Cool Tool: Diana Oglesby NSLP Nancy Masten Great Lakes 2006 NCHELP FALL TRAINING CONFERENCE.
NSLP for Quality of Service
-
Upload
nicole-morton -
Category
Documents
-
view
46 -
download
1
description
Transcript of NSLP for Quality of Service
NSLP for Quality of Service
Sven Van den Bosch (ed)Georgios Karagiannis
Andrew McDonald(et al)
draft-ietf-nsis-qos-nslp-03.txt
Changes from –02 version Addressed comments from early review Added text on receiver-initiated and bi-
directional reservations Extended description of session binding Added support for fate sharing Restructured message formats and processing
section Clarified refresh reduction mechanism Added assumptions on QoS model Added assumptions on operating environment
Receiver-initiated reservation Current proposal
Use QUERY to set up reverse path state Use QUERY to gather path information (OPWA) Use QUERY to refresh reverse path state
Question 1: Does QUERY need to carry QSPEC? It is optional in case OPWA is not needed
Question 2: Does QUERY need to carry RESPONSE_REQUEST?
It is not used when QUERY acts as an RSVP PATH message If ‘no’ on both questions
Do we need a separate (NULL) message for this ‘empty’ QUERY? Trade-off between QUERY complexity and additional message type
Is the NULL message generally useful across NSLPs? Question 3: Should GIMPS be responsible for refreshing
reverse path state?
Bi-directional reservation Current situation: two supported mechanisms
Sender-initiated reservation+Receiver-initiated reservation Two sender-initiated reservations
What happens when following conditions apply One of the end nodes does not have sufficient information,
and (some part of) the network does not install reverse path state
Question: Do we want to provide a solution for this situation?
Proposed solution: Carry necessary information (opaquely) in forward direction
Bundling of NSLP messages Provide indication to wait for subsequent NSLP messages
before sending?
Session binding (example) Aggregate reservations
If session B is torn down then session A may be torn down as well but not vice versa
QNI QNE QNE QNR
End-to-end session= ‘binding session’SESSION_ID = A
End-to-end session= ‘binding session’SESSION_ID = ABOUND_SESSION_ID = B
Aggregate session= ‘bound session’SESSION_ID = B
Session binding (example) Bi-directional reservations
X QNE QNE Y
End-to-end session (XY)SESSION_ID = ABOUND_SESSION_ID = B
End-to-end session (YX)SESSION_ID = BBOUND_SESSION_ID = A
Special ‘refresh’ cases New message with same SESSION_ID and
different MRI Default behaviour: Reservation is replaced Exception: NO_REPLACE flag set
Enter into resource sharing cases New message from a bound session
Default behaviour: all binding sessions share fate Exception: NO_FATE_SHARING flag set Only end/edge points use fate sharing information
Resource sharing Current situation
Resource sharing applies to sessions with same SESSION_ID, different MRI and NO_REPLACE flag set
Resource sharing is requested by QoS NSLP processing or RMF; Required information is contained in QSPEC
Question 1: Should it also apply to bound sessions? [Yes]
Question 2: What mathematical operations are useful on two or more reservations?
ADD, SUBTRACT, … Question 3: Do any of these have impact on QoS NSLP?
If yes, the impact is independent of session binding
Reserve/commit functionality Alignment needed
QoS NSLP: qualitative (commit flag) QoS model: quantitative (start/stop timing)
Is this a QoS NSLP issue anyway?
Priority Mailing list discussion
Reservation priority (preemption) is not a QoS NSLP function (see section 4.5)
Message priority is in scope for the QoS NSLP but relies on GIMPS (see section 7.7)
Question 1: Required number of levels for message priority?
Question 2: Is reservation priority applicable to different NSLPs? Should there be an generic NSLP priority object?
Refresh overhead reduction Current proposal:
Insert RESPONSE_REQUEST (to confirm state installation)
Refer to reservation with SESSION_ID and RSN So, still one refreshing RESERVE per
reservation But smaller and possibly easier to process
Question: Should we be able to send a RESERVE without MRI (and only pass MRI over the API)?
Mailing list Issue on use of SCOPING in RESPONSE
Need to clarify global RII significance versus local RSN significance
Proposed solution Use SCOPING only in QUERY/RESERVE make RII a separate object, carried in
QUERY/RESERVE and RESPONSE
Next steps Implement interim meeting
outcomes Complete
Error codes AAA Security