as it goes. By default it runs for 10s, but you can use options like --time 60 to run for 60 seconds.

bm and bmbm

bm and bmbm only run your benchmark once when timing, which isn't what benchmarks written for benchmark-ips expect. Therefore if a benchmark takes less than a tenth of a second it is run several times so that it takes about a second. The number of iterations is the same for all benchmarks in the same set and is printed as the set starts.

You can turn this off with --no-scale

benchmark-ips

We run with x.iterations = 3 by default.

If a benchmark takes more than a tenth of a second, the warmup and sampling periods of benchmark-ips will be increased, as by default benchmark-ips is set up for quicker benchmarks that complete many iterations in a second.

You can turn this off with --no-scale.

You'll need to manually install the benchmark-ips gem.

Bench9000

Using Bench9000 as a backend is a little more complex. You need to define a configuration file which runs benchmark with the benchmark file, the --bench9000 flag and the name of the benchmark, and then you separately run the bench9000 command.

benchmark 'clamp_a', 'benchmark examples/benchmark-interface.rb --bench9000 clamp_a'
benchmark 'clamp_b', 'benchmark examples/benchmark-interface.rb --bench9000 clamp_b'

$ bench9000 detail --config bench9000.config 2.3.0 clamp_a clamp_b --value-per-line

The Bench9000 has the same scaling functionality as bm, but here it switches to the micro-harness. Turn it off with --no-scale.

You'll need to manually install the Bench9000 gem.

Supported Ruby Implementations

Tested versions are in brackets.

• MRI (1.8.7-p371, 1.9.3-p547, 2.0.0-p648, 2.1.10, 2.2.5, 2.3.1, head)
• JRuby (1.7.25, 9.0.5.0, 9.1.0.0, head)
• JRuby+Truffle
• Rubinius (2.71828182, 3.29)
• Topaz

Benchmark-Interface is designed to be gentle on new implementations of Ruby and should hopefully be relatively easy to get working if you are writing one.

One major limitation is that the unparser gem doesn't work with Ruby versions below 1.9. You can get around this by caching the translation result with a supported version of Ruby (the --cache option) and then using that cached translation in the version of Ruby you are benchmarking (the --use-cache option).

You also need to install these gems manually if you want to translate MRI benchmarks.

Caveats

• This tool does nothing to address the risk of inputs being constant folded or results discarded.
• Don't compare results from two different benchmarking backends.
• Some automated things we do like scaling and setting iterations can complicate benchmarking.
• Use your own scientific judgement!