[ofa-general] Re: ARP in IPoIB

[Todd Rimmer]

________________________________

From: Amar Mudrankit



Michael,
thanks for your reply. But, this gives rise to couple of questions..

1] If such multicast routing protocol for IB routers is not yet specifid
by IBTA or IETF, then current implementation have IP subnet restricted
within a IB subnet. According to RFC 4391, section 9.1.1, the link layer
address is formed through combination of GID + QPN. If we are not
spanning across IB subnets what is the use of GID as we need to get LID
from GID? Probably, in that case ARP reply with LID,Q_Key and other path
information would be helpful which resolves path in 1 loop than 2 loops
in case of GID(first to resolve GID and then to get LID). 

[Todd Rimmer] Basing IPoIB on the GID keeps open the opportunity for
IPoIB to span IB subnets in the future.  Also this permits the SM to
manage the paths and PathRecord parameters appropriately even in
non-routed IB networks.  For example, if multi-pathing is used (LMC!=0,
hence giving multiple LIDs per port), the SM may respond to PathRecord
requests for a given Destination GID with a different LID depending on
Source GID.  Such a mechanism can be used to manage routes in the
fabric, etc.  That is just a simple example, since the PathRecord
includes lots of other information as well (QOS, routing info, MTU,
etc).  The SM can provide different PathRecord values for each Source
GID talking to a given Destination GID.

The IETF needed a "MAC Address" for IPoIB.  GID+QPN gave them a unique
endpoint with the potential to work through routers and still support
the full intentions of IB's QOS and routing options.  LID+QPN would
severely limit those capabilities.

In general it's a bad idea for end nodes to simply exchange LIDs as it
bypasses many of the intentions of the IB spec.  Such applications will
break in fabrics which use the more advanced IB QOS, routing, etc
options.

 

Todd Rimmer

Chief Architect 

QLogic System Interconnect Group

Voice: 610-233-4852     Fax: 610-233-4777

Todd.Rimmer at QLogic.com <mailto:Todd.Rimmer at QLogic.com>   www.QLogic.com

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20070726/011f99ac/attachment.html>