RE: Please read - proposed WG termination

> -----Original Message-----
> From: Vivek Kashyap [mailto:[email protected]]
> Sent: Tuesday, August 30, 2005 2:41 AM
> To: Yaron Haviv
> Cc: Michael Krause; [email protected]; [email protected];
> [email protected]
> Subject: RE: [Ipoverib] Please read - proposed WG termination
> 
> On Tue, 30 Aug 2005, Yaron Haviv wrote:
> 
> >
> >
> > By nature an IPoIB endpoint also contains QPs, so even theoretically
an
> > SM cannot deal with the problem
> >
> 
> Am I right in assuming that you are suggesting LID + QPN + GID
failover?
> LID is certainly beyond IPoIB. QPN + GID are in the ambit.
> 
> Vivek
> 

I'm not suggesting a solution, just pointing out the problem 
A sender sends to a LID+QP (it resolved the GID->LID before) 
When a node failed the sender keeps on sending to that LID+QP (address
vector)

If one suggests using the SM (e.g. re-route the LID) it won't solve the
problem, since the new receiver may not have the same QP
Something needs to tell the sender to redo the address resolution

It can either be an IB thing (e.g. IB GID out notifications that
requires the SM to now tell it to all the nodes using unicast, probably
not the most scalable approach), or UNARP (a simple null ARP broadcast
that deletes that entry in all nodes, require support for UNARP in IPoIB
drivers)

In any case that deserves a discussion since if we choose the later
option it relates to IPoIB and IETF 

Yaron