An evaluation of scheduling mechanisms for providing best-effort real-time communications in wide-area networks

Abstract

W e distinguish between four types of service that may be provided to real-time trafic b y packet-switched networks, ranging from "need-blind" and "need-based best-effort'' t o "guaranteed throughput" and "bounded delay jitter" services. W e evaluate a number of scheduling policies that offer need-based, best-effort service. W e introduce h o p -l a d y (HL) scheduling which is based on the time remaining until the packet must reach its destination as well as the number of hops separating it from the destination. H L scheduling is evaluated through simulation and has been implemented within a BSD-based kernel and tested on the DARTnet network. The results indicate that H L scheduling tends t o equalize delays between calls with large and small number of hops as compared to a FIFO dzscipline, reducing the 99.9% percentile of delay and the fraction of late packets. W e compare H L scheduling to the FIFO+ discipline suggested b y Clark et al.., and $nd that their delay properties are similar. Other disciplines, such as minimum laxity or transit priority, may actually do more harm than good.

Deadline-based scheduling in support of real-time data delivery

“…Simulation results showed that HL had the best performance in terms of the fraction of packets that are delivered on-time. These results were confirmed with an implementation of HL in a BSD-based kernel and its experimentation on the DARTnet network [11].…”

“…We will refer to this algorithm as T/H. T/H has also been studied in the context of packet-switched networks [9,10,11].…”

Type-of-service routing in datagram delivery systems

“…0, which imply the following necessary (but not su cient) conditions for convergence to the origin: 1 x N^0 y Mg. See Figure 3. 1 De ne the scalar function V ( x; y): V ( x; y) = x 2 + y 2 (11) We show that V ( x; y) is a Liapunov function in D 1 , which implies that starting from any point in D 1 , the trajectory stays inside D 1 , and converges to the origin.…”

“…That is: V ( x; y) is positive de nite; and for all ( x k ; y k ) D 1 ? f(0; 0)g, 4V ( x k ; y k ) < 0 and ( x k+1 ; y k+1 ) D 1 : 2 1 Although the equation of the left boundary of the domain D 1 is quadratic in x and y, the resulting solution turns out to be a straight line because one of the roots is infeasible. This is con rmed by montecarlo simulations.…”

Network Support for Turn-Taking in Multimedia Collaboration

“…Fairness is essentially a product of the underlying distributive mechanism for oors, corresponding to the transport mechanism and a speci c oor policy. 4 A policy 26 is implemented by a speci c ordering in the service of requests, together with revocation conditions and timeouts. It may support queueing and fairness conditions through request reordering, deferred execution, and preemption of activities.…”