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

Re: [PCN] CLE vs. admission state



Hi all,

I summarize the issue I raised during the spec writing process. It would be good to keep either the PCN egress or the decision point / PCN ingress unaware of CL or SM is used. This automatically leads to the same signalling protocol for CL and SM, too. Benefits of such an approach are obvious (one signalling protocol for CL/SM, simplified migration from SM to CL). This can be achieved in two different ways.

A) egress-based-PCN-calculation

The egress calculates a PCN-admission-state (admit/block) as well as the edge-to-edge supportable rate (ETM traffic in CL, u*ThM traffic in SM) and sends these values to the decision node. The decision node can do its operation (admit/block traffic, measure ingress PCN traffic and possibly terminate PCN traffic) without knowing whether the PCN reports from the egress node are based on CL or SM.

B) decision-point-PCN-calculation

The decision-point (ingress) calculates the PCN-admission-state and the edge-to-edge supportable rate. To that end, the egress transmits the measured rates of NM, ThM, and ETM traffic in regular intervals.


Advantages of egress-based-PCN-calculation (A):

1) The PCN calculation algorithm can be exchanged locally at the egress node without changes to the signalling protocol and the ingress node. Example: the egress may send "block" immediately when it observes a marked packet. It is basically a different edge behavior that is able to work with smaller ingress-egress-aggregates. Reason: small ingress-egress-aggregates need long measurement intervals for meaningful rate estimates and then it is good when a block-message can be sent before the end of the measurement interval.

2) If the PCN-admission-state does not change at the egress and the edge-to-edge supportable rate is zero, regular reports which are usually carried every 100 ms from egress to ingress can be suppressed until the next change. The saved signalling overhead can be significant when the rate of ingress-egress-aggregates is small.



Advantage of decision-point-based-PCN-calculation (B):

3) In case of SM, the CLE increases proportionally with the load. It is possible to use different CLE-admission-thresholds for high- and low-priority traffic so that high priority traffic has a lower blocking probability than low priority traffic when the system is near to pre-congestion.


Below the line, the question whether A or B is better boils down to the issue what's more important for PCN under the assumption that we want to keep either the ingress or egress unaware of CL/SM:

i) Potential support for smaller ingress-egress-aggregates

ii) Differentiated blocking probabilities for low- and high-priority traffic (only for SM)

Basically we need to decide which of the two issues is more important for PCN.

Regards,

   Michael


Tom Taylor schrieb:
The current versions of the edge behaviour drafts have stepped back slightly from the extreme of reporting only transitions in admission state. Now admission state is reported at regular intervals.

A fair amount of off-list discussion has been happening over the appropriateness of the current arrangement. For one, there is a certain amount of opposition to specifying exponential smoothing of the CLE. For another, Phil has raised the possibility of a CLE-dependent policy at the decision point, where more important flows are allowed in at higher CLE values than less important flows. That suggests that instead of admission state, the egress node should be reporting at least the CLE.

I believe the discussion should be happening on the list -- that's what it is for.

Michael Menth suggested an useful principle to which the present drafts adhere: actions at the decision point/ingress node should be exactly the same for CL and SM. He argued that the egress node unavoidably knows the difference between CL and SM in any case, so it might as well absorb all the differences.

It is true that at least one end has to know the difference, but it doesn't have to be the egress node. If the egress node simply reports rates of unmarked, threshold-marked, and excess-traffic-marked traffic each interval, the decision point/ingress node can apply the CL- or SM-specific algorithms and draw the correct conclusions.

It seems reasonable to design the system so one end is the same whether CL or SM is deployed. If we accept that, we have one-and-a-half questions really:

1) Which end should know that SM or CL is deployed?

1) a) If it is the egress node, should it report admission state or CLE? Either way, the egress node would have to report the estimated edge-to-edge supportable traffic rate whenever termination might possibly be required, and the decision point/ingress node would have to compare that with the admitted traffic rate to see if termination is really needed.

Tom
_______________________________________________
PCN mailing list
PCN at ietf.org
https://www.ietf.org/mailman/listinfo/pcn

--
Dr. Michael Menth, Assistant Professor
University of Wuerzburg, Institute of Computer Science
Am Hubland, D-97074 Wuerzburg, Germany, room B206
phone: (+49)-931/31-86644 (new), fax: (+49)-931/888-6632
mailto:menth at informatik.uni-wuerzburg.de
http://www3.informatik.uni-wuerzburg.de/research/ngn