Abstract

Abstract

Providing service guarantees requires strict admission control to ensure that the resources are not over-allocated. On another hand, limiting the core nodes' role has been recognized as a key aspect in providing scalable Quality of Service (QoS) solutions. This paper presents a scalable multiple-token mechanism that ensures a distributed and coordinated resource allocation along the logical ring of the ingress/edge nodes. The proposed mechanism alleviates the scalability problem by keeping all the admission control operations in the ingress/edge nodes while preserving guaranteed service semantics. By distributing the available bandwidth of the links over multiple tokens, increased robustness and low latency under light load conditions are achieved, while the zero-false positives property is maintained. The erroneous connection denials, which could be due to the distribution of available bandwidth information, are avoided by an accumulation procedure that concentrates the available bandwidth information whenever necessary.

SUMMARYIn order to achieve a quality of service (QoS) capable of satisfying an ever increasing range of user requirements, differentiated services (DiffServ) have been introduced as a scalable solution that emerges 'naturally' from today's best effort service approach. Mapping the packet treatment into a small number of per hop behaviours (PHBs) is the key idea behind the scalability of DiffServ but this comes at the cost of loosing some behavioural differentiation and some fairness between flows multiplexed into the same aggregated traffic. The paper proposes a novel simple and effective DiffServ approach, the 'Simple Weighted Integration of diFferentiated Traffic' (SWIFT), and uses it in a series of simulations covering a relatively wide range of local network conditions. Measured voice and video traffic traces and computer generated self-similar background traffic were used in simulations performed at various congestion levels and for in-profile and out-of-profile source behaviour. The resulted throughput, mean delay, maximum delay and jitter are used to asses SWIFT's capabilities}isolation of the in-profile traffic from congestion effects, treatment differentiation, increased resource utilization, fairness in treatment under congestion, and incentivity for nice behaviour. Comparisons with other approaches employing traffic control are also provided.