Antonis Dimakis

Sufficient conditions for stability of longest-queue-first scheduling: second-order properties using fluid limits

We consider the stability of the longest-queue-first scheduling policy (LQF), a natural and low-complexity scheduling policy, for a generalized switch model. Unlike that of common scheduling policies, the stability of LQF depends on the variance of the arrival processes in addition to their average intensities. We identify new sufficient conditions for LQF to be throughput optimal for independent, identically distributed arrival processes. Deterministic fluid analogs, proved to be powerful in the analysis of stability in queueing networks, do not adequately characterize the stability of LQF. We combine properties of diffusion-scaled sample path functionals and local fluid limits into a sharper characterization of stability.

Providing bandwidth guarantees over a best-effort network: call-admission and pricing

This paper introduces a framework for answering questions regarding the conditions on the network load that allow a best-effort network like the Internet to support connections of given duration that require a certain quality of service. Such quality of service is expressed in terms of the percentage of time the bandwidth allocated to a connection may drop below a certain level or the maximum allowable delay in placing the call through the network waiting for more favorable loading conditions. The call-acceptance conditions, which depend on the behavior of the system over the lifetime of accepted calls, are thus based on transient models for the congestion (instead of looking at the average behavior) and attempt to exploit the time-scales of the fluctuations of the number of connections competing for bandwidth. Extensions of the model consider the case of dynamic pricing which allows connections that pay more to get larger shares of the bandwidth, and investigate the trade-off between quality of service, the size of the acceptance region, and the charge to be paid by the connection. Under this framework we introduce an option contract that reduces the risk of quality disruption, if a user has a fixed budget at his disposal, and calculate its price. One potential use of this methodology is towards developing a simple admission control mechanism for placing voice calls through an IP network, where the decisions can be taken by edge devices.

Mechanisms for Efficient Allocation in Divisible Capacity Networks

We propose a mechanism for auctioning bundles of multiple divisible goods. Such a mechanism is very useful for allocation of bandwidth in a network where the buyers want the same amount of bandwidth on each link in their route. We first propose a single-sided VCG-type mechanism. However, instead of reporting types, the players only reveal a two-dimensional bid signal - the maximum quantity that they want and the per unit price they are willing to pay. We show the existence of an efficient Nash equilibrium in the corresponding auction game of the mechanism. We show through an example that not all Nash equilibria are efficient but provide a distributed algorithm that yields the efficient one. Further, we provide a sufficient characterization of all efficient Nash equilibria. We then present a double-sided auction mechanism for multiple divisible goods, and show that there exists a Nash equilibrium of the auction game which yields the efficient allocation

Traffic equivalence and substitution in a multiplexer with applications to dynamic available capacity estimation

For a multiplexer fed by a large number of sources, we derive conditions under which a given subset of the sources can be substituted for a single source while preserving the buffer overflow probability and the dominant timescales of buffer overflows. This notion of traffic equivalence is stronger than simple effective bandwidth equality and depends on the multiplexing context. We propose several applications of the above traffic substitution conditions. First, we show that fractional Brownian motion as a single source substitute can effectively model a large number of multiplexed sources using information obtained purely from traffic traces; this has direct application to simple but accurate traffic generation. Second, we focus on dynamic (i.e., on-line) estimation of available capacity and buffer overflow probability. This requires the solution of a double optimization problem expressed in terms of functions whose values are obtained from time averages of the traffic traces over a large range of timescales. We show how to solve this problem on-line by reducing it to the calculation of a fixed-point equation that can be solved iteratively by combining traffic substitution using fractional Brownian motion with dynamic measurements of the actual traffic. We have validated this approach by extensive experimentation with large numbers or real traffic sources that are fed to a high bandwidth link, and comparing our on-line estimation of available capacity and the resulting dynamic call admission control with other existing approaches. The superior accuracy of our approach also suggests that taking the buffer size into account, as does our on-line algorithm, may be vital for achieving approximations of practical interest.

Adaptive Quality of Service for a Mobile Ad Hoc Network

This paper presents a QoS routing system for MANET supporting multiple traffic classes. The system takes into consideration clustering and channel allocation. Simulation experiments show that our algorithms are convergent. The system also yields a higher total throughput compared to the case in which every interface uses the same channel.

Fair background data transfers of minimal delay impact

In this paper we present a methodology for the design of congestion control protocols for background data transfers that have a minimal delay impact on short TCP transfers and compete for a target share of the leftover average capacity with other background TCP transfers. We analytically compute the optimal policy and show that its delay performance can be well approximated by a weighted TCP policy, which maintains a target proportion of TCPs bandwidth at all times in order to achieve the same share of excess capacity. The relative approximation error is always less than 17.2% while it quickly decreases to zero as the number of coexisting background TCP flows increases. Next, we consider a general utility-based fairness criterion for sharing the leftover average capacity, including a penalty term capturing the delay impact to short flows. The criterion is jointly optimized over all allocations of excess capacity to background flows (including TCP ones) on long timescales, and all bandwidth sharing policies on fast time scales. Even though the delay optimal sharing policy that solves the above optimization problem does not lead to distributed congestion control algorithms and more significantly, requires knowledge of the number of competing background TCP flows, both problems disappear under the weighted TCP policy. A distributed weight adjustment policy is considered where, at equilibrium, the overall performance is nearly optimal, with a vanishing relative optimality error as the number of background TCP flows increases. We illustrate the methodology by giving two examples of congestion control algorithms for background transfers. Both achieve low delay for short flows relative to TCP, but at the same time they present strong incentives for adoption against incumbent low priority solutions in public environments.

Energy-Aware Pricing Within Cloud Environments

The Adapting Service lifeCycle towards EfficienT Clouds (ASCETiC) project aims to provide novel methods and tools to support software developers aiming to optimize energy efficiency resulting from designing, developing, deploying and running software at the different layers of the cloud stack architecture, while maintaining other quality aspects of software to meet the agreed levels. The Pricing Modeler is a component within the ASCETiC architecture, which is responsible for the price estimation and billing of cloud applications or Virtual Machines (VMs) based on their energy consumption. In this paper, we propose a set of novel energy-aware pricing schemes implemented within the Pricing Modeler component, as well as a set of envisaged service plans which aim to facilitate the gradual adoption of the ASCETiC architecture.

Economic Implications of Energy-Aware Pricing in Clouds

Cloud computing is a promising approach for delivering ICT services by improving the utilization of data centre resources. One candidate solution for accomplishing energy efficiency within clouds is the adoption of energy-aware pricing by the cloud service providers. In this paper, we compare the economic implications of the choice of pricing schemes under different scenarios.

Optimizing queueing cost and energy efficiency in a wireless network

We consider the joint minimization of queueing cost and power consumption in a wireless network, over all power control policies at the transmitters. Approximately optimal policies are derived through the use of a comparison to a single link system offering lower bounds for both performance metrics. The derived policies are of a simple form: the total power consumed at any instant is a monotonic function of a linear combination of the queue backlogs at that instant, while the portions of power allocated to each transmitter are such that the resulting link speeds are proportional to the queue backlogs. Simulations show that the lower bounds become tight -and so the policies become optimal- as the average power approaches the lowest value for which stability is maintained. Moreover, these policies are shown to perform better than gradient projection -based power adaptive algorithms such those arising in formulations using static optimization problems.

Congestion Control for Background Data Transfers With Minimal Delay Impact

Congestion control protocols for background data are commonly conceived and designed to emulate low priority traffic, which yields to transmission control protocol (TCP) flows. In the presence of even a few very long TCP flows, this behavior can cause bandwidth starvation, and hence, the accumulation of large numbers of background data flows for prolonged periods of time, which may ultimately have an adverse effect on the download delays of delay-sensitive TCP flows. In this paper, we look at the fundamental problem of designing congestion control protocols for background traffic with the minimum impact on short TCP flows while achieving a certain desired average throughput over time. The corresponding optimal policy under various assumptions on the available information is obtained analytically. We give tight bounds of the distance between TCP-based background transfer protocols and the optimal policy, and identify the range of system parameters for which more sophisticated congestion control makes a noticeable difference. Based on these results, we propose an access control algorithm for systems where control on aggregates of background flows can be exercised, as in file servers. Simulations of simple network topologies suggest that this type of access control performs better than protocols emulating low priority over a wide range of parameters.