Object allocation using literals vs. constants

object_allocation.rb

require "benchmark"
require "gc_tracer"
include Process

class WithLiteral
  def call
    []
  end
end

class WithConstant
  EMPTY_ARRAY = []

  def call
    EMPTY_ARRAY
  end
end

class WithFrozenConstant
  EMPTY_ARRAY = [].freeze

  def call
    EMPTY_ARRAY
  end
end

n = 10_000_000
run = ->(ex) { n.times { ex.new.call } }

gc_report = lambda do |label|
  s = GC.stat
  allocated = s[:total_allocated_object]
  freed = s[:total_freed_object]
  live = s[:heap_live_slot]

  <<-REPORT
  #{label} GC stats:
    Allocated: #{allocated}
    Freed:     #{freed}
    Live:      #{live}
  REPORT
end

[WithLiteral, WithConstant, WithFrozenConstant].each do |ex|
  fork do
    start_report = ""
    end_report = ""

    Benchmark.bm do |bm|
      bm.report(ex.to_s) do
        start_report = gc_report.("#{ex} start")
        run.(ex)
        end_report = gc_report.("#{ex} end")
      end
      puts "", start_report, ""
      puts end_report, ""
    end
  end # fork
  wait # for forked process
end

results.txt

> ruby lib/benchmarks/object_allocation.rb
       user     system      total        real
WithLiteral  2.930000   0.000000   2.930000 (  3.018267)

  WithLiteral start GC stats:
    Allocated: 49483
    Freed:     17301
    Live:      32182

  WithLiteral end GC stats:
    Allocated: 20049491
    Freed:     20016504
    Live:      32987

       user     system      total        real
WithConstant  2.290000   0.000000   2.290000 (  2.327953)

  WithConstant start GC stats:
    Allocated: 49484
    Freed:     17301
    Live:      32183

  WithConstant end GC stats:
    Allocated: 10049492
    Freed:     10016489
    Live:      33003

       user     system      total        real
WithFrozenConstant  2.300000   0.000000   2.300000 (  2.315593)

  WithFrozenConstant start GC stats:
    Allocated: 49485
    Freed:     17301
    Live:      32184

  WithFrozenConstant end GC stats:
    Allocated: 10049493
    Freed:     10016489
    Live:      33004

@elskwid also check out benchmark-ips for cases like this.