[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Diffserv] Final minutes from Washington...

...are attached. The only change is the addition of Yoram's extra notes.

Diffserv Working Group meeting, Washington D.C.

Chairs: Brian Carpenter and Kathie Nichols
Reporter: David Black (additional notes from Yoram Bernet)

----- Monday Evening Session, November 8, 1999 -----

-- Agenda Bashing and WG status

Core WG drafts today.  Tomorrow will take up unsolicited drafts.

-- New Terminology - Dan Grossman, Motorola

This document's purpose is to capture group decisions on terminology and taxonomy
	for use in future documents.  Summary of these changes:

(1) Use the terms Service Level Specification and Traffic Conditioning Specification, not
	Agreements (latter implies business contracts).
(2) The definition of PHB Group in RFC 2475 doesn't fit the entire set of AF classes.
	Hence, AF is a type of PHB Group.  AF1x, AF2x, AF3x, and AF4x are instances of this
	type, only the instances are PHB Groups.  The next update of RFC 2597 will reflect this.
(3) The DS Field is no longer the entire TOS octet.  DS Field is the 6 MSBs of the (former)
	IPv4 TOS Octet and IPv6 Traffic Class octet.  The other two bits are currently
	unused by diffserv.  DSCP is a value in the DS Field.  draft-bradner-iana-allocation-02.txt
	(soon to be an RFC) is an example of this usage.
(4) MPLS needs notion of a set of Behavior Aggregates with a common ordering constraint.
	Terms: PHB Scheduling Class = PHB group that has a common ordering constraint
	Ordered Aggregate = Set of Behavior Aggregates that share an ordering constraint.
	The words "All of the packets of an OA are members of the same PHB scheduling class."
	are problematic due to the use of "members of".  Francois Le Faucheur has provided
	an alternative: "An Ordered Aggregate is the set of Behavior Aggregates that correspond to
	the PHBs of a PHB Scheduling Class".  This will be incorporated into the draft.
The terms in (4) are actually applicable to any device that polices an edge interface,
MPLS just happened to encounter the need for these terms first, but they're not specific

Reference [3] in the draft (RFC 2597) needs to be corrected.

An issue was raised that an ordering constraint affects packets within the same microflow;
not reordering the Ordered Aggregate is sufficient to achieve this property, but not necessary.
On the other hand working involving the term microflow seems to be a bit strong to put in a core
diffserv document.  The next version of document will have a parenthesis around the word microflow
to reopen this issue.

All other issues in this document are considered to be closed.

-- PHB ID Draft - Scott Brim, Newbridge

This is a minor update to the PHB ID draft (draft-brim-...) from Oslo.
This is a teneral purpose approach to signal PHBs (as opposed to DSCPs).
Have added a flag to indicate whether PHBID is a single PHB or a set (Bit 14).
Sets are as defined in the appropriate document, for unusual sets (e.g., 2
	of the 3 AF 1x PHBs), a list of the PHBs should be signalled.
Also added a few examples.
Standard PHBs use recommended DSCP, otherwise IANA allocates a 12 bit ID value.
	Bit 15 is set to 1 to indicate latter case.

Next step is to last call this on the mailing list.  No opposition to this in the meeting.

-- A Conceptual Model for Diffserv Routers - Brian Carpenter, WG co-chair

The authors are Yoram Bernet, Andrew Smith, and Steve Blake, but Brian is presenting due
	to absence of Andrew and Steve

Changes since previous draft:
	- New Glossary
	- Allow all model elements to be associated with both ingress and egress interfaces
	- Queues are first class elements in the model
	- Routing core function characterized: zero loss/zero delay in conceptual model
		Actual loss/delay gets reflected in traffic conditioning elements of conceptual
	- MPLS LSRs acting as Diffserv routers: discussion added, although the DS Field
		is hidden from an MPLS router.  This creates some complications.
	- Meters are now separate with more discussion of simple and multi-bucket meters.
		Appendix A contains a concrete definition of a token bucket.
	- Mux, mirror, enqueue, and null action elements added.
A mirror makes a logical copy not a physical one, mirror may need to be replaced with
	another word as part of clarifying text, "tee" is one possible replacement,
	"inverse mux" is another, "splitter" is a third.  Both of these changes need
	to be made to the text.
	- Lots of details added on queue parameters and queue sets.  Shaping queue
		also defined.
Removal of items from a queue is not modelled well.  Queues can be thought of as
having both a buffering element and a scheduling discipline that can be considered
separately.  Dan Grossman will send some suggested language to the list.
	- Traffic conditioning blocks can be composed of any elements connected in any
		fashion, and TCBs are recursive, and hence can be connected with TCBs
		and other elements to form yet other TCBs.
	- TCBs have 1 input and one output:
This is an oversight and will be removed in the next version (e.g., Figure 8 is a TCB
with one input and multiple outputs).

The overall motivation for modelling queues is to have enough detail to be able to push
rules down to routers - need enough details of how the queue works to write these rules
in a standard fashion, but don't want to specify every last implementatation detail.

A discussion ensued about specific queue parameters, such as minimum service rate.
The motivation for what was done is that min and max rate are reasonably generic,
and going beyond there gets into implementation-specific details very quickly.
These are supposed to be primitive elements (e.g., could model a complex implementation
queue as being logically composed of multiple primitive queues in the model).

Overall caution - this is not supposed to be a prescription for hardware designers.

	- New section on security considerations.
	- Appendix A is a new example.

Open Issues ...
- Appendix A token bucket definition is fine as is.
- SRTCM meter cannot be built out of a set of that sort of token buckets.  This
is an important example as it's very close to Cisco CAR.  The meters in this
draft are examples, but need to check MIB to make sure that this meter can be
modelled adequately there.  Pointing to description of this meter in the AF
RFC is an alternative to describing it here, as the cascading token buckets
example is easier to describe.

Comment: Put some restrictions on the composition rules to avoid ridiculous things.

Monitoring elements will be generalized to count without an indication of
	direction so that a monitor can track queue size. 
Discussion of specific resolution text will have to be on the list where all the 
	authors can participate.

- Should traffic conditioner blocks be allowed to contain classifiers?
	Seems to be an ok thing to do given how definition of a TCB has evolved,
	but will not be changed to allow this, as there was no compelling reason to do so.
- RSVP module: discussed on list, RSVP is only an example, hence the module needs
	a more generic name.  Functionality is to handle QoS information that is dynamically
	signalled, potentially on a relatively fine grain.  MPLS LSP setup protocol can
	also be used as an example here.

Comment: Cable Device MIB has a large amount of policy extensibility, based on classifying
	packets to an arbitrary number that then has its own set of tables.

- Should MPLS LSRs be mentioned in the model?  Yes, MPLS classifiers are an important
	part of a router implementation.

- Do class selectors require additional queue parameters?  No comments in meeting.

Open issues, especially queue issues need to be taken to list for resolution.

Additional notes on the discussion provided by Yoram Bernet:

> 1. As noted by Steve, the term 'mirroring' needs to be replaced. Suggested
> alternatives included:
> splitter, tee, tap (as in wire-tap), inverse-mux... 
> I think that the point is to use a term that makes it clear that this is a
> logical copy, not a physical copy (as in for example, multicast).
> 2. Some expressed the opinion that the abstraction of queues in terms of
> max rate/min rate is too simplistic to implement the queuing mechanisms
> that they envisioned. Others (including at least one major router
> vendor/implementor) indicated that the queuing abstraction in the draft
> enables the expression of a variety of queuing mechanisms including those
> that are most used. Suggestion to those who are not happy with the current
> abstraction - write to the mailing list. Dan Grossman offered to submit
> text on this to the list. 
> Aside: A recurring theme related to the purpose of the draft and the
> nature of indirection intended. The draft attempts to identify the *types*
> of traffic conditioning elements (e.g. meter, classifier, etc.) and to
> offer a minimal set of *sub-types* of these types, with the parameter set
> that would be appropriate. For example - a 'classifier' is a TC element
> and a BA classifier is a sub-type of a classifier that has a six-bit field
> as its only parameter. The draft *does not* state that the set of
> sub-types is exhaustive. On the contrary, it suggests that anybody can
> define a new sub-type, by specifying the parameter set that is appropriate
> (and the name of the new sub-type). The draft *offers* a minimal set of
> sub-types that is believed to be useful and necessary for realizing basic
> implementations of a diffserv router. 
> Some suggested that the draft should not define any 'sub-types'. However,
> the counterpoint presented is that without specifying a minimal set of
> sub-types, the draft would not go far enough as it would not fulfill its
> purpose of  specifying the basis of a diffserv MIB needed to implement a
> diffserv router. Since the draft allows for additional sub-types to be
> defined in other drafts, there is no harm in offering a minimal subset in
> this draft.
> 3. There was a discussion of whether the queuing parameters should include
> a. a value and a flag indicating whether that value is a max rate or a min
> rate.  
> b. two values, one for max rate, one for min rate
> Suggestion was made that 'b' is preferrable, as it would allow for the
> specification of both a max and min rate simultaneously. It would also
> allow for a special value to indicate 'unspecified' min rate and/or max
> rate.
> 4. There are words in the Glossary  (section 2) to the effect that a TCB
> has a "single input and output". This conflicts with some of the sample
> TCB diagrams and was believed by the group to be overly restrictive. A TCB
> should be allowed to have an arbitrarty number of inputs/outputs. One of
> the slides presented, suggested a 1:1 nature of a TCB. Nobody really
> understood what was meant by this. We assumed that it referred to the 1
> input, 1 output restriction, which should be removed.
> 5. There are words in the draft to the effect that a monitor 'increments'
> for every packet passing through it. This was also believed to be overly
> restrictive and the suggestion was made that it be specified as something
> that simply 'counts', with no specification as to whether it counts 'up'
> or 'down'.
> 6. There was a suggestion that we look at the 'RFC for the cable device
> MIB' for an example of traffic conditioning abstractions.
> 7. There was some discussion regarding the emphasis/de-emphasis of RSVP as
> a configuration mechanism. The following points were made:
> a. to the extent that it is mentioned, it should be cited as an *example
> of one mechanism* to effect configuration.
> b. that it should be captured as a 'qos configuration box'.
> c. that it should not simply be captured as a 'qos configuration box'
> because there is already a box titled 'diffserv config and management
> interface' and that the box labeled RSVP is sufficiently different because
> it represents a configuration mechanism that tends to be much more dynamic
> and much finer grain (per-microflow) a mechanism than represented by its
> diffserv config counterpart.
> d. that MPLS is another example of the need for a dynamic, finer grain
> configuration mechanism that justifies using two boxes to represent the
> different types of configuration.
> 8. The question was asked as to whether MPLS support required anything
> additional. Response was that the only requirement for MPLS appears to be
> that the model support a classifier sub-type for MPLS.

-- Diffserv MIB - Kwok Ho Chan, Nortel

MIB draft didn't quite make cutoff, but really needs to be discussed.

Added Andrew Smith's IP 6-tuple MF classifier.
Added info on where the classifier configuration information came from.

The latter precipitated a long discussion on whether a more generic approach
to indicating the source of arbitrary information, not just classifier
configuration information was in order.  No clear resolution, as much
of the room did not understand the issues well enough to express an
opinion.  This is complicated by the forthcoming PIB discussion in
the OPS configuration management BOF, and hence the whole matter has
been taken under advisement by the MIB draft authors.

The queueing text has been clarified to allow multiple queues/scheduling
	disciplines to be used on the same interface.
Discussion on the list determined that the Meter Pass and Meter Fail objects
	should be allowed to point to anything - the draft will be updated to
	reflect this.

Queueing issues came up here again.  Andrew Smith is expected to provide
some text on the use of minimum and maximum rates.  Subclassing to add
additional queue parameters was mentioned as a possible structure.  Van
Jacobson will send a detailed proposals to the list on:
	- This subclassing approach to queues
	- A more general approach to configuring token buckets.
	- Changing the meter specification so that the count can never be negative.

Next version will add a table to specify an interface role.  This is a policy
	WG concept that makes higher level provisioning easier to do.

Final addition is that the MIB draft needs a diagram and usage section to clarify
	relationship of tables, how they are intended to be used, and how this relates
	to the model.

Still have work to do on this document -- there are issues to be taken to the list.
	May need an interim meeting to make progress faster than the list has been,
	but should try email on list first.

-- Diffserv and Tunnels - David Black, EMC

Your intrepid scribe presented this, hence all I can do is refer to the draft.

Two conclusions:
	(1) The final slide of the presentation raised open issues.  This will
		be taken to the list for discussion before preparing a new version.
	(2) WG consensus is that this is important, and the next version of this
		draft should be a WG draft (i.e., draft-ietf-diffserv- ...).

-- Closing Remarks, Kathie Nichols, WG co-chair

Would like to focus on issues that need to be resolved to encourage diffserv deployment.

----- Tuesday Afternoon Session, November 9, 1999 -----

This is the unsolicited draft session.  These notes capture the brief resolution
of what to do with each of the unsolicited drafts, as well as salient points that
were discussed.

-- SNMP-based QoS Programming Interface MIB for Routers

Portions of this draft are candidates to be merged into the diffserv MIB, especially
the classifier objects.  The authors of this draft need to take this up with the
MIB draft authors, and additional suggestions for things to merge should be sent
to the list.

-- Rate Adaptive Shapers (draft-bonaventure-diffserv-rashaper-01.txt)

This should be discussed on the list and then may be advanced as an Informational
	RFC as complement to RFC 2697 and 2698 if the WG approves via the list.

-- draft-bless-diffserv-multicast-00.txt

Multicast issues need to be taken up at some point, but WG has other more pressing
matters that need to be attended to first.  This draft is primarily about how to
add diffserv to multicast routing, and is proposed as material to incorporate
into the diffserv architecture RFC (2475).  Needs extensive WG discussion in order
to revise RFC 2575 (Section 5).

Provisioning material should be taken up with Traffic Engineering WG, not here.

There's also some issll WG work on a multicast branch in the middle of a diffserv
	network in the intserv over diffserv work.

Need to make sure that the scope is clear before taking up as WG work item,
	but problem area is important to address.

-- draft-bless-diffserv-lbe-phb-00.txt

Less than Best Effort PHB.  Part of the proposed solution to multicast problems
identified in previous draft.  Also useful for background data, penalty boxed flows,
and deploying multimedia apps in a way that won't disrupt existing traffic.

Significant discussion occurred both in the meeting and on the list about whether
a new PHB is necessary.  The WG needs to address this set of problems, but the need
for a new PHB is an open issue that will be taken to the list.  If a new PHB is
needed, then it can be decided whether this is the right base document to start from.

One proposal was to remap DSCP 0 to be the same as Class Selector 2, thus turning
Class Selector 1 into LBE.

-- Time Sliding Window Three Color Marker (draft-fang-diffserv-tc-tswtcm-00.txt)

This is a TCP-friendly marker that works better than token bucket for TCP running
through RIO-like droppers.  Comes in 2 and 3 color versions.  There are experimental
results that will be reported in Broadcom '99 (December).

The experimental results should be made available to the WG, and then it would be
reasonable to have this draft progress to become an experimental RFC.  The authors
will post a message to the list containing URLs for all of their results.

The diffserv WG is not currently progressing traffic conditioning documents as
	standards track RFCs.

-- draft-fu-diffserv-security-00.txt

Despite none of the the authors of this draft being at the session, comments were
made that the draft has some merit because it looks at problems caused by boundaries
not working correctly - diffserv assumes that edges do work correctly and interior
nodes depend on this.  Looking at what happens when this assumption is accidentially
or deliberately false is of interest, although this draft is at best very preliminary,
and in particular needs to have the configuration protocol material removed and 
taken up with the appropriate protocol WGs.

-- draft-lin-diffserv-gtc-00.txt

The authors of this draft were not present and did not provide any presentation
material, and hence the draft was not discussed.  If they would like the WG
to do something with this draft, send the request to the list.

-- draft-nadas-diffserv-experience-01.txt

Update of a previous draft presented at the Oslo decides BOF.  This is primarily
to provide information to the WG, atlhough an Informational RFC is not the best
way to publish this sort of results.  The WG thanks the authors for their efforts
in making their results available as an Internet-Draft, and encourages others
with similar results to do likewise.

-- Expedited Forwarding with Dropping PHB

EF for wireless to deal with peculiarities in that domain.  Wireless will always drop
	packets under some circumstances.  The proposed PHB eliminates congestion
	drops, but allows drops caused by media effects in wireless.  This PHB may
	also be useful for fixed networks, to absorb otherwise unused EF capacity.

Needs more discussion on list to decide whether to advance this as a standards track
	document, experimental document, or not at all.

-- The Multimedia Color Marker (draft-medina-mmcm-00.txt)

Adaptive color marker for audio/video streams.  Modification to Juha Heinanen's three
	color marker.  Adds 3rd Token bucket to limit inbound traffic and drop proactively
	on ingress rather than reactively when problems occur.
Goal is to increase delivery probability of green packets by dropping yellow/red at
	edge node that knows the difference as opposed to somewhere inside.

An opinion was expressed that marker documents should not also discuss dropping.

The draft needs more feedback, and support from group.  Relationship to the existing
	informational marker RFCs (2697 and 2698) is particularly important to discuss
	on the list.

-- Interoperability PHB group (draft-kilkki-diffser-interoperability-00.txt)

This draft reraises a proposal to define PHBs that can be used end-to-end rather
than edge-to-edge within a domain.  This idea was rejected by the WG almost a
year and a half ago at the Cambridge, MA interim meeting, due to opposition from
major ISPs.

Representatives of two medium sized network operators opposed consideration of
this material here, with one suggesting that the general topic of inter-ISP
service interoperability belonged in an ISP industry forum, not the IETF.

-- End of Meeting Comments

Overall, the areas that seem to be of strong interest are: Multicast issues,
	Less than Best Effort PHB, and EF with dropping PHB.

Diffserv over specific link layers (primarily PHBs) should be done via changes to the
	issll WG charter, not in the diffserv WG.

Multicast needs more discussion.  In addition to LBE, a separately provisioned and
	configured EF might also be useful.

A comment was made that the diffserv WG needs to start discussion of services to avoid
to avoid inventing new PHBs for services that can be implemented by appropriate
configuration of existing PHBs.

More discussion on these topics to come on the list.