Java subclass (instantiation) performance degraded in 9.3

[kares]

A feature which originated in 9.3 (#6422) seems to have caused a 20x performance degradation in micro-benchmarks compared to 9.2.

This has been observed in production as threads pile up due synchronization and the extra overhead each <init> call has.

The following script demonstrates a Ruby sub-class of a Java class which will have poor class.new performance :

puts RUBY_DESCRIPTION + "\n\n"

require 'benchmark'

class RubyList < java.util.ArrayList

  def initialize(*args) # method's actual logic is irrelevant here
    # the idea is simply that more instructions will slow-down execution more
    foo = args.first.is_a?(NilClass) && args.shift
    bar = args.first.is_a?(String) ? args.shift : self.class.name

    if foo
      super_args = args.dup.to_java(:string)
    elsif bar
      super_args = [bar, *args]
    else
      super_args = args
    end

    super(super_args)
  end

end

time = Benchmark.realtime do
  threads = []
  1000.times do |i|
    threads << Thread.start do
      ('a'..'z').each { |c| 1000.times { RubyList.new(i.to_s, c) } }
    end
  end
  threads.each(&:join)
end

puts "1 RubyList.new(...) elapsed real-time: #{time}"

time = Benchmark.realtime do
  threads = []
  1000.times do |i|
    threads << Thread.start do
      ('a'..'z').each { |c| 1000.times { RubyList.new } }
    end
  end
  threads.each(&:join)
end

puts "2 RubyList.new elapsed real-time: #{time}"

$ jruby slow_java_super.rb 
jruby 9.2.20.0 (2.5.8) 2021-11-02 1a3255440b OpenJDK 64-Bit Server VM 11.0.14.1+1 on 11.0.14.1+1 +jit [linux-x86_64]

1 RubyList.new(...) elapsed real-time: 3.4137503600068158
2 RubyList.new elapsed real-time: 2.6153695250031888

$ jruby slow_java_super.rb 
jruby 9.3.10.0 (2.6.8) 2023-02-01 107b2e6697 OpenJDK 64-Bit Server VM 11.0.14.1+1 on 11.0.14.1+1 +jit [x86_64-linux]

1 RubyList.new(...) elapsed real-time: 54.96447179099778
2 RubyList.new elapsed real-time: 54.98580553600186

[kares]

for the record: #7762 addressed the main (contention) issue.
but the degradation compared to 9.2 is still there and will require more work to avoid the extra code in Ruby initialize.

[kares]

a script that demonstrates single thread perf: https://gist.github.com/kares/f51f8a517d176fd2fdae52471aa7859e
(degradation seems to be ~ 10x)