تجزیه و تحلیل عملکرد نهایی با تجمع ترافیک

The provisioning of end-to-end services in the presence of traffic aggregation is evaluated in the framework of the Differentiated Services (diffserv) QoS architecture. The effect of stream multiplexing on delay and jitter-sensitive traffic is analyzed through an experimental approach. Aggregation is evaluated in three different scenarios: with different aggregate traffic loads, with a variable number of flows multiplexed in the same class, and with different packet sizes. Two different scheduling algorithms are considered: Priority Queuing (PQ) and Weighted Fair Queuing (WFQ).

The Differentiated Services (diffserv) QoS architecture, specified in RFC 2475, is based on the aggregation of streams into classes and on the provision of QoS to the class instead of the single flow [1], [2], [4] and [5]. The per-class packet treatment is defined by the Per-Hop Behavior (PHB), which describes the packet handling mechanism a given class is subject to [3]. Packets are classified by access routers or hosts through traffic filters: the class to which a packet is bound is identified by a specific code in the diffserv field called DSCP (DiffServ Code Point) as specified in RFC 2474. Classification and marking are performed at the ingress of a given domain. Only edge routers have access to the policy for the binding between packets and classes. Core diffserv nodes only manage classes; i.e., no per-flow information is deployed in the core. This is to move the network complexity from the core to the edge. In this paper, we focus on one aspect of diffserv: the provision of end-to-end services in the presence of traffic aggregation. While the mixing of different streams into one class is an inherent property of diffserv and is fundamental for improved scalability of the architecture, the interference between packets of multiple flows – that are treated in an undifferentiated way within a class – can have an influence on the end-to-end performance of the single stream. While aggregation is of relevance to diffserv, this issue does not arise in QoS architectures, such as ATM and intserv [7], [8] and [9], that provide finely grained QoS by supporting quality of service to the flow: both the stream and the reservation profile are advertised through a signaling protocol so that streams are treated independently. The problem of per-flow performance under aggregation is a topic that still needs more research for better understanding. In this work, we address the problem through an experimental approach by running tests over a diffserv test network. Given a reference stream to which measurement applies, the end-to-end performance of one stream in the class is evaluated in different scenarios: for different aggregation degrees, different loads, and different packet sizes. In order to keep the end-to-end behavior of a flow in accordance with the contract, policing, shaping or other forms of traffic conditioning may be adopted. In this work, we call stream isolation the capability of a network to preserve the original stream profile when it is transmitted over one or more diffserv domains in an aggregated fashion. The maximum degree of traffic isolation is similar to that achieved when a dedicated wire carries each stream. In this work, we estimate the traffic isolation supplied in different traffic and network scenarios in order to evaluate the feasibility of the QoS architecture being tested and to identify the conditions for maximum isolation.

The study of the effect aggregation has on end-to-end performance is fundamental in evaluating the capability of a diffserv network to preserve the original stream profile, in particular of delay and jitter-sensitive traffic, and to evaluate the need of traffic conditioning mechanisms like policing and shaping to limit the disruptive interaction between streams within a given class. Experimental results conducted in the presence of EF aggregation (without traffic conditioning) show that the limitation of the EF load and of the aggregation degree and the presence of small packet sizes greatly contribute to the minimization of the occurrence of occasional bursts. Burstiness is a linear function of class load. However, while the increase in the EF queue size – needed to store traffic bursts – has a minor effect on one-way delay, IPDV is clearly minimized in the case of limited load, and even in this case, IPDV distribution is spread in a larger range of values, and a few packet samples still experience high IPDV. Burstiness is proportional to the number of EF streams: it increases rapidly when the number of streams varies in the range 1–8, but then quickly converges to a stable value. WFQ is less prone to burstiness if the EF departure rate is approximately equal to the arrival rate, while it converges to PQ when the EF queue weight and, consequently, the corresponding service rate increase. With WFQ, a tradeoff between one-way delay minimization and burstiness avoidance has to be identified.