Performance of optical burst switched networks: A two moment analysis

Abstract

We describe a fairly accurate model for the analysis of optical burst switched (OBS) networks. We use a two-moment approximation to calculate the burst dropping probabilities and hence the throughput of OBS networks. We consider networks with no-deflection of a burst and also networks in which blocked bursts may be deflected from their primary paths. The algorithms that we describe are adapted from the models for circuit multiplexed networks. The adaptation consists of modeling the offered load on the links by accounting for the dropping of blocked bursts on the routes. We obtain a two moment description of the link loads which turn out to be superpositions of several Poisson and non Poisson flows. Further, since the offered load to a link is not Poisson, the dropping probabilities for the different routes on it are not identical and they are evaluated separately. Numerical results show that these features of the model provide a significant improvement in the accuracy of the link and end-to-end burst dropping probabilities. Obtaining accurate link blocking probabilities for each route is expected to help in the better dimensioning of the link capacities and routing.

Introduction

Traditionally, telecommunication networks use either circuit switching or packet switching. Both these switching techniques have significant disadvantages when used in optical networks. Circuit switching requires at least one roundtrip time between the source and the destination before actual data transfer can begin. This can cause significant delay overheads, especially if the transmission time for the payload is not significantly larger than the roundtrip time. Packet switching requires that switches have the ability to buffer packets. Optical buffers are provided by fiber delay lines. To be able to be used in an optical switch, the fiber delay lines should have the ability to be programmed at high speeds. This technology is not yet feasible. As an alternative to optical buffering, electronic buffering may be provided for packet switching but this requires the use of O/E and E/O conversion which significantly reduces the speed of operation. In addition to buffering, a packet switch should perform packet header processing. This has to be done electronically and requires O/E and E/O conversion which again reduces the speed of operation of the switch. Optical burst switching (OBS) [1], [2], [3], [4], [5] uses the ‘bufferlessness’ of circuit switching (while avoiding connection setup delays) and ‘on demand switching’ capability of packet switching techniques (while avoiding the requirement of buffering capability at the switch).

In a burst switched network, edge nodes assemble packets for the same destination into ‘bursts’ of data before beginning transmission. A source with a burst of data to send to a destination will first transmit a control packet on a logically separate channel. The control packet contains signaling and routing information that is used by the switches to set up the path for the impending burst. Thus the control packet has information similar to that carried by a packet header or a connection setup message. The data burst follows the control packet on the optical channel after a fixed delay. We can thus interpret burst switching as ‘reservation without confirmation’, i.e, although a control packet tries to reserve bandwidth on links on the path, the source does not wait for any confirmation about the availability of bandwidth on the path. Thus in optical burst switching, connection setup delay overheads are avoided by having the source transmit almost as soon as a burst of data is ready. If an all optical path cannot be set up for a burst between the source and the destination, then the burst may be dropped en route. Buffer requirements are avoided by dropping the bursts that cannot be accommodated on a link. We see that burst switching provides a transparent all-optical path between the source and the destination whenever such a path is available. Note that when a burst is dropped en route, it will have consumed capacity on the links till the point at which it is dropped. This is clearly an overhead of the simplicity of the protocol.

Two important issues in OBS are the control architecture and the contention resolution mechanism. The control architecture, also called the signaling architecture, defines how an edge node will inform the network about an impending burst and the kind of information about the burst that the source gives to the network. ‘Just-in-time’ [6] and ‘just-enough-time’ [7] are two of the many signaling schemes that have been proposed for OBS networks. An excellent discussion of the issues in the design of the control architectures for OBS networks is available in [7]. The analysis that we present in this paper is fairly independent of the control architecture.

Contention on a link occurs when more bursts than the capacity of the link need to use the same link at the same time. In packet switching all but one of the contending packets are queued at the switch while in circuit switching the excess calls are blocked and the blocked calls do not utilize any data transmission resources. In OBS, in the event of a contention, the bursts cannot be queued. A simple option is to drop the excess of the contending bursts. A second option is to do the following. Each burst is allocated a preferred path and a set of alternate paths branching out of the preferred path. If a link on the preferred path is blocking, then the burst is deflected onto an alternate path from the source node of the contention link. A burst will be dropped only if all the alternate paths are also blocked. Networks with the latter capability will be called ‘deflection networks’ while the former will be called ‘no-deflection networks’. Deflection routing has been studied extensively under the assumption that packets are never dropped. See [8], [9], [10], [11], [12], [13], [14] for analysis of time slotted networks and [15], [16] for unslotted networks. These analyses are not applicable to OBS networks because of burst dropping. Burst segmentation [17] is another contention resolution method to reduce losses during contentions. In burst segmentation, a long burst is segmented, and any number of the segments can be dropped without affecting the other segments. Hence it always ensures a ‘partial’ success during contentions.

In this paper, we present a unified analytical model for no-deflection and deflection networks that use optical burst switching. We will separately describe the model details for both no-deflection and deflection OBS networks but will not pursue burst segmentation. A nice discussion on burst segmentation is available in [18]. In the analysis of no-deflection networks, we will see that our model is more detailed than previously known models, e.g., [19], [20], [21] and provides more accurate, detailed performance characteristics. To the best of our knowledge, the only analytical model for deflection routed OBS networks is in [22], [23]. We will discuss the comparison of our models and those of [22], [23] in Section 5.

In the solution of our unified model, we borrow techniques from the analysis of circuit multiplexed networks. Note that in circuit multiplexed networks, blocked calls do not consume any resources in the data path while in OBS networks, when bursts are dropped they utilize bandwidth till the point they are dropped. Clearly, the analysis methods from circuit multiplexed networks are not directly applicable to the analysis of burst switched networks and we need to suitably adapt those models for OBS networks.

The rest of the paper is organized as follows. In Section 2 we describe our analysis technique for calculating the throughput of a no-deflection burst switching network. In Section 3 we extend it to deflection routed burst switching networks. In Section 4 we validate our analytical models with simulation results. In Section 5 we develop a simple intuitive measure to compare the burst dropping probabilities on the links for the different routes obtained from different models. We show that with this measure, our detailed results are more accurate than previous models.