Synchronous Stream Optical Burst Switching

Synchronous Stream Optical Burst Switching Oliver Yu, Ming Liao, and Yuan Cao Department of ECE, University of Illinois at Chicago 851 S. Morgan Street, Chicago, Illinois 60607 oyu@ece.uic.edu Abstract his paper presents a novel architecture of synchronous stream optical burst switching (SS-OBS), which employs a unified traffic control plane to enable a wavelength-routed optical network with integrated sub-wavelength grooming services to support sub-wavelength streaming traffic of constant bit rate (CBR) and variable bit rate (VBR) with guaranteed end-to-end delay and data loss rate. SS-OBS shapes sub-wavelength CBR streams into periodic burst trains; aggregates multiple sub-wavelength VBR streams with the same source and destination into a statistically multiplexed periodic burst train. By adopting periodic advanced scheduling of burst trains, SS-OBS substantially decreases recursive signaling processing in supporting sub-wavelength streaming traffic. SS-OBS employs burst train reservation blocking recovery by time-shifting a blocked burst train. Performance of SS-OBS is analyzed in terms of wavelength utilization and data loss. It is found that the performance of the SS-OBS under each type of traffic is compared favorably with other single-service-based OBS schemes. I. INRODUCION Built over wavelength-routed optical networks with photonic lambda switches and wavelength division multiplexing (WDM), lambda Grid systems have been increasingly deployed to support scientific applications with multi-gigabit rate bandwidth. While these wavelength-routed lambda Grid systems are optimal to support data Grid applications with constant wavelength traffic streams, they are suboptimal to support access Grid applications with constant and variable sub-wavelength traffic streams. Efficient implementation of sub-wavelength traffic routing requires a finer granularity of multiplexing. Employing one-way reservation, Optical Burst Switching (OBS)[1][2] was proposed to statistically multiplex optical bursts to reuse wavelength bandwidth. OBS separates Control Packet (CP) and data burst with offset time and reserves the network resources along the route of the CP. Various signaling schemes have been proposed for OBS networks, e.g. Just-in-time (JI)[3], Just-enough-time (JE)[4], and Horizon[5]. A comprehensive comparison of these schemes can be found in [6]. OBS is designed to support sub-wavelength bursty traffic without guaranteed end-to-end delay and data loss rate due to burst contention blocking. Furthermore, OBS cannot efficiently support sub-wavelength streaming traffic because excessive periodic reservation request overheads would be incurred. Employing two-way reservation and edge burst aggregation, Wavelength-Routed OBS (WR-OBS) [7][8] guarantees end-to-end delay and data loss rate due to aggregation buffer overflow, but it ignores data loss rate his work was supported in part by the U.S. Department of Energy (DOE) under Grant DE-FG02-03ER25566, and was supported in part by the NSF under Grant SCI-0225642. due to burst reservation blocking. Synchronous OBS [9] employs two-way advanced periodic reservation to support sub-wavelength CBR traffic streams with guaranteed data delay and loss. But it is not efficient when supporting sub-wavelength VBR traffic because it does not statistically multiplex different VBR traffic streams along the same route. Proactive reservation-based switching (PRS) [10] also employs periodic advanced reservation to support sub-wavelength traffic streams, but it is a one-way reservation protocol and therefore cannot guarantee data delay or loss. his paper presents the synchronous stream optical burst switching (SS-OBS) paradigm, which employs a unified traffic control plane to enable a wavelength-routed optical network with integrated sub-wavelength grooming services to support sub-wavelength streaming traffic of constant bit rate (CBR) and variable bit rate (VBR) with guaranteed end-to-end delay and data loss rate. SS-OBS shapes a sub-wavelength CBR stream into a periodic burst train; aggregates multiple sub-wavelength VBR streams with the same source and destination into a statistically multiplexed periodic burst train. By adopting periodic advanced scheduling of burst trains, SS-OBS substantially decreases recursive signaling processing in supporting sub-wavelength streaming traffic. SS-OBS employs burst train reservation blocking recovery by time-shifting a blocked burst train, which increases the offset delay between reservation request and burst train transmission. he performances of blocking probability and network utilization are analyzed through simulations and compared with those of JE-OBS and WR-OBS for both CBR and VBR traffic stream. he rest of this paper is organized as follows. Section II gives the detailed description of SS-OBS system. Section III presents the system modeling. Section IV shows the simulation results, and section V concludes the paper. II. SYNCHRONOUS SREAM OPICAL BURS SWICHING (SS-OBS) Fig.1 shows the System architecture of SS-OBS. At the edge node, the CBR or VBR sub-wavelength streams is aggregated and shaped into a periodical burst train through edge stream shaper (ESS). he burst train is then transmitted through the core network, which is composed of Wavelength-Routed Switches () with periodical advanced scheduling, burst train reservation blocking recovery and RWA with delayed reservation functionalities. SS-OBS enforces two-way acknowledgement, i.e., the burst train can be sent only after it receives the notification from the core that its Quality of Service (QoS) 0-7803-9277-9/05/$20.00/ 2005 IEEE 524

requirement can be satisfied. he switches used in SS-OBS are all-optical switches, which can provide the transparent transmission in optical domain along the entire route. By forming the traffic streams into burst trains, SS-OBS can support the sub-wavelength traffic grooming for both CBR and VBR traffic. Furthermore, unlike the most OBS-based scheme, SS-OBS does not require the equipped with wavelength conversion and buffering. he detailed control procedure is described in the following subsections. Station 1 Burst rains Edge Devices Photonic Switching Core Stream Station 2 Stream Edge control Aggregation/ Shaping ESS ESS Edge control Aggregation/ Shaping Periodic Advanced scheduling Core Control Burst rain Reservation Blocking Recovery Fig.1 SS-OBS System Architecture Multiplexed Burst rains Station 3 RWA with Delayed Reservation A. Core Control 1) Routing and Wavelength Assignment with Delayed Reservation he routing and wavelength assignment (RWA) in SS-OBS is done similarly as it in Wavelength-Routed network. In this paper, we assume there is a central controller. Although the implementation of such central controller may bring the scalability problem, it essentially eliminates the burst reservation contention, which is an unavoidable issue in distributed controlled network. With the help of admission control, central controlled scheme can provide the delivery guarantee of burst trains. On the other hand, the scalability problem can be solved by using hierarchy RWA. o achieve the best trade-off between efficient utilization and processing time, we will implement alternative routing (pick the route among a few pre-selected routes), segmentation (the selected route is divided into several segments based on the wavelength conversion ability of ), and First-Fit wavelength assignment (the wavelength is assigned for each segment). If the routing and wavelength assignment can be found successfully, each along the selected route will be notified, and the reservation table is then updated. With the knowledge of the propagation and switching delay along the route, the central controller can calculate the exact time of data burst arriving at a particular intermediate. hen it will forward this information to switch controller, so that the switch will only reserve the bandwidth when it is necessary (usually termed as delayed reservation in OBS-based scheme). In the mean time, an acknowledgement will be sent to the edge to notify the success of reservation. 2) Periodic Advanced Scheduling SS-OBS employs periodic advanced scheduling to provision sub-wavelength traffic stream. Upon the successful route and wavelength assignment for the first burst, the along the selected route will periodically set the switching fabrics to accommodate the successive bursts. Such feature eliminates the need for the later-formed burst to send the control packet, and accordingly, there is no extra processing in the central controller. 3) Burst rains a) Aggregation ime SS-OBS tries to time-multiplex bursts into same wavelength. Suppose bursts A and B are time-multiplexed for transmission at the same link, the wavelength holding time of each burst should at least be composed of the switching time, link propagation time, and its own transmission time. he overhead per burst is the switching time and propagation time, which is determined by the physical devices, not by the burst transmission time. o increase payload to overhead ratio, the transmission time for each burst should be increased, which is equivalent to increase the burst aggregation time. Although it would be better to increase the aggregation time for each burst, the aggregation time cannot be infinitely large. he upper bound of the aggregation time is determined by the end-to-end delay requirement. (Refer to section 3 for detailed derivation) b) Burst rain Overlapping Avoidance For different traffic streams, the end-to-end delay requirements could be different, which implies their aggregation times are also different. But when two burst trains with different aggregation times share the same link and wavelength, an overlapping may occur. o eliminate such overlapping, we need to design a scheme where a common period exists for all traffic streams. We defined this common period as the timeframe in this paper. By introducing timeframe, if a burst train can be supported in the first timeframe, it can also be supported in successive timeframes. Shaped raffic B Shaped raffic A Multiplexed traffic Fig.2 Periodical Scheduling and imeframe he timeframe is bounded by the maximum end-to-end latency requirement for all traffic streams. Each stream, depending on its latency requirement, may reserve one or more bursts per timeframe. he more bursts it reserves per timeframe, the smaller its aggregation time, thus the more 525

stringent latency requirement can be met. Meanwhile, the different delay requirements of different streams can still be satisfied. One simple temporal diagram is given in Fig. 2, in which refers to the timeframe. raffic A has more stringent delay requirement, so it reserves two burst per timeframe. c) Burst rain Reservation Blocking Recovery When two requests try to reserve the same link and same wavelength at the same time, there will be a contention. In the proposed SS-OBS scheme, such contention can be avoided by time-shifting the blocked burst train. More specifically, the central controller will change the offset times of competing burst trains, and try to avoid the contention. If the controller cannot find an appropriate offset time to accommodate the request, the whole traffic stream will be rejected. Fig. 3 gives an example of burst train reservation blocking recovery. When traffic A requests the bandwidth, there is no contention, and the offset time t B (A) is set as the minimum value, which is the sum of two-way signaling time between edge and central controller and the time needed to process the request. When traffic B s request arrives, the controller will add extra delay to its offset time to avoid the contention. raffic A time raffic B t B (A) Burst ransimission t B (B) Fig.3 Burst rain Reservation Blocking Recovery time time B. Edge Aggregation and Shaping Control Fig. 4 shows the ingress edge structure of SS-OBS. Each sub-wavelength CBR traffic stream will be shaped into a periodical burst train, which is composed of a series fixed- sized bursts. Such feature would enable the periodical reservation. But to accommodate the traffic with different bandwidth requirement, the burst size of different streams can be different. o support sub-wavelength VBR traffic, a periodic burst train will be reserved for each stream in advance, and VBR packets will fill up the reserved bandwidth in each timeframe. Note in the VBR case, the data aggregated in each inter-burst interval could vary, but the reserved burst size is fixed for all the burst from one stream. herefore it is very important to find an appropriate burst size based on the traffic profile. If the burst size is too large, there would be too much vacancy in the reserved bandwidth; if the burst size is too small, a lot of packets would be dropped because of insufficient burst size. o further decrease the data loss rate, the packets from same source and destination pair will share their reserved bandwidth with statistical multiplexing. sub-wavelength CBR/VBR data streams... Stream Buffers Light-path Scheduler Burst Generator Fig.4 SS-OBS Edge Architecture resource reservation and scheduling signaling synchronous optical bursts synchronous optical bursts III. SYSEM MODELING A. Description he network is modeled as a connected graph G ( V, E), where V is the set of network nodes, E is the set of links. Denote Λ as the set of wavelengths supported in each link, and W as the number of wavelengths per link, W = Λ. Assume each node (optical switch) has switching delay of t sw and assume there is no FDL in the optical switch. Assume each link represents a pair of fibers with opposite transmission directions. For dynamic sub-wavelength CBR and VBR streaming traffic, it is assumed that traffic arrival follows Poisson distribution with arrival rate of λ, and the requested service duration of each successful traffic request follows exponential distribution with unit mean. Assume that both traffic source s and traffic destination d are randomly selected among all nodes with uniform distribution. For CBR traffic, denote traffic bandwidth as R, which is a fraction of the bandwidth of wavelength (denoted as C ). For VBR traffic, Poisson on-off model is assumed: the status of data transmission follows on-off alternation, and the length of each on/off duration follows exponential distribution with means of 1/ µ on and 1 / µ off, respectively. he data rate of VBR traffic during ON status is denoted as R on ; and the data rate during OFF status is 0. Alternative routing and First-Fit [11] wavelength assignment are assumed in this paper. SS-OBS supports both CBR and VBR dynamic traffic through edge traffic shaping and periodic core network resource reservation. he objective is to minimize the overall traffic blocking with the constraint of channel capacity and end-to-end traffic delivery latency, which is explained as follows. B. Channel Capacity Constraint Assume as the length of timeframe in SS-OBS system; and assume δ as the temporal guard period between different sub-wavelength channels, which should be at least as large as the switching delay t sw. For a given 526

traffic t i, denote R avg ( t i ) as the average traffic bandwidth requirement (for CBR traffic R t ) = R( t ) ). Denote avg ( i i τ ( t i ) as the sum of burst size for traffic t i in one timeframe; obviously, Ravg ( ti ) τ ( ti ) =. (1) C Let { t, t2,, } t K 1 be the maximum set of sub-wavelength traffic requests that can be accommodated on a wavelength, then K i= 1 τ ( t ) K δ. he K δ efficiency η is: η =. (2) Specially, ifτ ( t1) = τ ( t2 ) = = τ ( ti ) = τ 1 i K, then K τ K δ thus η K = = = η τ + δ τ τ η τ and the efficiency η =. (3) τ + δ C. End-to-End Latency Constraint Given the QoS specification t QoS, which is the end-to-end delay guarantee. We need to determine the inter-burst interval int, such that the maximum possible end-to-end delay is within the guaranteed QoS specification t. QoS For a given traffic x, the end-to-end delay includes the following components: t S ( : Burst shaping delay, which is the time needed to shape a burst for traffic x ; t ( : Burst transmission delay, which is the burst length of traffic x ; t P (l) : Burst propagation delay on link l, burst propagation delay is the constant value for a particular link. hese delay components are shown in Fig.5. As shown in Fig.5, max{ t ( + t ( } S = i C R int Denote the total delay as t otal (, max{ t = max{ t = int + otal S ( } ( + t t P l route( ( l) ( + t P l route( ; ( l)} (4) o satisfy the end-to-end delay requirement, max { total } tqos. herefore, { total ( } + t P ( l tqos max = (5) i.e. int ) l route( int tqos t P ( l) l route(, which can be regarded as the upper bound of the system parameter int. Recall that timeframe is the largest inter-burst interval. raffic A raffic B Aggregated Bursts Propagation Delay B1 t (B) A1 B1 A1 t P δ t (A) δ t S (B) int int t S (A) B2 Fig.5 SS-OBS Delay Analysis Diagram IV. SIMULAION RESULS o compare the performance of blocking probability and utilization of the proposed SS-OBS against existing WR-OBS and JE-OBS, simulations are conducted for both CBR and VBR streaming traffic. he typical 14 node NSF network topology (Fig.6) is adopted and the parameters of the topology model are: Number of nodes N =14; Number of links E =40; Number of wavelength per link W = 4; and the bandwidth of each wavelength C =2.5Gbps. For CBR traffic, assume each traffic requests bandwidth of 50Mbps; for VBR traffic, assume each traffic request has the same expected value of ON and OFF durations, and that the bandwidth requirement of ON state is 62.5Mbps. he maximum end-to-end latency is 100msec for each traffic request. 5000 traffic requests are simulated for each sample point. he switching delay is assumed to be 3 msec, which is a typical value for all-optical switch. For JE-OBS, we assume that there is no buffer in the core network. his is reasonable since the optical buffer is still expensive thus unlikely to be implemented at every node. Fig.6 14-node NSF Network opology o further investigate the performance of the three schemes at different convertibility ratio of core network, A2 B2 δ A2 527

the simulation is done in three cases: no convertibility, where core node does not support wavelength conversion; sparse convertibility, where we randomly picked 3 wavelength convertible nodes out of 14 (in the simulation these 3 nodes are node 2, 6 and 12), all other node will not support wavelength conversion; full convertibility, where every core node is wavelength convertible, although this may not be real, since JE-OBS is based on such assumption, we will still do some simulation in this case. he results are presented in Fig. 7 and Fig. 8. Fig.7 Blocking and Utilization versus raffic Arrival Rate (CBR raffic) Fig.8 Blocking and Utilization versus raffic Arrival Rate (VBR raffic) 528

Fig.7 shows the blocking probability and utilization over the average traffic arrival rate for CBR traffic. In full convertibility case, as expected, JE-OBS achieves the best overall performance. he reason behind this is that JE-OBS can allow each burst try its luck to request the bandwidth, while SS-OBS will block all the bursts in a traffic stream if the first burst fails to find a RWA. hough WR-OBS has similar characteristic, it does not implement the Delayed-Reservation as SS-OBS and JE-OBS, and according to the simulation traffic, the improvement brought by Delayed-Reservation is obviously bigger. But when only some nodes, or even not a single node in core support wavelength conversion, the performance of JE-OBS will be greatly degraded. In these cases, SS-OBS will outperform both JE-OBS and WR-OBS. his proves the flexibility of proposed SS-OBS. Also note the primary motive of SS-OBS is not to achieve the best performance, but to provide the QoS guarantee for each accepted traffic stream, even in extreme case where full convertibility is assumed, the decrease of the performance is still an acceptable trade-off. he performance in VBR traffic scenario is shown in Fig. 8. he similar conclusion as CBR scenario can be drawn. SS-OBS can provide satisfactory performance regardless of the convertibility ratio in core network. It can also be seen from Fig. 8 that the performance is slightly degraded compared to CBR scenario. he reason for this is WR-OBS and JE-OBS implement advanced signaling, the actual size of the aggregated data may not equal the reserved burst size, therefore it involves either insufficient or inefficient reservation. he SS-OBS sets the burst size from the same traffic to be the same, so its performance will be degraded slightly more in VBR scenario. [3] Wei, and R. McFarland, Just-in-time signaling for WDM optical burst switching networks, IEEE J. Lightwave echnology, vol.18 no. 12, pp. 2019-2037, Dec. 2000. [4] M. Yoo and C. Qiao, Just-Enough-ime (JE): A high speed protocol for bursty traffic in optical networks, Proc. IEEE/LEOS Conf. on echnologies For a Global Information Infrastructure, pp. 26-27, Aug. 1997. [5] J. urner, erabit burst switching, J. High Speed Networks, vol. 8, no. 1, pp. 3-16, 1999. [6] J. eng, G. N. Rouskas, A Comparison of the JI, JE, and Horizon wavelength reservation schemes on a single OBS Node, Proc. 1st Workshop on OBS, 2003. [7] M. Duser, and P. Bayvel, Analysis of a dynamically wavelength-routed optical burst switched network architecture, J. Lightwave echnology, vol. 20, no. 4, pp.574-585, 2002. [8] Eugene Kozlovski, Polina Bayvel, QoS performance of WR-OBS network architecture with request scheduling, Proc. IFIP 6 th Conf. Optical Network Design Modeling (ONDM 2002), pp. 101-116, 2002. [9] S. Sheeshia, and C. Qiao, Synchronous optical burst switching, Proc. Int l Conf. on Broadband Networks (BROADNE 2004), pp. 4-13, 2004. [10] M. Elhaddad, R. Melhem,. Znati, and D. Basak"raffic shaping and scheduling for OBS-based IP/WDM backbones," Proc. SPIE Optical Networking Commun. Conf. (OptiComm 2003), Dallas, X, vol. 5285, pp. 336-345, Oct. 2003. [11] H. Zang, J. P. Jue, and B. Mukherjee, A review of routing and wavelength assignment approaches for wavelength-routed optical WDM networks, SPIE/Baltzer Optical Networks Magazine (ONM 2000), vol. 1, no.1, Jan. 2000. V. CONCLUSIONS In this paper, we proposed a new scheme named Synchronous Stream Optical Burst Switching (SS-OBS). SS-OBS can provide QoS guarantee of data loss rate and end-to-end latency to sub-wavelength CBR and VBR streaming traffic. Some issues in the design of SS-OBS are discussed in detail, such as the periodical advanced scheduling, burst train reservation blocking recovery. he performances of blocking probability and network utilization are compared through simulations with those of JE-OBS and WR-OBS for both CBR and VBR traffic. Simulation results show that SS-OBS can provide satisfactory performance regardless of the convertibility ratio in core network. REFERENCES [1] C. Qiao, and M. Yoo, Optical burst switching (OBS)-A new paradigm for an optical internet, J. High Speed Networks, vol. 8, no. 1, pp. 69-84, 1999. [2] Y. Xiong, M. Vandenhoute, and H. Cankaya, Control architecture in optical burst-switched WDM networks, IEEE J. Select. Areas Commun., vol.18 no. 10, pp. 1838-1851, Oct. 2000. 529