Streaming video continues to experience unprecedented growth. This underscores the need to identify user-centric performance measures and models that will allow operators to satisfy requirements for cost-effective network dimensioning delivered with an acceptable level of user experience. This paper presents an analysis of two novel metrics in the context of fixed capacity links: (i) the average proportion of a video’s playing time during which the quality is reduced and (ii) the average proportion of videos which experience reduced quality at least once during their playing time, based on an M/M/∞ system. Our analysis is shown to hold for the more general M/G/∞ system for metric (i), but not for (ii) and simulation studies show an unexpected form of sensitivity of metric (ii) to the flow duration distribution, contrary to the norm of increasing variance causing worse performance. At typical operational loads these new metrics provide a more sensitive and information rich guide for understanding how user experience degrades, than the widely used average throughput metric does. We further show that only the combined use of this existing and our new metrics can provide a holistic perspective on overall user performance.

Efficient fault detection and characterization are crucial requirements for automated network diagnosis systems. In this paper, we present normalized statistical signatures (NSSs), a network ‘soft-failure’ characterization technique for user devices (UDs) based upon exhaustive sets of aggregated statistical features extracted from transmission control protocol (TCP) packet streams. TCP streams are collected on-demand (e.g. upon user complaint) with fixed data transfer limits to capture the artifacts resulting from UD faults. NSSs created using live network and testbed data are able to uniquely characterize many faults and offer insight into how various types of features are affected by the faults and networks. We then introduce the link adaptive signature estimation (LASE) technique to reduce the quantities of collected NSSs required for generalized diagnostic systems with variable link parameters. We create feature estimator functions using multivariate regression techniques to generate artificial NSSs, which are subsequently used to train machine learning systems that have robust generalization capabilities. Performance of a prototype fault classifier system based on NSSs shows that an overall detection accuracy of 98% can be achieved for eight types of faults in a live network environment. In this paper, we specifically focus on formulating the basic framework of NSS and LASE, and limit the analysis to wired networks. This work can later be extended to encompass more complex fixed and mobile wireless networking environments. We expect that the combination of NSSs and LASE can serve as the foundation of next-generation automated network diagnosis systems.

We present an automated solution for rapid diagnosis of both known and unknown “soft-failures” in network User Devices (UDs). A multiclass classifier is first trained with the known faults and during diagnosis, the unknown faults are clustered to determine the existence of a new fault. Then, in an iterative process, the classifier is re-trained with the newly detected fault. The system relies on 410 features long Normalized Statistical Signature (NSSs) for fault characterization. Since, the high dimensionality of the NSS can create model overfitting, we propose EigenNSS, a transformed signature with lower dimensions and minimum information loss. The system is evaluated with live network data of 17 emulated UD faults. The results show an overall detection accuracy of 97.2%with minimum false positives and dimensionality reduction of 93.9%. Also, compared with the NSS, the EigenNSS has faster training and diagnosis times suitable for on-demand as well as real-time diagnostic applications.

This paper studies throughput metrics for elastic traffic in a beyond 3G wireless system with QoS differentiation. We extend previous work by studying the capacity for elastic traffic subject to simultaneous variation due to the loading of higher priority traffic and the radio channel. The results yield novel insights that show: (i) load dependence in determining an optimal throughput metric, (ii) call-average throughput (thetasmacr) is not affected by the radio channel variability while flow-average (Phi) and the median throughputs, as well as the sojourn time are affected, and (iii) the proximity of thetasmacr to the median over a range of typical system loads (les0.7) confirms it as the best throughput estimator for elastic traffic, except at extreme loads (>0.85) where Phi and the median converge.

The paper presents results from an analysis of traffic measurements taken at the G/sub b/ interface of a GPRS network. Data was captured in one-week long data sets between November 2002 and March 2004. The paper provides new results, based on three of these data sets, and extends our previous work (Ivanovich, M. et al., 16th ITC Specialist Seminar, 2004). The data was analysed to characterise the aggregate workload statistics, including marginal distribution and autocorrelation, and identify any trends over this timeframe. A study of models that adequately characterise the traffic was also undertaken We find that the aggregate work process (bytes "departing" the SGSN) has a marginal distribution closely matching a negative exponential; however, this does not allow us to conclude that the packet arrival or departure processes are Poisson. The "burstiness" of the traffic (i.e. the coefficient of variation) was found to become smaller over time, i.e. the traffic became smoother. This occurred due to a greater level of aggregation, as a direct result of more "low traffic" users connecting to the network. This may be indicative of an increase in WAP and mobile e-mail usage. Four potential models for the aggregate work process were investigated. The latest data shows that the trend towards a more Gaussian-type marginal distribution model continues. Of the four models studied, the Poisson Pareto burst process gave the best fit to the queueing performance compared with the recorded traffic traces. The fit to the most recent data traffic measurements has substantially improved compared with the earlier results.

We present an empirical 3-D user satisfaction model dealing with quality of service (QoS) in 2.5/3G wireless data networks. An OSI-based hierarchical framework is proposed, with mappings between performance attributes of adjacent network layers in a wireless environment. This concept is effectively a data-centric variant of a voice-service mean opinion score (MOS) to quantify user-perceived QoS, albeit in a multidimensional sense. The proposed approach is validated with a comprehensive 166-participant experiment encompassing the relationships between the top three levels of the framework: User-perceived, application- and Internet protocol (IP)-layer QoS. Finally, a wireless access technology mapping method is synthesised, aimed at giving engineers and application developers insight into the approximate bounds of an application's user-level performance for a given access technology.

Cellular radio networks supporting both circuit switched voice and packet switched data services aim to maximise the resource utilisation through efficient sharing of the resource. These sharing algorithms can use pre-emption to give voice service priority over data, thus preserving the grade of service for voice at the expense of delaying data calls. This paper develops models for a pre-emptive voice and data system, with a particular focus on GPRS. The model is based on an assumption of independence between the voice and data traffic. These models allow estimates of (i) the probability that a voice call causes pre-emption of a data call, as well as (ii) the probability that a data call will be delayed due to either being pre-empted by a voice call while in progress or due to not finding free resources at its initial attempt. The results show that a Poisson model for data traffic gives the best results for pre-emption probability when compared with measurements taken from a GSM/GPRS network.

We investigate performance improvements through TCP window size optimisation achievable when TCP Reno is used over a highly heterogeneous network, such as an 802.11b wireless LAN or a GPRS-based internet connection. Initially, our modelling focuses on a constant rate, buffered access link, based on a loss-less wireless channel and with a bandwidth at least one order of magnitude less than its fixed-network continuation into the core. We implement the model guidelines on wireless LAN and GPRS access links, confirming that TCP throughput can be increased by optimising the TCP receiver window size to the link characteristics, and more significantly, that there exists a broad range of near-optimal awnd values.

Medium access protocols for HFC and wireless ATM networks often use a collision based capacity request signalling channel which may rely on the slotted Aloha multiaccess principle. This paper studies the performance of a p-persistence slotted Aloha contention resolution algorithm (CRA), subject to extreme interstation correlation, by means of a discrete-time Markov chain analysis. We examine in detail the conditions leading to a deadlock-a situation where the time to collision resolution becomes unacceptably high and the system is practically unstable. We analyze two disaster scenario deadlock models, and study the effect of channel error probability, signalling traffic load, and the contention resolution algorithm used. We show that the key factor of the CRA is the collision rate and not channel errors. We propose and test three signalling channel capacity allocation schemes. We identify the best-performing of these three schemes as the cyclic contention mini-slot (CMS) sharing employing multiple CMSs per data slot. Finally, we demonstrate the need for implementation of an added scheme, which dynamically adjusts the p-persistence parameter.

Modern multi-service medium access protocols use a collision based capacity request signalling channel. Such signalling channels may be based on the slotted Aloha multiaccess principle. This paper studies the performance of slotted Aloha subject to extreme inter-station correlation by means of a discrete-time Markov chain analysis. We study conditions whereby the time to collision resolution becomes unacceptably high (defined as deadlock). Three signalling channel management schemes for alleviating the deadlock problem are evaluated. Of these, the cyclic contention mini-slot (CMS) sharing technique employing multiple CMSs per data slot is the one that extends the protocol's useable load region the furthest. We find that implementation of a scheme, which dynamically adjusts the p-persistence parameter towards its optimal value, is desirable. Both error free and error prone conditions are studied. The results highlight the fact that the critical signalling load is largely unaffected by the presence of errors, so that even in extremely error prone environments, the limiting performance factor is still the collision rate.

We study a new scheme for provision of priorities within the framework of a MAC protocol which is a potential candidate of the (as yet unpublished) IEEE 802.14 standard. A comprehensive simulation study, based on real traffic traces, shows that this new method provides better protection for the high priority traffic than an earlier proposed scheme. Except for physical and MAC layer overheads, the scheme exhibits comparable behaviour to that of an ideal multiplexer, for traces with very diverse correlation structures. The significance of using realistic (correlated) traffic streams for modelling purposes is demonstrated and discussed.

The paper provides insight into the performance of a candidate IEEE 802.14 MAC protocol for hybrid fibre/coax (HFC) networks, by simulation. The simulations make use of two types of input traffic: measured traffic traces, and randomly generated memoryless traffic. It is confirmed that the protocol does not reach congestion collapse under any traffic load, and this is found to hold regardless of traffic type. The paper demonstrates that memoryless traffic streams cannot be used for dimensioning guidelines as they underestimate delay as compared to real traffic traces. The paper also compares between the performance of the protocol loaded by traffic traces with that of a benchmark based on an ideal multiplexer under the same traffic conditions. It is demonstrated that both the throughput and delay performance of the protocol converge to that of an ideal multiplexer, with an increasing number of active stations.