From jonathan.gagnon at croesus.com  Wed Mar 19 12:39:15 2008
From: jonathan.gagnon at croesus.com (Jonathan Gagnon)
Date: Wed, 19 Mar 2008 12:39:15 -0400
Subject: [Mono-gc-list] -DPARALLEL_MARK and GC_enable_incremental()
Message-ID: <00c001c889df$c7c72a90$1b36a8c0@PROG159>

Hi,
 
I wanted to know how the mono GC behaves on multiprocessor machines.  After
reading some documentation on the Boehm GC, I'm wondering if mono compiles
the GC with -DPARALLEL_MARK.  This allows the collector to run the mark
phase in parallel, and thus use all the available processors on a
multi-processor machine (instead of using just one).  More details can be
found here :
 
http://www.hpl.hp.com/personal/Hans_Boehm/gc/scale.html
 
It seems like they also have generational collection, that can be enable by
calling GC_enable_incremental.  It doesn't seem to be called anywhere in the
mono sources, so I assume it is not used.  Is there a reason why?
 
Thanks,
 
Jonathan
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-gc-list/attachments/20080319/8485bfbe/attachment.html 

From lupus at ximian.com  Thu Mar 20 06:48:50 2008
From: lupus at ximian.com (Paolo Molaro)
Date: Thu, 20 Mar 2008 11:48:50 +0100
Subject: [Mono-gc-list] -DPARALLEL_MARK and GC_enable_incremental()
In-Reply-To: <00c001c889df$c7c72a90$1b36a8c0@PROG159>
References: <00c001c889df$c7c72a90$1b36a8c0@PROG159>
Message-ID: <20080320104850.GC22705@debian.org>

On 03/19/08 Jonathan Gagnon wrote:
> I wanted to know how the mono GC behaves on multiprocessor machines.  After
> reading some documentation on the Boehm GC, I'm wondering if mono compiles
> the GC with -DPARALLEL_MARK.  This allows the collector to run the mark
> phase in parallel, and thus use all the available processors on a
> multi-processor machine (instead of using just one).  More details can be
> found here :
>  
> http://www.hpl.hp.com/personal/Hans_Boehm/gc/scale.html

Some time ago I tested a mono version enabled with -DPARALLEL_MARK
on a workload that should have benefited from it and I saw basiccally no
performance improvement on my core 2 duo (there was a slight regression).
For kicks I just tested another case and with parallel marking
enabled performance is orrible (twice as slow).
You can easily test your workload configuring mono with the
--enable-parallel-mark option.
I didn't investigate the cause of the slowdown.

> It seems like they also have generational collection, that can be enable by
> calling GC_enable_incremental.  It doesn't seem to be called anywhere in the
> mono sources, so I assume it is not used.  Is there a reason why?

GC_enable_incremental has a number of issues that are detailed in the
documentation (mainly unexpected interrupted syscalls: we could deal
with them in the runtime, but the p/invoke code of many programs would
break unless we copy potentially lots of data around: not practical).

lupus

-- 
-----------------------------------------------------------------
lupus at debian.org                                     debian/rules
lupus at ximian.com                             Monkeys do it better