[ofa-general] [RFC] upstream IB router support

Sean Hefty sean.hefty at intel.com
Thu Oct 25 10:50:23 PDT 2007


>> The patches do not introduce any new
>> protocols or SA attributes.  Visible changes are limited to setting the
>> DLID field in the CM REQ message to an invalid value that the passive
>> side keys off of to determine the correct value.  A node which does not
>> support this would simply reject the connection with an invalid LID.
>
>How might this affect end node operation when there are standard based
>routers ? If there are other larger changes for that, then this
>particular issue is a red herring.

Assuming that the CM protocol does not change, the standard will need to define
a way for the active side to obtain correct CM REQ values.  (The patches handle
this btw.)  This likely requires new host to SA interactions.  For now, the
patches use the defined path record query, which is likely inadequate based on
previous discussions.

>I do think it's important to try to keep in mind if it is possible to
>smooth a migration path for end nodes (and SMs) in terms of prestandard
>and standard routers. That's not to say that there should be no changes;
>just that it would be nice to be able to tell the two apart and make
>intelligent choices based on this.

IMO, some of this falls into the routing architecture.  Does it change the CM
protocol or modify/add SA attributes (path record)?

But there's no sure way for one side of a connection to know beforehand if the
other side is following the standard.  I would consider these patches purely
experimental with no guarantee that they interoperate with any defined standard.
They give people developing routers something that they can use to test with,
and care was taken to try to avoid potential interoperability issues.

- Sean



More information about the general mailing list