From: Paul Brannan <pbrannan@...>
Date: 2010-01-11T23:52:07+09:00
Subject: [ruby-core:27536] Re: better GC?

On Fri, Dec 11, 2009 at 06:55:53AM +0900, Roger Pack wrote:
> Could I put in a small plea for a better GC?
> 
> With this code:
> 
> (1..7).each{|n| puts n; a = ['a']*(10**n); a.inspect;}
> 
> when n gets high enough almost 50% of the time is spent in GC.
> Somewhat frustrating.

I modified this program slightly (imo cases n=1..6 aren't useful for a
profile):

  pbrannan@random:~/tmp> cat test.rb
  100.times do |n|
    puts n
    a = ['a']*(10**6)
    a.inspect
  end

The results aren't at all surprising.  Callgrind reports 52% of the
application's run time is spent in gc_mark.  21% of the time is spent in
gc_mark_children, which means the majority of the time the objects
are immediate objects or have already been marked (I'm guessing the
latter in this program).

This probably isn't very representative of a real-world application due
to the small number of objects, but the large size of the array.

It makes me think two things though:

1. Is gc_mark getting inlined?  If it were, marking the array could be
   done without so many function calls.

2. If there were many objects, what would be the cost of resetting the
   FL_MARK flag on all the objects in gc_sweep?  Could this cost be
   eliminated by using a 2-bit counter instead of a single flag?

   RBASIC(v)->flags = 0 /* when the object is created */
   counter = ((counter + 1) % 3) + 1; /* before marking */
   if (p->as.basic.mark_counter == counter) /* to check if marked */
   p->as.basic.mark_counter = counter; /* to mark */
   ; /* no need to reset all the flags at the end now */

Hoping to spark some thought/discussion,

Paul