Sie sind auf Seite 1von 11

Efficient Distributed Control Protocols for WDM All-Optical Networks.

Yousong Mei and Chunming Qiao Department of ECE SUNY at Buffalo Buffalo, NY 14260 f ymei,qiaog @eng.buffalo.edu Abstract Path Multiplexing (PM) and Link Multiplexing (LM) [3, 5] are two approaches for establishing connections (or lightpaths) in optical networks. This paper describes distributed control protocols for establishing lightpaths in WDM networks using LM and PM. We propose and evaluate the performance of two classes of protocols, namely Source Initiated Reservation (SIR) and Destination Initiated Reservation (DIR). It is found that DIR protocols generally perform better than SIR protocols. However, the impacts of DIR protocols on the performance of a network using LM and PM are different. Key Words Optical Network, WDM, Wavelength Reservation, Distributed Control. 1 Introduction In WDM networks, there are two basic approaches that can be used to select wavelengths to establish lightpaths. One approach establishes a connection using the same wavelength along the path, and is called Path Multiplexing (PM). The other approach can establish a connection using possibly different wavelengths on different links along the path, and is called Link Multiplexing (LM) [3, 5]. LM is more flexible but may require wavelength conversion at intermediate nodes. In either approach, the network control (or signaling) required for establishing a connection may be either centralized or distributed. Under centralized control,each requestisprocessedbyacentralcontroller,which chooses a path and assigns an appropriate wavelength on each link along the path to establish the connection. In large networks, centralized control is not feasible and therefore, it is essential to study distributed control protocols. Most studies on all-optical WDM networks focus on wavelength assignment algorithms under centralized control [1, 2, 7, 8], and only a few studied distributed control . This research is supported in part by a grant from NSF under contract number MIP-9409864 [4, 6, 9]. This paper studies two classes of distributed wavelength reservation protocols, namely Source Initiated Reservation (SIR) protocols and Destination Initiated Reservation

(DIR) protocols. We investigate the performance of these protocols in both LM and PM. The rest of the paper is organized as follows. In Section 2, we discuss various SIR and DIR protocols, which is followed by the performance evaluation of these protocols in Section 3. We conclude the paper in Section 4. 2 Distributed Control in Optical Networks In a large network, a central controller is often the performance bottleneck of the system. Distributed control can not only eliminate such a bottleneck but also improve the system reliability. However, distributed control may introduce more complexity in the protocol design and a higher control latency. An optical network with distributed control may be considered as having a data network consisting of the optical switches interconnected by several data wavelengths, and a control network consisting of the control units (CUs) interconnected by one or more control wavelengths. Each optical switch is controlled by a CU, which also maintains the status information of the data wavelengths accessible to the switch. A CU exchanges the control information (e.g. requests and acknowledgements) with other CUs by sending and receiving control packets. 2.1 Source Initiated Reservations (SIR) In this section, we first describe a basic SIR protocol for PM, then we describe some variations applicable to both LM and PM. In the following discussions, the term channel refers to a data wavelength. Under PM, the same channel has to be used on different links along the path. A basic SIR protocol reserves all potentially suitable channels on every link along the path and then releases the unnecessary ones later. Specifically, when the source wishes to establish a connection, it sends a request packet (REQ) to the destination. At the source node,

all available channels are considered as candidates, and are reserved in parallel. When an intermediate node receives the REQ, it checks the status of these candidate channels on the outgoing link, and only reserves the subset of these candidate channels which are currently available. This subset becomes the new set of candidate channels and the information is carried by the REQ to the next node. Note that if the subset should become empty, a negative acknowledgement (NAK) will be sent back to the source and the REQ will be dropped (see Figure 1(a)). The NAK releases all the channels reserved by the corresponding REQ and informs the source node of the failure. The source will send another REQ later in an attempt to establish the connection. Source Intermediate Destination Source Destination REQ NAK REQ ACK Reservation Retransmit Time { Data REQ REL Transfer (a) (b) Figure 1. A basic SIR protocol If the REQ does arrive at the destination with a nonempty set of candidate channels, the destination will select one channel to be used for the connection and send a positive acknowledgement (ACK) back to the source node identifying the channel. The ACK follows the reverse of the path taken by the corresponding REQ and releases all the channels reserved by the REQ except the one chosen by the destination. Once the source gets the ACK, it knows that the connection has been established and can start transferring data. After transferring all the data, the source sends a release packet (REL) to the destination to tear down the connection. This is illustrated in Figure 1(b). The basic protocol described above is essentially a parallel reservation protocol which adopts the dropping policy. Some of its variations are described briefly as follows. Holding Policy: When the source sends a REQ, it specifies the lifetime of the REQ. At an intermediate node, if the REQ cannot progress (e.g. because no candidate channels is available on the outgoing link), it will not be dropped right away as in the basic protocol. Instead, the intermediate node will buffer the REQ, hoping that some suitable channels

will become available before the REQ expires. If a suitable channel does become available, the REQ continues its journey. On the other hand, if the REQ expires, a NAK will be sent back to the source to inform the failure as in the basic protocol. Sequential Reservation: In the basic protocol, the REQ reserves all suitable channels on each link. This seems to increase the chances of success for a specific connection. However, it wastes a lot of bandwidth and may decrease the chances of success for other connections. An alternative is to use a certain heuristic to select only one candidate channel and then reserve it on the first and subsequent links. If the reservation fails, the source will try another suitable channel later. Note that in LM, any available channel on a link is a suitable candidate. Hence, parallel reservations only apply to PM but do not make sense in LM. However, sequential reservations (using either dropping or holding) using a different wavelength selection criterion can be applied to LM. 2.2 Destination Initiated Reservation (DIR) In all SIR protocols, the channels are reserved by REQ on its way from the source to the destination, and hence bandwidth on the reserved channel is wasted during the reservation period, as shown by the light shaded area in Figure 1(b). This overhead may become significant in high-speed WANs. One way to tackle this problem is to use Destination Initiated Reservation (DIR) protocols to reduce the overhead. Source Intermediate Destination Source Destination REQ RESV NAK NAK REQ REL ACK Reservation Retransmit Time { ACK (a) (b) Figure 2. A basic DIR protocol Figure 2 shows how a DIR protocol adopting sequential reservation and dropping policy works in general. After the source sends a REQ to the destination, the REQ collects the wavelength usage information on its way to the destination but does not reserve any channel. When the destination gets the REQ, it can make an intelligent decision on which channel to reserve on each link based on the information provided by REQ. In addition, if there is no wavelengths available along the (primary) path, the destination may try an alternate path. In any case, an ACK is then sent back to the source to reserve the appropriate channel on each link along the path. If the ACK cannot progress at an intermediate

node, the intermediate node sends an NAK to the destination to release the partially established path, and may also send an NAK to inform the source of the failure so that the source may try again later. 3 Performance Evaluation In this section, we will evaluate the performance of the protocols based on the average throughput of a network, which is the amount of data (e.g. in Kbits) transferred in a time unit (e.g. 100 s). Under distributed control, it is very difficult to evaluate the performance analytically even for some simple topologies without making much simplified

and less realistic assumptions. Hence, we evaluate the performance through simulations. In our simulation, we assume that each host generates a connection request according to a Poison Process. The destination of each connection is uniformly distributed. The connection duration (i.e. the amount of data to be transferred) has an exponential distribution with an average of 256 time units, and the time needed to process a control packet at an intermediate node is one time unit. The simulationresults are obtained for an 88 mesh with 8 data wavelengths on each direction of the full-duplexed links. The propagation delay between any two adjacent nodes is assumed to be the same, and has two possible values, namely 1 time unit (for MANs) and 64 time units (for WANs). In the following presentation of the simulation results, we use S for sequential reservation, P for parallel reservation, H for holding, and D for dropping. Accordingly, SIR-SH, for example, refers to the SIR protocol adopting sequential reservation and the holding policy. The lifetime of each REQ (meaningful only with holding policy) is set to be a multiple of the sum of the propagation and processing delays along the path. Two DIR protocols adopting sequential reservation and the holding policy, namely DIRPRIME, which uses the primary path only, and DIR-ALT, which may use an alternate path, are evaluated. 20 40 60 80 100 120 140 160 180 200 0 0.005 0.01 0.015 0.02 Throughput (x transmission rate per wavelength) Connection Requests Per Node Per Time Unit LM, SIR-SD LM, SIR-SH PM, SIR-PD PM, SIR-PH PM, SIR-SD PM, SIR-SH Figure 3. SIR with propagation delay = 1 Figures 3 and 4 shows the throughput of various SIR protocols with different propagation delays. In all the cases, protocols adopting the dropping policy has a slightly higher throughput than protocols adopting the holding policy 1 . One notable difference between the two figures is that parallel reservation performs better than sequential reservation with a small propagation delay but is worse with a large propagation delay. This is because the bandwidth wasted by

parallel reservation in SIR protocols is insignificant (com1 1 Note that, holding may be better than dropping under some other cir cumstances [4, 9] 5 10 15 20 25 30 35 40 45 50 0 0.0005 0.001 0.0015 0.002 0.0025 0.003 0.0035 0.004 0.0045 0.005 Throughput (x transmission rate per wavelength) Connection Requests Per Node Per Time Unit LM, SIR-SD LM, SIR-SH PM, SIR-SD PM, SIR-SH PM, SIR-PD PM, SIR-PH Figure 4. SIR with propagation delay = 64 pared to its advantage) in the former case but becomes dominant in the later case. Figure 5 compares the throughput of the DIR protocols with the best SIR protocols with a small propagation delay under LM and PM. One can see that under LM, DIRPRIME protocol does not significantly improve the performance over the best SIR protocol (which is SIR-SD in LM), and DIR-ALT protocol can improve the performance by about 10%. On the contrary, under PM, DIR-PRIME protocol improves the performance over the best SIR protocol (which in this case happens to be SIR-PD) by about 10% but DIR-ALT cannot improve the performance further. 20 40 60 80 100 120 140 160 180 200 220 0 0.005 0.01 0.015 0.02 Throughput(x transmission rate per wavelength) Connection Requests Per Node Per Time Unit LM, DIR-ALT LM, DIR-PRIME

LM, SIR-SD PM, DIR-ALT PM, DIR-PRIME PM, SIR-PD Figure 5. DIR vs. SIR (propagation delay = 1) Figure 6 compares the performance of different protocols with a large propagation delay. Note that the best SIR protocols used in this comparison become SIR-SD in both LM and PM. The situation is totally different from that of a small propagation delay. Specifically, under LM, both DIR protocols improve the performance over the best SIR protocol (SIR-SD) considerably. However, under PM, neither DIR

protocols improves the performance significantly over the best SIR protocol (SIR-SD). 0 10 20 30 40 50 60 70 0 0.0005 0.001 0.0015 0.002 0.0025 0.003 0.0035 0.004 0.0045 0.005 Throughput (x transmition rate per wavelength) Connection Requests Per Node Per Time Unit LM, DIR-ALT LM, DIR-PRIME LM, SIR-SD PM, DIR-ALT PM, DIR-PRIME PM, SIR-SD Figure 6. DIR vs. SIR (propagation delay = 64) The results in Figures 5 and 6 can be explained based on the following discussions. In general, DIR protocols have two useful features when compared to SIR. One is the reduced bandwidth waste during the reservation period, and its usefulness is proportional to the propagation delay. The other is that the destination can make a better decision based on the wavelength usage information along the primary path provide by REQ. However, the accuracy of the information is inversely proportional to the propagation delay. For LM, what a destination node needs to know is whether any wavelength is available on the primary path, but the accuracy of the information is less important. Specifically, the second feature is useful to DIR-ALT only since it can use an alternate path if there is no available wavelength on the primary path (there is nothing DIR-PRIME can do in such a case). This is why the first feature helps DIR-PRIME improve the throughput over a SIR protocol with a large propagation delay, and the second feature helps DIR-ALT improve the throughput further. For PM, on the other hand, a protocol needs to know which wavelength is available for reservation on all the links along the path, and hence the accuracy of the information is of great importance. This is why DIRPRIME can improve the throughput with a small propagation delay (relative to a parallel SIR protocol) but not with a large propagation delay (relative to a sequential SIR protocol). Furthermore, DIR-ALT cannot do better than DIRPRIME since no wavelength usage information is available on the alternate path at all. Finally, it is also obvious from the results that LM performs much better than PM under all cases. In some specific situations where the propagation delay is large and a DIR protocol is used, LM can achieve twice the throughput of PM. This suggests that the cost of extra hardware for wavelength conversion required by LM may be justified un

der distributed control. 4 Concluding Remarks In this paper, we have described protocols for wavelength reservation in WDM all-optical network with distributed control. We have examined both SIR and DIR protocols and simulated their performances in a mesh network. It is found that DIR protocols generally performs better than SIR protocols. However, the impacts of DIR protocols on the performance of LM and PM are different. In all cases, LM can result in a significantly better performance than PM. References [1] I. Chlamtac, A. Ganz, and G. Karmi. Lightpath communications: an approach to high-bandwidth optical WANs. IEEE Transactions on Communications, 40:1171 1182, July 1992. [2] K.C. Lee and Victor O.K. Li. A circuit rerouting algorithm for all-optical wide-area networks. In Proc. of IEEE Infocom, pages 954 961, 1994. [3] C. Qiao and Y. Mei. On the multiplexing degree required to embed permutation in a class of interconnection networks. In Proc. of the IEEE High Perfermance Computer Architecture(HPCA), pages 118 129, February 1996. [4] C. Qiao and Y. Mei. Wavelength reservation under distributed control. IEEE/LEOS summer topical meeting on Broadband Optical Networks-Enabling Technologies and Applications, August 1996. [5] C. Qiao and R. Melhem. Reducing communication latency withpathmultiplexinginopticallyinterconnected multiprocessor systems. IEEE Trans. on Parallel and Distributed Systems, 8(2):97 108, February 1997. (A preliminary version appeared in HPCA 95). [6] R. Ramaswami and A. Segall. Distributed network control for wavelength routed optical network. In Proc. of IEEE Infocom, pages 138 147, 1996. [7] R. Ramaswami and K.N. Sivarajan. Optimal routing and wavelength assignment in all-optical networks. In Proc. of IEEE Infocom, pages 534 543, June 1994. [8] S. Subramaniam, M. Azizoglu, and A.K. Somani. Connectivity and sparse wavelength conversion in wavelength-routing networks. In Proc. of IEEE Infocom, pages 148 155, March 1996. [9] X. Yuan, R. Melhem, and R. Gupta. Distributed path reservation algorithm for multiplexed all-optical interconnection networks. In Proc. of the IEEE High Perfermance Computer Architecture(HPCA), February 1997.

Das könnte Ihnen auch gefallen