方法定义中的Ruby Splat运算符占用更多内存

问题描述

在优化代码库时,我们尝试使用bang方法减少有意义的对象分配,但是我们在基准测试中观察到分配的对象数量减少了,但总体内存大小却增加了。

复制脚本:

# frozen_string_literal: true

require 'bundler/inline'

gemfile(true) do
  source "https://rubygems.org"

  git_source(:github) { |repo| "https://github.com/#{repo}.git" }

  gem 'benchmark-memory','0.1.2'
end

require 'benchmark/memory'

def with_bang(*methods)
  methods.tap(&:flatten!)
end

def without_bang(*methods)
  methods.flatten
end


Benchmark.memory do |x|
  x.report("with_bang") { with_bang(:a,:b,:c,:d,:e,:f,:g,:h,:i,:j,:k,:l,:m,:n,:o) }
  x.report("without_bang") { without_bang(:a,:o) }
  x.compare!
end


# Output
# Ruby version: ruby 2.7.2p137 (2020-10-01 revision 5445e04352) [x86_64-darwin19]

# INPUT: (:a,:o)
# Calculating -------------------------------------
#            with_bang   160.000  memsize (     0.000  retained)
#                          1.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)
#         without_bang    80.000  memsize (     0.000  retained)
#                          2.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)

# Comparison:
#         without_bang:         80 allocated
#            with_bang:        160 allocated - 2.00x more


# INPUT: (:a,[:f,:g],:o)
# Calculating -------------------------------------
#            with_bang   240.000  memsize (     0.000  retained)
#                          3.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)
#         without_bang   480.000  memsize (     0.000  retained)
#                          3.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)

# Comparison:
#            with_bang:        240 allocated
#         without_bang:        480 allocated - 2.00x more

在我的实验中,我相信这是由于splat运算符转换为数组所致。以下是提示我该结论的脚本。

# frozen_string_literal: true

require 'bundler/inline'

gemfile(true) do
  source "https://rubygems.org"

  git_source(:github) { |repo| "https://github.com/#{repo}.git" }

  gem 'benchmark-memory','0.1.2'
end

require 'benchmark/memory'

def with_splat(*methods)
  methods.flatten!
end

def without_splat
  methods = [:a,:o]
  methods.flatten!
end


Benchmark.memory do |x|
  x.report("with_splat") { with_splat(:a,:o) }
  x.report("without_splat") { without_splat }
  x.compare!
end

# Output
# Ruby version: ruby 2.7.2p137 (2020-10-01 revision 5445e04352) [x86_64-darwin19]

# INPUT: (:a,:o)
# Calculating -------------------------------------
#           with_splat   160.000  memsize (     0.000  retained)
#                          1.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)
#        without_splat    40.000  memsize (     0.000  retained)
#                          1.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)

# Comparison:
#        without_splat:         40 allocated
#           with_splat:        160 allocated - 4.00x more


# INPUT: (:a,:o)
# Calculating -------------------------------------
#           with_splat   240.000  memsize (     0.000  retained)
#                          3.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)
#        without_splat   240.000  memsize (     0.000  retained)
#                          3.000  objects (     0.000  retained)
#                          0.000  strings (     0.000  retained)

# Comparison:
#           with_splat:        240 allocated
#        without_splat:        240 allocated - same

我缺少什么才能了解​​这种行为?为什么会以这种方式运行?

谢谢!

编辑: 我向包含嵌套数组的基准比较添加了新的输入。通过新的输入,我们看到的结果与以前的基准测试有所不同,我感到更加困惑!

解决方法

让我们更仔细地检查两个数组:

require 'objspace'

def with_splat(*methods)
  ObjectSpace.dump(methods,output: open('with_splat.json','w'))
end

def without_splat(methods)
  ObjectSpace.dump(methods,output: open('without_splat.json','w'))
end

with_splat(:a,:b,:c,:d,:e,:f,:g,:h,:i,:j,:k,:l,:m,:n,:o)
without_splat([:a,:o])

ObjectSpace.dump_all(output: open('all_objects.json','w'))

该脚本生成3个文件:

  • with_splat.json包含有关数组阵列的数据
  • without_splat.json包含有关非分割数组的数据
  • all_objects.json包含有关所有对象的数据(很多!)

with_splat.json :(格式化)

{
  "address": "0x7feb941289a0","type": "ARRAY","class": "0x7feb940972c0","length": 15,"memsize": 160,"flags": {
    "wb_protected": true
  }
}

without_splat.json :(格式化)

{
  "address": "0x7feb941287e8","shared": true,"references": [
    "0x7feb941328d8"
  ],"memsize": 40,"flags": {
    "wb_protected": true
  }
}

如您所见,后一个数组确实消耗较少的内存(40 vs 160),但是它也设置了"shared": true,并且引用了内存地址0x7feb941328d8上的另一个对象。

让我们通过jqall_objects.json中找到该对象:

$ jq 'select(.address == "0x7feb941328d8")' all_objects.json
{
  "address": "0x7feb941328d8","frozen": true,"flags": {
    "wb_protected": true
  }
}

这是具有与上面的第一个数组相同的内存大小的实际数组。

请注意,此数组已设置"frozen": true。我假设Ruby在遇到数组文字时会创建这些冻结的数组。然后可以根据评估结果创建便宜的(更)共享数组。