[ofa-general] Re: [PATCH] opensm/osm_mcast_mgr.c: Only sendMCG delete notice if not previously done

[Eitan Zahavi]

In OpenSM we separated "MCG Registration" from "MCG Routing". 
The reason was that on very large networks it is better to avoid the need 
to re-compute the tree for each joining node. Instead the MC routing is done for 
a set of changes.

Once a MCG group is deleted by the last member leaving the group (immediately) it 
is required to still exist until the routing engine deletes its routing.
We mark the group with "to be deleted" flag so we know it will be deleted soon.

EZ

> -----Original Message-----
> From: Hal Rosenstock [mailto:hrosenstock at xsigo.com] 
> Sent: ד 02 יולי 2008 16:39
> To: Sasha Khapyorsky
> Cc: general at lists.openfabrics.org; Eitan Zahavi
> Subject: Re: [ofa-general] Re: [PATCH] 
> opensm/osm_mcast_mgr.c: Only sendMCG delete notice if not 
> previously done
> 
> On Fri, 2008-06-27 at 19:48 +0300, Sasha Khapyorsky wrote:
> > BTW do you remember what was the reason to not delete MCG on a last 
> > leave request? I know that spec allows this (but requires 
> to post trap
> > 67 on a last leave), but is there any practical reason for 
> doing this 
> > in OpenSM?
> 
> >From the spec perspective, there was no need to make it 
> other than lazy
> as that's the way multicast works so implementations can take 
> advantage of a "convenient" time to do this. I really haven't 
> looked at this specifically in OpenSM and am not sure but 
> think this was added for a practical reason (a real life scenario).
> 
> I believe Eitan wrote this code so he is probably best to 
> comment and hopefully he will :-)
> 
> -- Hal
> 
> 
>