Response for Measuring Service in Multi-Class Networks

The paper seeks to develop a methodology for inferring multi-class QoS mechanisms and parameters from the network edge. The main criticism of the reviewers is practicality and applicability to real networks, and we address different aspects of this criticism below.

(1) Multiple and heterogeneous QoS mechanisms. Reviewers 1 & 2 correctly point out that the paper does not explicitly address cases of multiple routers with heterogeneous components (e.g., one EDF router and the next WFQ). There are two cases:

i) If an analytical characterization of the concatenated components is possible, i.e., if we can derive a statistical service envelope, then our approach is extendible to multiple nodes. For example, we have derived a multi-node envelope for CJVC, core stateless -jitter virtual clock, in ICNP 2000.

ii) If the multiple elements are simply intractable, we cannot make a rigorous statement. Here, we can merely resort to the common assumption of a 'single bottleneck'. Yet, we do note that a network operator has little incentive to concatenate nodes for which the end-to-end behavior is unknown (such as the EDF/WFQ example is not solved) when well understood alternatives are available.

In revisions, we have explicitly addressed issues of multiple routers with cross traffic to make the scenario more realistic.

(2) Practical application. Reviewer 2's main concern (2.1) is that service providers will document their internal behavior to clients. While this point may indeed prove true, the premise of our work is that providers cannot be relied upon to reveal their internal behavior. Our only evidence of this is their current reluctance to reveal traces for fear of poor service being discovered and documented. It is possible that an provider claiming loss free virtual bandwidth guarantees would not want it discovered that they have over-engineered, rather than implement any particular controls.

(3) Implications. Reviewer three asks what are the actions end-points would take on the basis of the results.

This is an interesting issue with a number of possibilities. One is admission control. In fact, in our implementation of scalable admission control, ( we have used the technique so that edge routers can perform admission control with implicit control of the path. Other possibilities are client capacity planning: if the client knows whether it is strictly policed to an SLA bandwidth, or allowed some borrowing, the client can better determine its future usage needs.

We have added a reference to the above implementation paper and improved the discussion of applications in the introduction.

1.A. is addressed in (1).

1.B. We added more detail on the discussion of backlogging and precisely describe how to determine backlogging with cross traffic.

1.C. The role of time scales is a theme throughout, namely that some mechanisms are best inferred at small time scales and others at larger ones. We changed the quoted text to be the more precise "If only a single time scale is used to infer the scheduler, erroneous answers can be obtained. However, by using majority rule over all time scales, the correct inference is attained 94\% to 100\% in our experiments."

2.1 is addressed in (2) and 2.2 and 2.3 are addressed in (1)

3.1 is addressed in (3).

3.2 We agree that the Gaussian approach is applicable in high aggregation regimes, and also agree that the technique is extensible to the non-Gaussian case. In particular envelopes can be derived using large deviations theory [5] and higher moment measurements incorporated into estimations. We revised the text to explicitly make this point.

3.3 We hadn't thought of this, but agree it is worthwhile and believe it can be achieved. We are pursuing this and hope to include it in the final version if finished in time.