GC in MRI 1.9 and REE

luislavena · January 10, 2012, 4:51pm

Hi all,

While playing with the GC in different ruby implementations i’ve got a
pretty odd situation with MRI GC.

here is the code i’ve used with the results -

gist.github.com

https://gist.github.com/MasterLambaster/1589617

Results

ree-1.8.7-2011.03
-----------------

1. Set instance value to nil
-------------------- Stats for Allocate --------------------
Live Objects before: 4025
Live Objects after: 1504032
------------------------------------------------------------

---------------- Stats for Set iVal to nil -----------------

This file has been truncated. show original

test.rb

def grab_stat
  if RUBY_DESCRIPTION =~ /Enterprise/
    ObjectSpace.live_objects
  else # >= 1.9
    ObjectSpace.count_objects[:TOTAL] - ObjectSpace.count_objects[:FREE]
  end
end

def gc
  GC.enable

This file has been truncated. show original

If i understand correctly the both implementations(MRI and REE) uses the
same GC strategy - mark and sweep.

The problem is that the MRI is not deallocating objects after the
reference is removed.
I’m not certain is it a bug or I misunderstood some GC concepts.

PS: The allocations are confirmed with vmmap tool so it’s not an live
objects calculation problem.

digitalpodcast · January 11, 2012, 10:23am

On Tue, Jan 10, 2012 at 4:51 PM, Alex N. [email protected] wrote:

The problem is that the MRI is not deallocating objects after the
reference is removed.
I’m not certain is it a bug or I misunderstood some GC concepts.

I think it’s in your understanding. Same strategy does not imply
identical algorithm. Even if both implementation use M&S they still
might have different algorithms to detect liveness of objects. It may
actually make sense to be more conservative about liveness if the
check is faster then. That would result in less objects being
collected.

Bottom line is there are no guarantees as to when an object is
collected. It can long be unreachable before it is actually
collected.

Kind regards

robert