Reviewers Familiarity: 9 Evaluation of work and contribution: 7 Significance to theory and practice: 6 Originality and novelty: 5 Relevance for NET-COOP: 7 Readability and organization: 7 Overall recommendation: 7 **(0 = no mark set by reviewer)** Summary: This paper is really about an online scheme for controlling a queueing parameter $t$ using utilisation as an observed quantity and trading off delay against utlisation, with a form of welfare maximisation as an objective. The paper is clearly written, although there are one or two minor errors (eg grammatical). The paper is a different slant on familiar topic, and interesting from this point of view. Ns simulations give some insights, however it is difficult to draw conclusions from these small experiments, with stable and synthetic traffic traces Comments to the author: This assumes there is a clear link between the queuing parameter (eg buffer size) and delay / utilisation. It is not clear the relationship is so obvious in real networks: eg, a certain amount of buffering is needed to cope with "packet-level" -effects , namely small bursts of back-to-back pkts arriving at different input ports of a switch; beyond this, adding further buffering has little effect on "burst-scale" effects, which occur on large timescales. If $t$ relates to buffering, given that this is fixed in a real system, you would be adjusting within the max and min values. At present, is there a real cost to the providers (rather than users) of delay?? There is a real-time feedback effect on users and elastic traffic caused by delay, which it is not easy to capture, in simulation or in the model. Bringing fairness in (Simulation 4) is off-tract wrt rest of the paper, especially since TCP has inbuilt RTT bias. Have you looked at any of the implentation issues? There are hourly variations of arrival patterns. --------------------------------------- Reviewers Familiarity: 9 Evaluation of work and contribution: 5 Significance to theory and practice: 5 Originality and novelty: 5 Relevance for NET-COOP: 7 Readability and organization: 9 Overall recommendation: 5 **(0 = no mark set by reviewer)** Summary: The paper formulates the utilization/delay trade-off as a "benefit" maximiation problem. Under the assumption that the objective function is concave, the paper proposes and analysis an online algorithm to maximize the benefit. The key observation of the paper is that network traffic conditions do not change too much on a time-scale of seconds, and that estimates of the average delay/utilization over this time scale are meaningful. The theoretical contribution of the paper is marginal. Under the assumption that the objective function is convex and "good" estimates of the average delay/utilization are available, then the optimization scheme is rather straight-forward. The contribution of the paper is to show through simulation that the network traffic indeed does not seem to change too much over the time-scale of seconds, and that "reasonably good" estimates of the average delay/utilization can be obtained. Comments to the author: --------------------------------------- Reviewers Familiarity: 4 Evaluation of work and contribution: 6 Significance to theory and practice: 6 Originality and novelty: 5 Relevance for NET-COOP: 7 Readability and organization: 7 Overall recommendation: 7 **(0 = no mark set by reviewer)** Summary: Defining benefit as the difference between utilization and cost associated to queuing delay, the authors present a generic algorithm ODU-t (Optimal Delay-Utilization control of t) which is maximizing benefit. Their approach is easy to follow and interesting, complemented by numerical simulations. Comments to the author: Some typographical errors must be corrected in the final version, e.g.: "the problem with employing these strategy", "vise versa", etc.