Re: Patch: Bitmap marking

[Jeff Sturm <jeff dot sturm at appnet dot com>]

Bryce McKinlay wrote:
> I've cleaned up, tested, and written ChangeLog entries for the Boehm
> GC bitmap descriptor marking patch. This is basically the same as Hans'
> version of the patch, but with formatting fixes, and a fix for an
> off-by-one error in the vtable copying code in _Jv_FindArrayClass. The
> patch is working well with the current tree.

I tested this patch on a few different systems.  It looks good, with one
mysterious exception: on Alpha it caused a strange abort() in _Jv_ThreadWait. 
Although it may not be related to the bitmap marking, I've never seen that
happen, so I'll continue to investigate.

I also tried GCBench
<uri:http://www.hpl.hp.com/personal/Hans_Boehm/gc/gc_bench.html> to get an idea
of the improvements with bitmap marking.  I tested the original marking code
along with pure conservative marking and the new bitmap marking on three
development machines I have:

                        | original  | conservative  | bitmap marking
------------------------+-----------+---------------+------------------
i686-pc-linux-gnu       |  16.9s    |  15.1s (-11%) |  12.8s (-25%)
sparc-sun-solaris       |  43.4     |  33.1  (-24%) |  33.3  (-24%)
alpha-unknown-linux-gnu |  39.5     |  26.4  (-34%) |  28.5  (-28%)

On x86 bitmap marking seems to be a clear win, but on these two RISC machines it
is actually slower than simple conservative marking.  Not quite what I expected.

--
Jeff Sturm
jeff.sturm@appnet.com