Performance evaluation of admission policies in ATM based embedded real-time systems

Abstract

We study the effect of the output link scheduling discipline of an ATM switch on the ability of an ATM LAN to admit real-time connections. Three output link scheduling policies are studied: first come first served (FCFS), round robin (RR), and packet-by-packet generalized processor sharing (PGPS). We derive connection admission criteria for the three scheduling policies. To evaluate the performance of the three scheduling policies, we introduce the metric of admission probability. The admission probability gives the probability that a randomly chosen set of real-time connections will be admitted into the network. The admission probability allows system designers to study the performance of different scheduling policies over a wide range of network loads. We observe that the performance of the three scheduling policies is sensitive to message deadlines. When the deadlines are small, PGPS outperforms both RR and FCFS, and RR outperforms FCFS. When the deadlines are large, all three scheduling policies perform the same. We also note that although PGPS is better than RR and FCFS most of the time, its improved performance is achieved at the cost of high implementation complexity and run time overheads.<>