Performance of different call admission schemes in a QoS Diffserv domain

Two types of call admission strategies have been proposed for real time support in the Internet: per flow, reservation (i.e., RSVP) based, and measurement based. We evaluate these strategies for a video streaming application. We use a simulation environment (based on a Parsec platform) which supports DiffServ tools as well as QoS routing. In our experiments, we compare the performance of RA-CAC with M-CAC. We show that the selection between these alternatives is influenced by tradeoffs between throughput and delay performance; and that different classes of users may opt for one or the other technique. We also consider a hybrid situation in which both RA-CAC and M-CAC coexist in the network and are reciprocally "protected" by a weighted fair queueing (WFQ) trunk scheduler. A key enabling technology for the M-CAC scheme is QoS routing. We use Q-OSPF not only to find paths that satisfy the quality constraints (delay and bandwidths) but also to advertise current traffic, delay and available bandwidth on all network links. We also use leaky buckets for shaping and policing. Leaky bucket parameters are selected as a function of effective bandwidth, buffer space and end-to-end delay bound. If the traffic (after shaping and policing) conforms to the negotiated traffic profile, and the effective resources are allocated to the connection, the packets are guaranteed delivery within a given worst-case delay bound. In our experiments we verify that when bandwidth and buffers are explicitly allocated (with RA-CAC), the delay bounds are satisfied. The same is not true for M-CAC, where only a statistical allocation is carried out.