Attaining per flow QoS with Class-based Differentiated Services

The Differentiated Services (DiffServ or DS) framework takes an edge over IntServ because it is scalable and lesser complex. On the other hand, the application level end-to-end quality o f service, in DiffServ, may get compromised because: 1) network resources are not allocated at microflow level (a data stream pertaining to a single connection) but at aggregate level (collection of one or more microflows), 2) the DiffServ working group does not specify algorithms for PHBs but their output behavio urs and 3) end-to-end quality is function of Service Level Agreements (SLAs) between the adjacent domains transited by the connection and a large diversity in SLAs is quite evident as each DS domain would have different service provision policies. We focus, in this paper, on the first two issues. Our goal is to ha ve DiffServ deployed with all its simplicity and still be able to provide application level end-to-end qu ality of service. For that, we study a PHB for AF classes. A PHB comprises a packet scheduler and a packe t accept/discard algorithm. For packet scheduler, we use the Extended-VirtualClock (Ex-VC) algorithm. Ex-VC perfo rms delay-based service differentiation among the aggregates while selecting a packet for service. The reasons for having delay-based definition for service differentiation are: it is adaptable to load per aggr egate and it does not need to be microflow aware. Other definitions like bandwidth and loss may also be u sed but the former requires microflow aware management and the latter lacks in simplicity. For packet accept/discard algorithm, we use RED when all packets have the same drop precedence level and DI-RO (Deterministic for In-RED for Out) when packets are policed at the ingress DS node and packets viol ating the Service Level Agreement (SLA) are marked OUT. In DI-RO, IN packets are always accepted (except buffer overflow) whereas OUT packets are accepted probabilistically.