IP over IB WG minutes 7/14, as taken by Rob Thurlow
Jerry Chu co-chair present and presenting for others
Agenda items:
WG last call - three drafts out for last call now; link and multicast,
encapsulation and DHCP over IB. Please add your comments - last good chance! You
have until end of July. WG wants to put these drafts on standards track and to
move these to proposed standard.
IPoIB MIBs - Sean Harnady's slides presented by Jerry. Six MIBs defined in WG
charter. Textual conventions MIB version 1 available. IPOIB Infiniband
Interfaces MIB version 3 available. IPOIB Subnet Management MIB version 2
available. [one more version 2, missed the name] IPOIB Baseboard Management
Agent MIB not yet published, and IPOIB Sample Counters MIB yet to be completed.
Need to complete two drafts at least to initial version, and the others need a
thorough review. Also need some implementation experience, though this is not
yet necessary to advance. Also: do we need more MIBs?
Poll for new agenda items met with silence.
Recharter discussion - current WG scope is limited to basic encapsulation
method. Some advanced features have been discussed - should this be added to
scope? IETF rules: WG needs to register sufficient interest, must have a concise
definition and scope, deliverables and milestones need to be clear, and IESG
must agree.
Work items: IP datagrams in non-UD IB transports; Demuxing IP datagrams over
multipls QPs based on some TBD criteria; other client capability, e.g. SDP;
skipping/disabling TCP/UDP checksum when running over RC/UC (this is
controversial!).
Area director - do we have any idea of the ordering and desired delivery of
running IPoIB on top of a non-UD transport? Are we being clear enough about what
we might do? Answer: No, not clear enough. Supporting some of this on reliable
connections is the main thing. If we can use RC, a larger MTU could be used and
we could have better performance. Some ideas come from a previous draft. Take
this back to the mailing list?
Audience in room is not really familiar with the issues. Multiple QPs - could
introduce routing and session establishment problems, unclear if it is solvable.
Jerry polled the room for whether this item should be added to the WG charter.
Two were against it. Reasoning from Roger Cummings who is one of the two - see
no use of it. only want enough IP to be able to set up signalling, but want to
run SDP (Session Description Protocol) instead. Jerry asked if this feature is
only optional so only those who think it's useful will implement it, will he
oppose it? The reply is no. AD commented that there is a cost vs benefit
tradeoff one must weigh. "Other client capability - SDP" - is this even in
scope? Maybe not. Jerry - SDP out of scope, but a few bits to be able to
facilitate this could be OK. AD - bar is higher than "could be useful".
Bypassing checksums - lots of people have had this idea, and they always found
they had to add them back later on, so why? Attendees with an opinion all agreed
that bypassing was bad.
Next steps: Move 3 IPoIB drafts plus IB Interfaces MIB to standards track.
Need more reviewers plus volunteers for other drafts, Finally, expand WG charter
to address advanced features.
