Skip to content

Latest commit

 

History

History
288 lines (205 loc) · 8.6 KB

README.md

File metadata and controls

288 lines (205 loc) · 8.6 KB
Raw

benchmark-ips

Gem Version Build Status Inline docs

DESCRIPTION:

An iterations per second enhancement to Benchmark.

FEATURES/PROBLEMS:

  • benchmark/ips - benchmarks a blocks iterations/second. For short snippits of code, ips automatically figures out how many times to run the code to get interesting data. No more guessing at random iteration counts!

SYNOPSIS:

require 'benchmark/ips'

Benchmark.ips do |x|
  # Configure the number of seconds used during
  # the warmup phase (default 2) and calculation phase (default 5)
  x.config(warmup: 2, time: 5)

  # Typical mode, runs the block as many times as it can
  x.report("addition") { 1 + 2 }

  # To reduce overhead, the number of iterations is passed in
  # and the block must run the code the specific number of times.
  # Used for when the workload is very small and any overhead
  # introduces incorrectable errors.
  x.report("addition2") do |times|
    i = 0
    while i < times
      i += 1

      1 + 2
    end
  end

  # To reduce overhead even more, grafts the code given into
  # the loop that performs the iterations internally to reduce
  # overhead. Typically not needed, use the |times| form instead.
  x.report("addition3", "1 + 2")

  # Really long labels should be formatted correctly
  x.report("addition-test-long-label") { 1 + 2 }

  # Compare the iterations per second of the various reports!
  x.compare!
end

This will generate the following report:

Warming up --------------------------------------
            addition     3.572M i/100ms
           addition2     3.672M i/100ms
           addition3     3.677M i/100ms
addition-test-long-label
                         3.511M i/100ms
Calculating -------------------------------------
            addition     36.209M (± 2.8%) i/s   (27.62 ns/i) -    182.253M in   5.037433s
           addition2     36.552M (± 7.8%) i/s   (27.36 ns/i) -    183.541M in   5.069987s
           addition3     36.639M (± 4.8%) i/s   (27.29 ns/i) -    182.994M in   5.009234s
addition-test-long-label
                         36.164M (± 5.8%) i/s   (27.65 ns/i) -    181.312M in   5.038364s

Comparison:
           addition2: 36558904.5 i/s
           addition3: 36359284.0 i/s - same-ish: difference falls within error
addition-test-long-label: 36135428.8 i/s - same-ish: difference falls within error
            addition: 34666931.3 i/s - same-ish: difference falls within error

Benchmark/ips will report the number of iterations per second for a given block of code. When analyzing the results, notice the percent of standard deviation which tells us how spread out our measurements are from the average. A high standard deviation could indicate the results having too much variability.

One benefit to using this method is benchmark-ips automatically determines the data points for testing our code, so we can focus on the results instead of guessing iteration counts as we do with the traditional Benchmark library.

You can also use ips_quick to save a few lines of code:

Benchmark.ips_quick(:upcase, :downcase, on: "hello") # runs a suite comparing "hello".upcase and "hello".downcase

def first; MyJob.perform(1); end
def second; MyJobOptimized.perform(1); end
Benchmark.ips_quick(:first, :second) # compares :first and :second

This adds a very small amount of overhead, which may be significant (i.e. ips_quick will understate the difference) if you're microbenchmarking things that can do over 1 million iterations per second. In that case, you're better off using the full format.

Custom Suite

Pass a custom suite to disable garbage collection during benchmark:

require 'benchmark/ips'

# Enable and start GC before each job run. Disable GC afterwards.
#
# Inspired by https://www.omniref.com/ruby/2.2.1/symbols/Benchmark/bm?#annotation=4095926&line=182
class GCSuite
  def warming(*)
    run_gc
  end

  def running(*)
    run_gc
  end

  def warmup_stats(*)
  end

  def add_report(*)
  end

  private

  def run_gc
    GC.enable
    GC.start
    GC.disable
  end
end

suite = GCSuite.new

Benchmark.ips do |x|
  x.config(:suite => suite)
  x.report("job1") { ... }
  x.report("job2") { ... }
end

Independent benchmarking

If you are comparing multiple implementations of a piece of code you may want to benchmark them in separate invocations of Ruby so that the measurements are independent of each other. You can do this with the hold! command.

Benchmark.ips do |x|

  # Hold results between multiple invocations of Ruby
  x.hold! 'filename'

end

This will run only one benchmarks each time you run the command, storing results in the specified file. The file is deleted when all results have been gathered and the report is shown.

Alternatively, if you prefer a different approach, the save! command is available. Examples for hold! and save! are available in the examples/ directory.

Multiple iterations

In some cases you may want to run multiple iterations of the warmup and calculation stages and take only the last result for comparison. This is useful if you are benchmarking with an implementation of Ruby that optimizes using tracing or on-stack-replacement, because to those implementations the calculation phase may appear as new, unoptimized code.

You can do this with the iterations option, which by default is 1. The total time spent will then be iterations * warmup + iterations * time seconds.

Benchmark.ips do |x|

  x.config(:iterations => 3)

    # or

  x.iterations = 3

end

Online sharing

If you want to quickly share your benchmark result with others, run you benchmark with SHARE=1 argument. For example: SHARE=1 ruby my_benchmark.rb.

Result will be sent to benchmark.fyi and benchmark-ips will display the link to share the benchmark's result.

If you want to run your own instance of benchmark.fyi and share it to that instance, you can do this: SHARE_URL=https://ips.example.com ruby my_benchmark.rb

Advanced Statistics

By default, the margin of error shown is plus-minus one standard deviation. If a more advanced statistical test is wanted, a bootstrap confidence interval can be calculated instead. A bootstrap confidence interval has the advantages of arguably being more mathematically sound for this application than a standard deviation, it additionally produces an error for relative slowdowns, which the standard deviation does not, and it is arguably more intuitive and actionable.

When a bootstrap confidence interval is used, a median of the interval is used rather than the mean of the samples, which is what you get with the default standard deviation.

The bootstrap confidence interval used is the one described by Tomas Kalibera. Note that for this technique to be valid your benchmark should have reached a non-periodic steady state with statistically independent samples (it should have warmed up) by the time measurements start.

Using a bootstrap confidence internal requires that the 'kalibera' gem is installed separately. This gem is not a formal dependency, as by default it is not needed.

gem install kalibera

Benchmark.ips do |x|

  # The default is :stats => :sd, which doesn't have a configurable confidence
  x.config(:stats => :bootstrap, :confidence => 95)

    # or

  x.stats = :bootstrap
  x.confidence = 95

  # confidence is 95% by default, so it can be omitted

end

Output as JSON

You can generate output in JSON. If you want to write JSON to a file, pass filename to json! method:

Benchmark.ips do |x|
  x.report("some report") {  }
  x.json! 'filename.json'
end

If you want to write JSON to STDOUT, pass STDOUT to json! method and set quiet = true before json!:

Benchmark.ips do |x|
  x.report("some report") {  }
  x.quiet = true
  x.json! STDOUT
end

This is useful when the output from benchmark-ips becomes an input of other tools via stdin.

REQUIREMENTS:

  • None!

INSTALL:

$ gem install benchmark-ips

DEVELOPERS:

After checking out the source, run:

$ rake newb

This task will install any missing dependencies, run the tests/specs, and generate the RDoc.