[ofa-general] Re: [PATCH] ib/ipoib: Reduce comparison size in data path

[Hal Rosenstock]

On Thu, 2008-01-03 at 09:49 +0200, Eli Cohen wrote:
> On Wed, 2008-01-02 at 10:16 -0800, Roland Dreier wrote:
> > > In the majority of cases, if the neighbour will change, it will
> >  > be reflected in the guid part of the GID (bytes 8-15). If the GID
> >  > prefix will change as well (bytes 0-7) it will be because the master
> >  > SM has changed, in which case we will get an SM change event resulting
> >  > in all paths flushed.
> > 
> > Is it guaranteed that an active SM can't change a GID prefix?
> I know opensm has a fixed, hard coded subnet prefix.

OpenSM's subnet prefix is determined via a config file (or defaults if
not overridden there). There has been discussion about supporting
reconfiguration without OpenSM restart. (One can also envision a restart
scenario here where the subnet prefix is changed). Also, OpenSM is not
the only SM out there and IMO ideally we would/should rely on only what
the IBA architecture requires and not on current implementation.

-- Hal

> > Especially if we're using a GID at an index != 0?
> I think ipoib uses only the GID from index 0, isn't it?
> >   In other words, is
> > this change definitely 100 percent safe?
> It looks safe to me but I wanted to hear other opinions.
> > 
> > Also I assume this change is coming from performance tuning.  For
> > patches like this it is always helpful to include hard data like "this
> > gives a speedup of X on test Y on system Z."
> > 
> > Thanks...
> Indeed I am working on performance and on the branch I am working on,
> which is different than main branch, it does makes a slight difference.
> I am trying to improve the throughput of small (up to 128 bytes) UDP
> messages where I am CPU bound so everything counts. But I believe that
> if it is correct than we should use it even if the improvement is not
> outstanding.
> 
> 
> _______________________________________________
> general mailing list
> general at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general