Comparative performance of interleaved and non-interleaved pipelining in ATM terminal adapters

Abstract

In this paper we compare the end-to-end delay performance of two service disciplines that an Asynchronous Transfer Mode (ATM) multiplexor can use to multiplex pipelined synchronous native protocol frames arriving over low speed access lines onto higher speed ATM trunks. The ATM multiplexor must convert native protocol frames on the access lines (typically hundreds of bytes in length) into ATM cells on the trunks (of fixed length 53 bytes) at the network access and the corresponding ATM (de)multiplexor must reassemble these multiple ATM cells into a bit continuous synchronous native protocol frame at the network egress. In one service discipline, the multiplexor (hereafter referred to as a terminal adaptor, or TA) pipelines one synchronous frame at a time onto the high speed trunk, waiting, if it needs to, for the successive ATM cells from this one synchronous frame to accumulate. This scheme introduces a larger queueing delay at the ingress TA but no playout delay is required at the egress TA. In the other scheme the TA interleaves the ATM cells from the various synchronous frames arriving concurrently over the low speed lines. An interleaving scheme has a smaller queueing delay, but introduces a large variability in the network insertion times of the ATM cells corresponding to a native protocol frame. Hence the egress TA must enforce a delay before playing out the original bit-synchronous native frame over a low speed line, so as to increase the probability that the ATM cells are available when required to properly reassemble the native synchronous frame into a bit continuous stream. Using an analytical model for the first scheme and a simulation model for the second scheme we analyze the end-to-end delay of a synchronous frame, and compare the mean end-to-end delay for various trunk speeds (TA-to-network switch) to "access" line speed ratios. We conclude that, for the cases studied in this paper, if this ratio is greater than or equal to 4 then, over the range of useful loading, interleaved pipelining typically yields smaller mean delays than noninterleaved pipelining. If the ratio is less than 4 then interleaving typically yields larger mean delays for native protocols due to increased playout delays at the egress TA.