Ractor newobj cache

[ko1]

Now object allocation requires VM global lock to synchronize objspace.
However, of course, it introduces huge overhead.
This patch caches some slots (in a page) by each ractor and use cached
slots for object allocation. If there is no cached slots, acquire the global lock
and get new cached slots, or start GC (marking or lazy sweeping).

[ko1]

On the Linux box,

Warning[:experimental] = false

def task
  i = 0
  while i < 1_000_000
    a = [[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]
    i += 1
  end
  # 100_000.times{ [[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]] }
end

MODE = (ARGV.shift || :serial).to_sym
TN = 8

case MODE
when :serial
  TN.times{ task }
when :r_serial
  TN.times{
    Ractor.new{
      task
    }.take
  }
when :r_parallel
  TN.times.map{
    Ractor.new{
      task
    }
  }.each{|r| r.take}
else
  raise
end

This program shows:

                               user     system      total        real
serial/master_mini         0.000085   0.000028   0.321040 (  0.321028)
serial/miniruby            0.000136   0.000045   0.289354 (  0.289312)
serial/master_ruby         0.000144   0.000048   0.333495 (  0.333450)
serial/ruby                0.000134   0.000044   0.333663 (  0.333622)
r_serial/master_mini       0.000097   0.000032   0.564617 (  0.564514)
r_serial/miniruby          0.000142   0.000047   0.309408 (  0.309311)
r_serial/master_ruby       0.000134   0.000044   0.595155 (  0.595050)
r_serial/ruby              0.000119   0.000040   0.348166 (  0.348066)
r_parallel/master_mini     0.000079   0.000027   7.840724 (  2.451340)
r_parallel/miniruby        0.000061   0.000020   1.708715 (  0.777945)
r_parallel/master_ruby     0.000092   0.000031   8.217047 (  2.541274)
r_parallel/ruby            0.000103   0.000035   1.921474 (  0.868992)

master_mini: master's miniruby
master_ruby: master's installed ruby
miniruby: modified miniruby
ruby: modified ruby

This patch improved the object allocation performance, especially on multi-ractor mode.
However it is still slow with parallel allocations.

[ko1]

With another task:

def task
  i = 0
  s = '0,' * 1_000
  while i < 1_000
    a = s.split(/,/)

    i += 1
  end
end

                               user     system      total        real
serial/master_mini         0.000094   0.000023   0.733983 (  0.733979)
serial/miniruby            0.000144   0.000036   0.738881 (  0.738846)
serial/master_ruby         0.000138   0.000035   0.798197 (  0.798159)
serial/ruby                0.000139   0.000035   0.803850 (  0.803837)
r_serial/master_mini       0.000142   0.000035   1.131381 (  1.131291)
r_serial/miniruby          0.000144   0.000036   1.051713 (  1.051624)
r_serial/master_ruby       0.000131   0.000033   1.336663 (  1.336582)
r_serial/ruby              0.000146   0.000037   1.211739 (  1.211644)
r_parallel/master_mini     0.000093   0.000023  12.059595 (  3.394471)
r_parallel/miniruby        0.000103   0.000026   9.873731 (  2.765747)
r_parallel/master_ruby     0.000095   0.000024  13.050518 (  3.613259)
r_parallel/ruby            0.000083   0.000021  10.168195 (  2.848708)

It is still slower. Maybe because VM lock contention on the Encoding.

[ko1]

Another task (Array#product to make many arrays):

def task
  i = 0
  a = (1..100).to_a
  while i < 1_000
    as = a.product(a)
    i += 1
  end
end

                               user     system      total        real
serial/master_mini         0.000185   0.000031   2.461938 (  2.461924)
serial/miniruby            0.000154   0.000025   2.452719 (  2.452708)
serial/master_ruby         0.000151   0.000025   3.474300 (  3.474296)
serial/ruby                0.000145   0.000025   3.396823 (  3.396814)
r_serial/master_mini       0.000151   0.000025   3.183569 (  3.183501)
r_serial/miniruby          0.000113   0.000019   3.173086 (  3.173036)
r_serial/master_ruby       0.000144   0.000024   4.357124 (  4.357084)
r_serial/ruby              0.000114   0.000019   3.029095 (  3.029025)
r_parallel/master_mini     0.000089   0.000015  30.343168 (  9.423811)
r_parallel/miniruby        0.000097   0.000017   5.655085 (  3.710885)
r_parallel/master_ruby     0.000103   0.000017  32.811944 ( 10.600854)
r_parallel/ruby            0.000093   0.000015   5.451007 (  3.810941)

This patch is effective (compare with master).