Performance analysis of the pi-persistent protocol in unidirectional bus networks
Introduction
In a series of papers Mukherjee et al. 1, 2, 3, 4, 5 proposed a medium access protocol for high-speed time-slotted unidirectional bus networks, known as the pi-persistent protocol. This is a straight adoption of the well-known p-persistent protocol of broadcast busses (cf. [6]), in which stations linked by unidirectional (folded or dual) busses have different station-dependent probabilities of accessing empty slots. The pi-persistent protocol, with its very simple flow control mechanism, has been seen as an alternative for solving fairness problems associated with the original IEEE 802.6 standard for the distributed queue dual bus (DQDB) for high-speed networks 7, 8, 9, 10, 11. Conceptually similar techniques, based on accessing slots according to station-dependent priorities, had been earlier considered in [12] for single and in 13, 14 for multiple unidirectional busses.
Protocols of this type are based on open loop flow control schemes and can be attractive in high-speed networks since in such networks, even in the case of a folded bus, feedback information in the upstream direction is not available in a timely manner. Various alternative solutions for medium access control (MAC) protocols in slotted unidirectional busses, including those implemented in CRMA, Fasnet and Hangman networks, can be found surveyed e.g. in [15].
Under the pi-persistent protocol each slot is able to carry one data packet of fixed size plus some overhead bits necessary for synchronization and control of the network. If station i has a packet to send, it persists with its attempt to transmit the packet in the next free slot with probability pi (see Fig. 1). The closer a station is located to the origin of the bus the larger is the probability that it encounters a free slot. If, as an extreme case, station 1 always has packets to transmit and its access probability per slot is 1, then there will be no free slots available for subsequent stations.
Packet delays under the pi-persistent protocol have been investigated in [16] by modeling contents of subsequent slots as a two state switched Bernoulli process. By matching corresponding moments an approximate analysis is carried out for any number of coupled stations. The exact upper and lower bounds on the average packet delay and queue length for the original pi-persistent protocol and its deterministic counterpart, in which station i can access only each ⌈1/pi⌉ empty slot, are given in [17]. If the access probabilities are properly selected, the rate of accessing the channel by different stations can be balanced.
In a network with N stations the probabilities pi, 1≤i≤N, can be chosen also adaptively by sensing the actual load on the bus. Corresponding methods have been suggested by [3]. Besides flexibility concerning load changes, this also allows for dynamic integration of new stations. After a short sensing period, where an appropriate access probability pi is estimated, new stations simply use empty slots for transmission. However, this induces a certain management overhead, particularly if the whole network is reconfigured. This price has to be paid to achieve fair network access, independent of the position on the bus.
Reference [5] deals with fairness of the pi-persistent protocol by introducing different criteria, e.g. equal mean packet delay, equal blocking, and equal throughput. It is assumed that each station has equal finite buffer capacity M to store packets, arriving at station i according to a Poisson process with intensity λi. If an arriving packet encounters a filled buffer, it is blocked and lost. Of course, if high blocking rates are tolerated, low delay times are easily achievable, for instance by reducing the buffer size. To compromise these two criteria is the most critical issue of the protocol. The network and the access bridges should be designed in such a way as to cope with the major part of the entire traffic load, which is also stressed in [1].
It is obvious that mechanisms used for enforcing fairness in communication networks can potentially worsen their performance. Thus, detailed studies of tradeoffs between fairness and efficiency of different networks are needed. The main purpose of this paper is to investigate the relationship between the fairness enforcement mechanism and performance of a network operating under the pi-persistent protocol. Pursuing this goal we look at the network from an operator's point of view, and assume that a certain guaranteed capacity of the network is bought by clients, just enough to satisfy their individual needs. If there is extra unused capacity available, it is offered to all subscribers in order to improve their throughput and delay times. New subscribers receive their ordered capacity, hence the possibility that others' quality of service will be reduced, but never below the guaranteed threshold. A network control of this type can be implemented by dynamically steering the access probabilities pi, which offers an elegant way of warding off overload.
A surprising point emerges from the case when the bus is able to carry all offered traffic from Poisson arrival streams, or more generally, from (non-batched) arrival processes with independent increments. Then, if an active upstream station i reduces its access probability pi to allow fairer treatment of downstream stations, this deteriorates the waiting times at station i, but does not improve waiting times at subsequent stations i+1,…,N. Hence, fairness can be achieved only by sacrificing a certain part of overall performance.
We set out to describe briefly the topology of the bus which was basically developed in [5]. Empty slots are generated by a control station, and slots drop off the bus at the end (black boxes in Fig. 1). N stations are using the bus for communication by sensing the outbound and inbound bus channel. Stations are labeled 1,…, N according to their relative position on the bus, starting with the station closest to the control station. The packet arrival stream at station i is assumed to be a process with independent increments with mean λi (packets per slot), where packets arrive entirely at time instants, e.g. in bulks of 64 bits in parallel. Packets are sequentially copied bit-by-bit to the channel. To simplify the analysis we assume that each station has an infinite buffer to store locally generated packets. If a station has a packet to send, it independently attempts to transmit it in the next slot with probability pi, provided this slot is empty. With probability 1−pi it leaves an empty slot, which is passed for use by subsequent stations. We furthermore suppose that the arrival processes at different stations are stochastically independent.
A new packet arriving at a random instant τ at a buffer already filled with k packets has to wait a random number of slots until the k packets ahead are cleared. Moreover it has to wait a random number of slots until its own transmission starts, and one slot until it is completely copied to the channel, counted from the beginning of the next slot after its arrival time τ (see Fig. 2). Hence, the minimum waiting time for a packet arriving at an empty buffer is 1.
The basic problem now is how to choose the pi such that the above-described requirements apply. We start by analyzing the queue length and packet delay at each station and their dependence on the access probabilities.
Section snippets
Access probabilities under heavy traffic
If the bus capacity is C bps, then user i may reserve (or buy) a certain portion αi, 0<αi≤1, to satisfy its needs. In summary, N users share a certain portion of the channel capacity by choosing (or getting assigned) α1,…,αN>0, ∑i=1Nαi≤1, each demanding ⌊αiC⌋ bps on average. To guarantee this portion for each station under heavy traffic the equationhas to be solved. The left-hand side gives the probability that a slot is used by station i, if all stations always have
Queue length and waiting times
It is important for each station to know its queue length distribution and the mean waiting times of packets as a function of the actual access probabilities pi and the traffic loads λi. Obviously, the performance parameters of each station are influenced by the traffic load and the presently used pi of other stations.
We first assume that there is only one station on the bus with channel access probability p and packet arrival times according to a one-dimensional Poisson process with intensity λ
Multiple access
The general case of N stations, each persisting to access an empty slot with probability pi, is now considered. Let qi denote the probability that the buffer at station i is nonempty at a slot beginning instant in steady state. By Eq. (9)it holds for the first station that q1=λ1/p1. For station i the probability that a waiting packet is transmitted in a slot can be approximated byHence, with Eq. (9)This system can be solved
Conclusions
In this paper we have achieved an exact analysis of the queue length distribution, the expected delay time of packets and the corresponding variance when focusing on a single station in a unidirectional bus. For n coupled stations we assume independent slot occupancies, which is shown to be quite accurate for moderate load by simulation. A criterion for choosing the pi is given in order to guarantee a certain portion of bus capacity for individual users. There is an interesting effect for
Acknowledgements
This work was partially supported by the Deutsche Forschungsgemeinschaft under Grant Ma 1184/5-1 (R. Mathar), and by the Alexander von Humboldt Foundation (K. Pawlikowski). We thank Nurul Sarkar for programming the extensive simulation tasks. We gratefully acknowledge the reviewers' helpful comments which improved the paper.
Rudolf Mathar was born in Germany in 1952. He received his Dipl.Math. and Dr.Rer.Nat. degree in Mathematics from Aachen University of Technology in 1978 and 1981, respectively. In 1986–1987 he worked at the European Business School as a lecturer in computer science, and in 1988–1989 he joined a research group in applied optimization at the University of Augsburg. In October 1989 he joined the faculty at Aachen University of Technology, where he is currently a Professor of Stochastics. He is
References (18)
- et al.
The continuation-bit approach and the pi-persistent protocol for scheduling variable-length messages on slotted high-speed fiber optic LANs/MANs
Computer Networks and ISDN Systems
(1994) A journey through the DQDB network literature
Performance Evaluation
(1992)Media access techniques: the evolution towards terabit LANs and MANs
Computer Networks and ISDN Systems
(1994)Algorithm for the pi-persistent protocol for high-speed fibre optic networks
Computer Commun.
(1990)- et al.
Dynamic control and accuracy of the pi-persistent protocol using channel feedback
IEEE Trans. Commun.
(1991) - et al.
Integrating voice with the pi-persistent protocol for unidirectional broadcast bus networks
IEEE Trans. Commun.
(1988) - et al.
The pi-persistent protocol for unidirectional broadcast bus networks
IEEE Trans. Commun.
(1988) - J.L. Hammond, P.J.P. O'Reilly, Performance Analysis of Local Computer Networks, Addison-Wesley, Reading, MA,...
- IEEE Standard 802.6/D15, IEEE standards for local and metropolitan area networks: distributed queue dual bus (DQDB)...
Cited by (1)
A delay-throughput performance improvement to the p<inf>i</inf>-persistent protocol
2001, IEEE Symposium on Computers and Communications - Proceedings
Rudolf Mathar was born in Germany in 1952. He received his Dipl.Math. and Dr.Rer.Nat. degree in Mathematics from Aachen University of Technology in 1978 and 1981, respectively. In 1986–1987 he worked at the European Business School as a lecturer in computer science, and in 1988–1989 he joined a research group in applied optimization at the University of Augsburg. In October 1989 he joined the faculty at Aachen University of Technology, where he is currently a Professor of Stochastics. He is especially interested in applications of computer science. His research interests include mobile communication systems, performance analysis and optimization of networks, and applied probability.
Krzysztof Pawlikowski is currently an associate professor of Computer Science at the University of Canterbury in Christchurch, New Zealand. He received his Master of Electronic Engineering and Ph.D. in Computer Engineering from the Technical University of Gdansk, Poland, and worked there until 1983. Since then Professor Pawlikowski has worked at universities and research labs in Australia, Germany and the USA. He is the author of over 70 research publications, including four books on computer communication networks. Professor Pawlikowski's research interests are in the area of stochastic discrete-event simulation, performance modeling of ATM and optical telecommunication networks, and information theory.
- 1
The paper was written when the author was visiting the Department of Computer Science, University of Canterbury, Christchurch, New Zealand.