FLOW AGGREGATION AND BUFFER MANAGEMENT IN MULTI-SERVICE INTERNET

Abstract

This thesis describes a scheme that solves the scalability problem of integrated services and Resource Reservation Setup Protocol (RSVP) by: (a) aggregating RSVP reservations in the backbone network where the traffic load is potentially high, (b) providing a simple and fast forwarding mechanism for routers on the high speed backbone. In particular, flows that share the same ingress router, egress router and service class are grouped into a flow aggregate, thus reducing the number of RSVP states. Aggregated RSVP is used to make reservations on behalf of these flows. An aggregation region is a network that is capable of aggregating data flows. Within an aggregation region, four traffic classes are defined to support four types of service, namely guaranteed, controlled-load, assured and best effort services. Each class is implemented as a separate queue with Random Early Detection-(RED) like mechanism and scheduled by priority based queueing. Data flows are assigned to the traffic classes according to their network resource requirements as well as their willingness to pay. An edge router, which is placed at the link between an RSVP capable stub network and an aggregation region, translates the RSVP flows into one of the traffic classes by marking the DS fields of their data packets. Policing is performed by the ingress border router, which sits at the boundary within an aggregation region, to ensure incoming flows do not exceed their allocated bandwidth. Interior routers in an aggregation region make forwarding decision based on the DS field rather than the source address and port number, destination address and port number, and protocol number.

RSVP messages of individual flows are transmitted transparently through the aggregation region. In place of these per-flow RSVP messages are aggregated RSVP mes­ sages. A method is proposed to hide the individual RSVP messages from the interior routers so that they are not burdened by message processing. To ensure that the aggregated RSVP messages traverse the same path as the data flows, we propose to use minimal encapsulation for the aggregated flows. This method has the added advantage of being able to adapt to a route change automatically using refresh timeout, without resorting to route query. Besides that, discovery of an egress router by an ingress router is done by simple extensions to RSVP. This thesis shows that the proposed packet scheduling and buffer management scheme is able to provide bounded delay and rate guarantee. The buffer size needed for loss­ less service is also derived. Furthermore, the proposed scheme has low implementation complexity and operational overhead. Clumping of packets, which is a side effect of any aggregation scheme, is eradicated by employing reshaping. This also reduces the buffer requirement and does not increase the maximum delay experienced by the packets. As with any aggregation scheme, there are some penalties associated with simplification. Making an aggregated reservation for a group of flows may result in the inefficient usage of bandwidth. By mixing traffic that can tolerate occasional long delay with traffic of guaranteed service it is possible to improve the bandwidth utilization. Proper demarcation of an aggregation region is important to prevent overloading the border routers. If the flows are short-lived, the proposed scheme cannot reduce the amount of RSV messages in the aggregation region significantly.