Вы находитесь на странице: 1из 4

APPLYING TIME-DIVISION MULTIPLEXING IN STAR-BASED OPTICAL NETWORKS

M. Jin, O. Yang, Y. Zhang, A. G.P. Rahbar and W. Yang The Computer Communication Network Research Lab School of Information Technology and Engineering, University of Ottawa Ottawa, Ontario, Canada K1N 6N5 email: {mjin, yang}@site.uottawa.ca Abstract
Time-division-multiplexing (TDM) technique provides good granularity and response in high-speed networks. This paper presents the recent work of CCNR Lab in the study of TDM design and implementation in a star-based network. Node and network architectures are considered for the support of both the centralized and distributed algorithms. Performance evaluations have been carried out, and optimization is considered. Keywords: optical TDM, time slot assignment, star network, distributed and centralized algorithm, AAPN. (e.g. [8]). In a mesh topology, devices are connected with many arbitrary inter-connections. Ring topologies are widely used in the optical networks (e.g. SONET rings), because it can offer a higher bandwidth efficiency, and network failure protection. A simple physical tree topology requires the minimum number of links to interconnect the whole network, and can thus reduce the fiber expenditure in an optical network. Being a special form of the tree, the star configuration inherits the same reliability problem of a simple tree because any single failure may partition the network. Multiplicities/redundancies can be built into the network to improve reliabilities. For example, the AAPN (All-Agile Photonic Network) [9] adopted an overlaid star topology. Tree variants can be used to increase the network reliabilities [10], avoid complex routing, enforce network robustness and facilitate enabling technologies in the near future. The CCNR (Compute Communication Network Research) Lab at the University of Ottawa has been doing research work on the design, modeling, analysis and performance evaluation of computer communication networks, their protocols, services and interconnection architectures for the support of future Information Highway. The CCNR Lab has been working on various projects in the area of photonic networks, wireless networks, traffic characterization and control. Details can be found in Prof Yangs website. Prof Yang is a participant in the AAPN Research Network [11] and has been contributing to the AAPN research in switch architecture, bandwidth sharing and control techniques, and network failure control issues. These issues are inter-related especially for the first two. We have first worked on the OBS techniques [12], and lately the OTDM technique. The objective of this paper is to report on the current status of our work on OTDM. Section 2 summarizes the features of the star-based network architecture. Section 3 presents two interesting types of OTDM scheduling techniques. Section 4 provides the core and edge switch architectures to supports them followed by some performance results in Section 5. Section 6 presents the optimization works related to such operation. Work to be completed is also discussed in the Concluding Remarks (Section 7).

1. Introduction
Several classical technology candidates are applicable to alloptical switching. In circuit switching, a switching path is set up before the payload message is sent out from the source node. This method generates long latency and low efficiency. An example is the WDM (Wavelength Division Multiplexing) optical networks [1]. Packet switching splits the payload into sequential blocks, and uses a store-and-forward technique on any physical path that need not be pre-configured, e.g., [2]. This method requires more processing ability on switching components and their functionalities, such as header processing and message storage, but it goes very well with TDM (TimeDivision-Multiplexing) technique as the slot time is usually designed to take on fixed-size data units. OBS (Optical Burst Switching) [3] has combined and modified the previous two technologies to facilitate todays optical technology, but it is still unable to provide a good granularity due to its bursty characteristics. Optical Time-Division-Multiplexing (OTDM), e.g. [4], is a combination of the traditional TDM and the WDM technology. The bandwidth is shared by many sourcedestination pairs. Time slots are interleaved for high-dynamical and high-capacity photonic networks; thus it is a useful choice for WDM networks. The major challenges of an OTDM network are solving the routing, and the time slot and wavelength assignment problems. For example, optimization method for grooming traffic between the optical and electronic domains [5]. In general, there are three major types of physical topologies for transport networks: ring (e.g. [6]), tree (e.g. [7]) and mesh

0-7803-8886-0/05/$20.00 2005 IEEE CCECE/CCGEI, Saskatoon, May 2005

1141

period through a control channel. During the transmission of the last slot of the current frame, the slot assignment can already be performed for the next frame. Details can be found in a companion paper in this conference [13]. Unlike the centralized algorithm, the nice feature of the distributed algorithm is that a data slot does not have to wait for confirmation from the core switch and therefore no synchronization problem among all edge nodes. No feedback from the core node is needed for its transmission.
Edge node Core node

Figure 1: An Overlaid-Star Topology

4. Node Architecture
We shall use the following edge and core node architecture to support the two algorithms above. For micro-seconds switching, we think that acoustic-optic switch devices can be used while LiNbO3 [14] may be used in nano-seconds switching.

2. Network Architecture
Figure 1 depicts an over-laid star network consisting of edge nodes and all-photonic core switching nodes. Edge nodes are responsible for traffic aggregation and distribution, and core nodes relay traffic between any two edge nodes. We do not use any optical buffers or wavelength converters throughout the network, which are considered to be infeasible in current applications. The number of edge nodes can be in the order of hundreds. Since it is a big challenge to build such a large alloptical switch with current technology, an alternate solution is to add some selectors for traffics between the core and edge nodes. In this way, the size of a core switch in terms of the number of switching ports can be reduced dramatically.

4.1. Edge Node


Fig. 2 shows the general three stages of our switch in a sending edge-node. The first (queueing) stage aggregates the traffic from accessing networks and sorts them based on their destinations. This stage consists of VOQ (Virtual-OutputQueueing) buffers, one for each of the destination nodes. The traffic is assigned to distinct buffers according to their destination address. At the same time, incoming messages are organized into slots each of which has the length of one time slot. The second (forwarding) stage delivers slots held in buffers to output ports. The last (transmission) stage implements the electronics/ optics conversion and sends out optical signals. All operations of these stages are under the control of a control unit which is also responsible for the synchronization and header packet generation. Same switch architecture is used for a receiving node to receive messages from the core node.

3. OTDM Schemes
We would like to study both the centralized and the distributed schemes in order to find out the tradeoff in their design requirements and performance. In a centralized scheduling algorithm, each edge node stores the arrived payload messages and sends their slot requirements to a core node. The core node would allocate wavelengths and data (time) slots based on its knowledge of the available network resources and traffic demand, and sends back scheduling information. Upon receiving this information, an edge node sends out the payload on the wavelength channel at the designated time slots. If a data slot does not receive a grant signal, it has to wait further and then dropped when timed out. One can see that each data slot has to wait at least one round-trip time before its transmission. In order to synchronize slots from different edge nodes, this waiting time is determined by the longest distance from any edge node to the core node. Frame-by-frame and call-by-call are two scheduling methods at different levels to alleviate the slot-byslot method and result in different performance. In a distributed edge-based TDM protocol, bandwidth is provided in frame boundaries at the edge switches to support Diff Service. Each edge switch uniformly distributes sessions (traffic streams) over wavelengths, and uses a slot assignment algorithm to randomly distribute slots on wavelengths through the frame. The core switch would receive the data slot control information from all the edge switches within a slot-offset

4.2. Core Node


Figure 3 is a core switch architecture using optical TDM. Input messages are demultiplexed into different switching fabrics depending on their wavelengths. Control packets are sent to the control unit where the optics-electronics conversion and electronic processing are performed. Using the information included in the control packets, the control unit modifies the status of the switching fabric at the epoch of each time slot to transfer payload slots. A new control packet is also generated after integrating the routing information. Then new headers and switched payload slots are combined by a multiplexer and sent out through the output port.

4.3. Operation
When centralized scheduling is employed, the core node is responsible for transferring messages from f edge nodes to g edge nodes where each edge node i in general has respectively Ki and Hi fiber links to and from the core node. The

1142

scheduling of this scenario is equal to the routing of an asymmetric three-stage multi-connection Clos network. During each time slot, the core node removes the requests that are out of time and builds a traffic matrix. By repeating finding maximum flow across the network (e.g. the PREFLOW-PUSH algorithm in [15]), a series of switching matrix are built for each wavelength switch. Then the scheduling results are sent back to edge nodes. In our distributed algorithm, a core node would do its best to accommodate the request from the edge nodes by switching the data (time) slots without conflict. Otherwise, the slots have to be dropped, and a notification is sent back to the edge node for retransmission. Again details can be found in [13] in this conference.
queueing 1 1 2 Legacy Networks 2 To Core Nodes NI NO N E/O multiplexer forwarding transmission

time slots. The number of fiber links between the 8 edge-core connections is 1, 2, 3, 4, 1, 2, 3 and 4 respectively. Each fiber consists of four wavelengths. We measure the delay (defined to be the interval from when a packet arrives at an edge node until when the slot it belongs to is sent out from the edge), the throughput and the loss ratio.
average packet delay (sec)
0.00075 0.0007 0.00065 0.0006 0.00055 0.0005 0.00001

0.0001

0.001

maximum assebly delay (sec)

average packet delay (sec)

W 1
1

0.0009 0.0008 0.0007 0.0006 0.0005 0.0004 0.0003 10 15 20 25 30 35 40

0
Control Unit

distance (slots)

Figure 2: Architecture of an edge switch


0
O/E Control Unit E/O 1

1
1

W
Input Output

1
FI

Switching Fabric FO

demultiplexer

multiplexer

Figure 3: Architecture of a core switch

5. Performance Evaluation
We have conducted simulation study on our network and switch architecture and the scheduling algorithms on an 8x8 star network. We assume each channel carries the same traffic load with Poisson arrivals. The arrivals consists 46% of 40byte packets, 18% of 522-byte packets, 18% of 576-byte packets, and 18% of 1500-byte packets. Each edge node has equal distance to the core node, and each edge node is equipped with an infinite electronic buffer. Each time slot is 10 us, in which 9 us is used for the payload message and 1 us is supposed to be switching time of the core optical switch. The available bandwidth is 10Gbps. The edge to core distance is equal to 20 time slots. The packets MAD (Maximum Assembly Delay allowed for slot assembly) is 100 us and the slot maximum scheduling delay (for time-out constraint) is 100

Figure 4: Average Packet Delay The delay performance is decided by the slot assembly method, the transmission distance and the allowable waiting time. As shown in Fig 4, increasing MAD does not affect the delay significantly when it becomes twice or more than the slot length. It is demonstrated that the average packet delay increases almost linearly with respect to the edge-core distance. In our parameter setting, the MAD has little effect on the total delay comparing to the round trip time which is much longer than it. The total delay a packet is imposed in an edge node has a linear relationship with the distance and the slot waiting time. Since the distance is fixed for a given network, we may achieve required delay by choosing appropriate combination of these parameters. Other performance results are not shown due to limited space. Performance of distributed algorithm can be found in [13]. The loss performance there appears to be quite high and need to be much curtailed by proper choice of architecture parameters.

6. Optimization
Based on the Lagrangean Relaxation and Subgradient Method obtained in our previous research [16], we have proposed a modified algorithm and studied various network configurations and the influence of resources [17] in the AAPN network. Our algorithm is shown to be very computational effective, and the bounds generated are mostly within 1% of the final objective value. We have furthered it to study optimized grooming techniques [18] on Synchronous Optical Network (SONET) over WDM networks for static traffic that

1143

can combine low-speed traffic streams onto high-speed wavelengths in order to maximize the network throughput and profit. The optimal traffic grooming scheme for a star- network is a very complicated optimization problem because the problem is composed of two levels of closely related optimization subproblems, i.e., the routing and wavelength assignment of the lightpaths over the physical topology and the traffic demand routing from the electronic domain over the virtual topology. We will extend our grooming technique to formulate the problem as a constrained integer linear programming problem and use network profit as an innovative optimization objective. The routing cost in the electrical and optical domains as well as the revenue generated by accepting traffic demands will be modeled. Through the optimization process, traffic demands will be selectively accepted based on the profit (i.e., the excess of revenue over network cost) they generate. To generate a feasible network routing scheme, a heuristic algorithm is being proposed based on the dual solution. Like before, we like to present a systematic approach to obtain theoretical performance bounds

References
[1] T. E. Stern and K. Bala, Multiwavelength Optical Networks: A Layered Approach. Addison-Wesley, 2000. [2] M. J. OMahony, D. Simeonidou, D. K. Hunter and A. Tzanakaki, The Application of Optical Packet Switching in Future Communication Networks, IEEE Commun. Mag., vol. 39, no. 3, pp.128-135, Mar 2001. [3] J. S. Turner, Terabit Burst Switching, Journal of High Speed Networks, vol. 8, no. 1, pp.3-16, Jan 1999. [4] Soung Y. Liew, and H. J. Chao, On Slotted WDM Switching in Bufferless All-Optical Networks, in Proc. HOTI'03, 2003. [5] K. Zhu, H. Zang, B. Mukherjee, A Comprehensive Study on Next-Generation Optical Grooming Switches, IEEE Journal on Selected Areas in Communications, vol. 21, no. 7, pp. 1173-1186, Sept 2003. [6] Tsong-Ho Wu and Richard C. Lau, "A Class of Self-Healing Ring Architectures for SONET Network Applications", in Proc. IEEE GLOBECOM '90. [7] O. W. W. Yang, "Two-center Tree Topologies for Metropolitan Area Networks", IEEE Proc. Communications, vol. 141, pp. 280-288, Aug 1994. [8] Rainer R. Iraschko, M. H. MacGregor, "Optimal Capacity Placement for Path Restoration in STM or ATM Mesh Survivable Networks", IEEE Transactions on Networking, vol. 6, no.3, pp. 325-335, June 1998. [9] G.V. Bochmann, M.J. Coates, T. Hall, L. Mason, R. Vickers and O. Yang, The Agile All-Photonic Network: An architectural outline, in Proc. The 22nd Symposium on Communications, Queen's University, Canada, 2004. [10] Zhenxiao Liu and Oliver W. W. Yang , Terminal-Pair Reliability Analysis of Overlaid-Star Networks, in Proc. CCCT2004, Austin, Texas, Aug 14-17, 2004. [11] http://www.aapn.mcgill.ca/. [12] Mushi Jin and Oliver Yang, Parameter Choices and Their Effect on the Performance of the OBS Networks, in Proc. SPIE APOC, Beijing, China, Nov 7-12, 2004. [13] Akbar Ghaffar Pour Rahbar and Oliver Yang, Slot Contention Resolution for Distributed TDM Scheduling in an Optical Star WAN, to be presented at CCECE2005. [14] Stamatios V. Kartalopoulos, DWDM: Networks, Devices, and Technology. Wiley-IEEE Press, 2002. [15] R. K. Ahuja and James B. Orlin, A Fast and Simple Algorithm for the Maximum Flow Problem, Operations Research, vol. 37, no. 5, pp. 748-759, Sept-Oct 1989. [16] Yiming Zhang, Oliver Yang and H. Liu, "a Lagrangean Relaxation and Subgradient Approach to the Routing and Wavelength Assignment Problem in WDM Networks", IEEE J-SAC OCN Series, pp. 1752-1765, Nov 2004. [17] Yiming Zhang, Oliver Yang and Haomei Liu, "The optimization issues in an agile all-photonic backbone network", in Proc. SPIE APOC (Asia-Pacific Optical Communications), Beijing, China, Nov 7-12, 2004. [18] Y.M. Zhang, J. Wu, O. Yang, M. Savoie, A LagrangianRelaxation Based Network Profit Optimization for Mesh SONET-over-WDM Networks, (accepted) Journal of Photonic Network Communication.

7. Conclusions
This paper has discussed the current work of CCNR to implement OTDM in star-based networks. We have designed the node and network architectures to support both centralized and distributed algorithms. Some performance results are presented for use with the companion paper [13] The centralized scheduling operation appears to introduce a pretty high time complexity when the network becomes large. Some heuristics may be required to speed up the scheduling at the cost of wasting some slots instead of finding a perfect matching for all traffic requests. This has been confirmed indirectly turn by the performance of the distributed algorithm and would require further investigation. Work remains to study the tradeoffs between a centralized and a distributed system. As pointed out, the round-trip delay and the slot loss probabilities appear to be major performance measures. We would like to conduct more comparison on a common set of parameters in order to provide tradeoff study in the design. Finally, we hope our work in traffic grooming on the star network may shed lights into the scheduling algorithm as well.

Acknowledgements
This research is supported in part by the National Science and Engineering Research Council (NSERC) and industrial partners through the Agile All Photonic Networks (AAPN) Research Network. We would also like to thank Y.F. Gu and Wail Mardini for their technical support of this project. However, it should also be mentioned that the present paper is simply reporting on CCNR work relevant to the AAPN, but cannot be construed as an official position of AAPN.

1144

Вам также может понравиться