drop millions of allocations by using a linked list

[tenderlove]

I'm sending a PR because the reduction in allocations is so surprising to me that I'm afraid it's wrong. All tests pass on my machine, and I think it's backwards compatible, but I need review. /cc @evanphx @drbrain

Use a linked list to drop array allocations. The previous
implementation would dup arrays on every call to traverse. This patch
uses a linked list so that any block that is interested in keeping a
reference to trail can just keep the trail node around (the trail node
points at it's parents).

This approach reduces allocations from 11173668, to 2940.

Here is the test I used:

require 'stackprof'
require 'allocation_tracer'
require 'rubygems/test_case'
require 'rubygems/ext'
require 'rubygems/specification'
require 'benchmark'

class TestGemSpecification < Gem::TestCase
  def test_runtime
    make_gems do
      StackProf.run(mode: :wall, out: '/tmp/out.dump') do
        assert_raises(LoadError) { require 'no_such_file_foo' }
      end
    end
  end

  def test_alone
    make_gems do
      tms = Benchmark.measure {
        assert_raises(LoadError) { require 'no_such_file_foo' }
      }
      p tms.total
      assert_operator tms.total, :<=, 10
    end
  end

  def test_memory
    make_gems do
      ObjectSpace::AllocationTracer.setup(%i{path line type})
      r = ObjectSpace::AllocationTracer.trace do
        assert_raises(LoadError) { require 'no_such_file_foo' }
      end

      r.sort_by { |k,v| v.first }.each do |k,v|
        p k => v
      end
      p hash_alloc: ObjectSpace::AllocationTracer.allocated_count_table[:T_HASH]
      p array_alloc: ObjectSpace::AllocationTracer.allocated_count_table[:T_ARRAY]
      p :TOTAL => ObjectSpace::AllocationTracer.allocated_count_table.values.inject(:+)
    end
  end

  def make_gems
    save_loaded_features do
      num_of_pkg = 7
      num_of_version_per_pkg = 3
      packages = (0..num_of_pkg).map do |pkgi|
        (0..num_of_version_per_pkg).map do |pkg_version|
          deps = Hash[(pkgi..num_of_pkg).map { |deppkgi| ["pkg#{deppkgi}", ">= 0"] }]
          new_spec "pkg#{pkgi}", pkg_version.to_s, deps
        end
      end
      base = new_spec "pkg_base", "1", {"pkg0" => ">= 0"}

      Gem::Specification.reset
      install_specs base,*packages.flatten
      base.activate

      yield
    end
  end
end

Before:

{:TOTAL=>11173668}

After:

{:TOTAL=>2940}

[tenderlove]

I think this will fix #1169 and #1167

[tenderlove]

Also the runtime test:

master:

[aaron@TC rubygems (master)]$ ruby -I lib:test test.rb -n test_runtime 
Run options: -n test_runtime --seed 52721

# Running tests:

.

Finished tests in 34.727601s, 0.0288 tests/s, 0.0288 assertions/s.

1 tests, 1 assertions, 0 failures, 0 errors, 0 skips

This branch:

[aaron@TC rubygems (omg)]$ ruby -I lib:test test.rb -n test_runtime 
Run options: -n test_runtime --seed 7522

# Running tests:

.

Finished tests in 0.050260s, 19.8965 tests/s, 19.8965 assertions/s.

1 tests, 1 assertions, 0 failures, 0 errors, 0 skips

[kuldeepaggarwal]

@tenderlove: Awesome.. 👍

[jcutrell]

This is one of the coolest merge requests I've seen for so many reasons.

[AnatoliiD]

Yeah! Really awesome! 👍 👍 👍

[evanphx]

I'd believe it. But look at the linked list that resolver uses for exactly the same reasons, maybe we can use it here too.

[tenderlove]

@evanphx where can I find it? I don't know the resolver code very well.

[tomciopp]

@tenderlove I believe the code you are looking for is here:
https://github.com/rubygems/rubygems/blob/master/lib/rubygems/util/list.rb

[tenderlove]

@tomciopp thanks!

I've updated the PR to use the built-in linked list class. I think we can implement the current instance method traverse in terms of the class method I've introduced (though ultimately I'd like to deprecate the instance method), but I will do that if this PR is acceptable for merging.

[skunkworker]

I may be wrong but shouldn't this be _traverse instead?

[libo]

Adding a?

private_class_method :_traverse

[libo]

👍

[chus]

👍

[byroot]

If you include Enumerable you don't need the explicit find implementation anymore. The to_a implementation could go too if Struct didn't have a default one.

[tenderlove]

I'd rather do that in a different commit. 😨

[ms-ati]

@byroot This isn't quote true. You have to remove Struct (which has its own Enumerable inclusion and meaning), and reverse the to_a. See PR #1200 ;)

[pawitk]

👍

[libo]

Is the merge button you have to click @tenderlove,

[tenderlove]

@libo hah, I closed and reopened to get travis to build again.

[libo]

:-) eheh

By this PR was a hell of a #fridayhug thanks @tenderlove !

[canweriotnow]

[draffauf]

Mind blown. Great job.

[sstelfox]

Wow awesome. Nice find!

[renatomoya]

Way to go @tenderlove

[vcavallo]

8 O

[amaltson]

Awesome work @tenderlove, can't wait for this to be released 😄

[gbxl]

Dude that's awesome! Good job.

[zeevallin]

@tenderlove No matter what people on the internet say, what you're doing is fantastic. Please don't get discouraged. Thank you for spotting and optimising this! #FridayHug

[seandmccarthy]

I may have missed something, but I think this should be traverse rather than _traverse as the former is the only one (I can see) with a 3 element signature.

[cheald]

I think you have that backwards; traverse takes 1 param + block, _traverse takes 2 + block. And in any case the Gem::List is generated prior to the invocation, so traverse wouldn't be appropriate anyhow.

[seandmccarthy]

Yeah sorry, scratch that. I was looking at an earlier version and the call is correct now.

[cheald]

👍

[ghost]

👍

[tenderlove]

@mfazekas you are totally right. Think we should revert?

[mfazekas]

@tenderlove i think it have to be reverted, and we need to add some test that fails with that implementation.

[tenderlove]

@mfazekas do you have any idea how to make a test that fails with this implementation? I don't want to revert until we can prove it's wrong (no doubt it is wrong, just need to make a test).

[mfazekas]

i'll try to came up with a testcase sometime later today

[mfazekas]

#1191 shows is a testcase, demonstrating the issue with your optimization. But it also shows that the original one despite all the multiple reverse's and exponential algorithm will prefer earlier and not later versions.

So we use find_in_unresolved_tree in case we cannot find a required file in any of the activated gems. The idea is that we check indirect dependencies of the activated gems and try to select a gem that contains the file, but doesn't cause a conflict. And from all of the possibilities we should prefer latest gem versions.

Current implementation is has many issues as its exponential, might select conflicting gems (#1169) and might not use the latest versions of the possibilities (#1191).

[twelve17]

Some 👏 for @mfazekas too!

[ghost]

A valid reason to learn linked lists 👍

[ms-ati]

Hi @tenderlove, very cool.

After seeing this commit on Ruby Weekly News, I took a look at this code for the first time. I just submitted a PR to hopefully improve the linked list implementation used here (#1200).

Specifically, in the #to_a code path that is frequently called, Array#unshift was called for every list element. I believe this is quadratic performance, since unshift is linear. Is that right?

[007lva]

👍