[ofa-general] PATCH [0/3] osm: adding root and compute node guid files options for fat-tree

[Hal Rosenstock]

On Thu, 2007-06-14 at 09:45, Sasha Khapyorsky wrote:
> On 15:36 Thu 14 Jun     , Yevgeny Kliteynik wrote:
> >  Sasha Khapyorsky wrote:
> > > Hi Yevgeny,
> > > On 11:19 Thu 14 Jun     , Yevgeny Kliteynik wrote:
> > >>  The following three patches are adding root and compute node guid files
> > >>  options for fat-tree routing,
> > > Is there any reason to not share root guids file option with up/down?
> > 
> >  There are two new options for fat-tree: roots and compute nodes (CN).
> >  These two will be very "tightly coupled" and would have more implication
> >  on the routing than in case of up/dn roots. For instance, having root
> >  file but not CN file means that the topology doesn't have to be pure 
> >  fat-tree,
> >  but all the CAs are considered CNs and have to be on the same level of the 
> >  tree.
> >  And there is similar implication of all the combinations of these two 
> >  options.
> > 
> >  Because of this coupling I wanted to differentiate these two options from
> >  the up/dn roots.
> > 
> >  Thoughts?
> 
> I still not have strong option about two options against common one.

Me neither.

> Hypothetically if in some days we will implement routing engine chains
> (so failed algo will fallback to next in chain and not just to default)
> separate options could be useful.

So is this a(nother) reason to keep the roots separate or would that be
dealt with when the routing fallback strategy changes ?

-- Hal

> > > Also the way how root guids are handled (in both up/down and ftree)
> > > doesn't look very optimal - guids are loaded to dynamic list, the list
> > > is converted to map, this map is matched and root nodes are marked as
> > > roots. Isn't it would be easy just to mark root nodes during file parsing?
> > 
> >  The only thing you can save here is converting list to map:
> 
> I don't think the root guids map is needed - you can just set is_root
> field for sw nodes by guid(s) specified in the file, since you already
> have sw by guid map.
> 
> >  You have to parse the guids file anyway, and you have to build all the
> >  fat-tree data structures anyway. So if you parse the file and fill the
> >  map right away instead of filling the list first, you will save the list2map 
> >  conversion.
> >  But then up/dn and fat-tree can't use the same function to parse the guid 
> >  file,
> >  and since the list2map conversion is not a big deal (we're talking about 
> >  list
> >  of roots, which is couple of hundreds of guids at max), I prefer to leave it
> >  and not to use separate parsing functions for up/dn and fat-tree.
> 
> You can pass custom callback to common parser.
> 
> >  BTW, since we're on this subject, how about removing the list2array 
> >  conversion
> >  in the same place in up/dn routing?
> 
> Sure, similar junk should be cleaned up in up/down too (and my original
> complain was about both root guids users).
> 
> Sasha