Editor's note:  These minutes have not been edited.
 
Subject: AgentX WG at 37th IETF, reported by Dale Francisco

Agentx WG at 37th IETF, December 9-13, 1996
-------------------------------------------
At the 37th IETF in San Jose, December 9-13, 1996, the SNMP
Agent Extensibility Working Group (agentx) met on
Wednesday, December 11, 9:00-11:30.
 
The most active participants in the discussion (a few
names may have been missed due to confusion on the part of
the note-taker) were: Uri Blumenthal, Linda Cabeca, Jeff Case,
Dale Francisco, Maria Greene, Jeff Johnson, Deirdre Kostick,
Bobby Krupczak, Bob Natale, Dave Perkins, Randy Presuhn, and
Don Ryan.  About 80 people were present at the meeting.
 
Meeting notes were taken by Dale Francisco, wg editor.
 
AgentX WG Meeting (Wed, December 11, 09:00-11:30)
-------------------------------------------------
Working group chair Bob Natale opened the meeting with a
statement that he wanted to concentrate first on a review
of section 11 ("Questions and Issues") of the most recent
protocol draft, and then move on to discussion of
implementations and interoperability testing.
 
There was some discussion of 11.1, the decision to proceed
with multiple varbinds per AgentX PDU. Jeff Case proposed
that we require subagents to be able to accept
multiple-varbind PDUs, but not require master agents to
emit them. Others felt that only in the case of AgentX
GetNext and GetBulk PDUs should the master be allowed to
emit fewer than all the relevant varbinds in a single
AgentX PDU.  The issue was not resolved.
 
The next item to be discussed (11.3) was the decision to
remove the value "all contexts" for the context field in
the AgentX header. Some felt that by not allowing the
possibility of subagents registering for "all contexts", a
significant percentage of the market for extensible agents
would be unable to use AgentX. It was agreed that there
needed to be more discussion on the list of whether
registration for "all contexts" is necessary.
 
Several people voiced concerns about how changes in the
protocol draft were tracked from version to version. There
was a request, agreed to by the chair and the editor, that
either a longer rationale section or a separate rationale
document be added to explain the motivation behind various
design decisions.  There was also a request, agreed to by
the editor, that instead of posting interim versions of the
protocol draft to the mailing list, each new iteration
would be submitted as a new internet draft.
 
The question was raised as to whether AgentX was meant to
be able to run over a connectionless transport, and if so,
how.  The consensus was that AgentX is only specified over
connection-oriented transports.
 
Several people felt that, though by itself it would not
solve the problem of running on a connectionless transport,
the idea of a connection or session ID in the AgentX header
was a good one.
 
There was a great deal of discussion about index allocation
and instance reservation. Many felt that there was
insufficient support for multiply-indexed tables in the
current protocol draft, and that, in fact, the current
index allocation scheme only worked for singly-indexed,
arbitrary integer tables. The example was given of a table
with two index objects, the first of which is shared. Since
indices must be allocated one at a time, it would be
impossible for subagent A to get an allocation for, say,
"1.2", and subagent B to allocate "1.3".
 
Many people felt that any model that relied on the master
agent alone for index allocation was inherently unworkable,
since, by design, only the subagent has detailed MIB
knowledge. Various solutions were proposed, including a new
entity, the "index allocation server"; ad hoc,
outside-the-protocol subagent-to-subagent communication;
and new AgentX protocol operations that would allow
subagents to discover existing indexes and to request
arbitrary length (not just single integer valued) index
objects.
 
There was also discussion of the unionized registration
concept that had been deleted, for lack of a compelling
defense, from the most recent protocol draft. It was
asked how, without unionized registration, it would be
possible to do row creation in tables supported by multiple
subagents, where only one of the subagents would be
capable of creating a particular row.
 
It became clear, given the intensity of the discussion and
the cogency of some of the counterexamples that had been
brought forward, that further discussion of the protocol
was necessary before general implementation could go
forward (though in fact, many of the questions that were
raised were the result of initial implementations of the
current protocol draft).  Deirdre Kostick, NM Area
Director, asked that everyone with proposals for changes to
the protocol draft post them to the mailing list by a
specified time, and that we agree on a date sometime after
that for achieving consensus on the protocol.
 
It was agreed that proposed changes to the protocol be
posted to the list by January 10, 1977; that we determine
final consensus on those issues by February 14; and that
a new version of the protocol draft would be submitted to
the IETF by February 28.