End-to-end performance evaluation of datagram acceptance control in DQDB-ATM-DQDB CL network

Abstract

In this paper we extend a buffer overflow control scheme called datagram acceptance control (DAC) to shared buffers, then evaluate its end-to-end performance and compare it with another two methods. To manage CL traffic in an interconnected CL overlaid network operated in on-the-fly mode, one can use loss mechanisms by discarding cells at a congested buffer. However, since the size of a typical datagram is larger than the payload size of a cell, cell discarding without taking into account the integrity of a datagram can produce corrupted datagrams. As a corrupted datagram is dropped at its final destination, the network resources used on its route are wasted. Moreover, it is observed in our studies that when the network is overloaded even with a simple buffer overflow control scheme the waste still can be a large portion. Thus, cell discarding at congested buffers has to be treated with the consideration of concentrating discarded cells upon as few datagrams as possible to achieve reducing the waste but not at the expense of the overall datagram loss ratio. It is shown by our simulation results that DAC, by managing a buffer together with its associated outgoing bandwidth, can achieve the stated objectives end-to-end. It is also observed in our studies that increased buffer size can only improve the datagram loss ratio for load levels up to 100%. Once overload occurs an eightfold/sixteenfold increased buffer size, which is 10 times of the mean frame length, has virtually no impact on the datagram loss ratio. Thus, losses due to overload cannot be avoided by large buffer capacity.