Packet Rescheduling in Real-Time Using Token-passing Protocol in WDM Ring Access Networks

Transcription

1 Tamkang Journal of Science and Engineering, Vol. 9, No 2, pp (2006) 107 Packet Rescheduling in Real-Time Using Token-passing Protocol in WDM Ring Access Networks I-Shyan Hwang*, San-Nan Lee and Yen-Pin Kuo Department of Computer Engineering and Science, Yuan-Ze University, Chung-Li, Taiwan 320, R.O.C. Abstract This paper proposes a novel real-time packet rescheduling algorithm using a token-passing protocol to achieve QoS on WDM (Wavelength Division Multiplexing) Ring Access Networks. The proposed rescheduling algorithm is called PEM (PDS-EATS-MTD) algorithm. The PDS (Priority- Differentiated Scheduling) algorithm deals with real-time packets and allows them to be transmitted first, such that the front line of the prescheduled nonreal-time packets can be inserted into the queue. The EATS (Earliest Available Time Scheduling) algorithm selects the earliest available data channel, independently of the availability of the destination. The MTD (Minimum Time Difference) algorithm selects the minimum-time-to-wait channel to transmit the nonreal-time packet and is quick to establish the path of the real-time packets. The updated information including the Scheduled Data Table and the Channel Available Time Table, is then broadcasting to other access nodes using token-passing protocol to maintain the status of packet scheduling/rescheduling consistency. Overall, the PEM algorithm outperforms NPEM, PEE, and EATS in terms of average delay time for different traffic loads and number of channels. Key Words: Packet Rescheduling, Token-passing Protocol, Quality of Service, WDM Ring Access Network, PEM Algorithm 1. Introduction The growth of next-generation multimedia and realtime applications is demanding more and more from the access networks, such as that they support broadband Internet access and offer good Quality-of-Service (QoS) [1,2]. Recently, many MAC (media access control) protocols [3] have been proposed in the literature [3 6]. A scheduled approach to optical flow switching is proposed in [7]. This method divides time into fixed-length timeslots and schedule transmissions to reach viability and scalability of a scheduled connection setup approach. Mondiano [8] proposes a centralized scheduler that uses first-come-first-serve input queues for unicast *Corresponding author. ishwang@saturn.yzu.edu.tw traffic and a window selection policy to eliminate headof-line blocking, while multicast traffic is scheduled using a random algorithm. Chu [4] proposed a PDS algorithm to reschedule realtime packets to preempt nonreal-time packets in the single-hop WDM (Wavelength Division Multiplexing) star network. He showed that the PDS algorithm can dramatically reduce the delay of real-time packets, without sacrificing the performance of nonreal-time packets. However, long-haul access networks using WDM [9 11], shown in Figure 1, become commercially available as the backbone. The WDM ring access network actually consists of two independent networks a feeder network and distribution networks, which are connected by access nodes. The access node is composed of reconfigurable Optical Add/Drop Multiplexer (OADM) [12]. It routes full opti-

2 108 I-Shyan Hwang et al. Figure 1. WDM access ring network. cal-wavelength channels and data packets inside the wavelength channels toward their destination nodes. The reconfigurable OADM is selected to direct optical signals to one of the two routers or to completely bypass the routers. The characteristics of OADM are as follows: (1) the transmission of packets among nodes is fully optical, (2) packets in different channels can be transmitted or received simultaneously, and (3) each distribution network aggregates data from local end-users. This paper proposes a PEM (PDS-EATS-MTD) algorithm that uses a token-ring protocol to schedule/reschedule real-time packets and schedule nonreal-time packets on the WDM ring access network, which includes one control channel and data channels. The access node with token can reschedule real-time packets based on the PEM algorithm, and the control packet, which includes Scheduled Data Tables and Channel Available Time Tables, is updated and are broadcasted to the other access nodes through the control channel, to keep the information of the packets transmission consistent. The token is released when the access node has finished the scheduling/rescheduling packets. If the access node has no token, then it only transmits or receives prescheduled packets to/from data channels. The proposed PEM algorithm incorporates three algorithms, PDS, EATS [5], and MTD algorithms [7], to handle the different types of packets (real-time packets and nonreal-time packets). The PDS and EATS algorithms together deal with the realtime packets: the PDS algorithm allows real-time packets to preempt the prescheduled nonreal-time packets and the EATS algorithm selects the earliest available data channel to transmit real-time packets. The priority of the nonreal-time packet in the queue is upgraded after some time, to avoid the infinite waiting. The MTD algorithm is used to select the minimum-time-to-wait channel to transmit nonreal-time packets. Hence, the purpose of the MTD algorithm is to improve the channel utilization to help allow real-time packets to be transmitted through more available channels and to accelerate the establishment of lightpath efficiently. For example, there are two data channels (a and b). The available times of them are 10 and 0. When a packet is transmitted by EATS algorithm and the starting time is 6, it will be assigned channel b because channel b is the earliest available channel. Otherwise, the packet is assigned channel a by MTD algorithm since its is the minimum-time-to-wait channel. The rest of this paper is organized as follows. Section II describes the design and operation of the system, and the proposed PEM algorithm. Section III presents the simulation results and compares the system performance with the performance of three other algorithms NPEM, PEE and EATS algorithms in terms of average delay time and average channel utilization. Section IV draws the conclusions. 2. Design and Operation of System The WDM Ring Access Network has n access nodes and w channels (a control channel and the rest are data channels). Each access node stores information about the transmission of all packets from access nodes i, i = 1, 2,,n, in two tables the Scheduling Data Table (SDT) and the Channel Available Time Table (CATT), i.e. SDTi and CATTi, i = 1,2,,n. Figure 2(a) shows the format of SDT for each packet. Each packet has a packet id (PI) with a packet length (PL) and a packet priority (PP). The transmission from the source node (SN) to the destination node (DN) is through the selected channel (SC) and the delay time is from the starting time (ST) to ending time (ET). The CATT is a table to record the available time (AT) when the channel is free, shown in Figure 2(b). These two tables, restored in each access node, must be kept consistent at all times. Figure 3 details the operation of the system as follows. Step 1. When an access node holds the token and a new packet is generated, it determines the priority of packet(s) is high or low. If the packet is high, go to step 2(a), else go

3 Real-Time Packet Rescheduling in WDM Ring Access Networks 109 Packet Id Channel Id Available Time Source Node Wavelength 1 Destination Node Wavelength 2 Selected Channel Wavelength 3 Packet Length Starting Time Ending Time Wavelength n-1 Packet Priority Wavelength n (a) (b) Figure 2. (a) Format of SDT for each packet and (b) CATT in each access node. Figure 3. Flowchart of system operation. to step 2(b). Step 2(a). The real-time packet preempt the prescheduled nonreal-time packet and the lightpath is selected by EATS algorithm. Step 2(b). The MTD algorithm allocates the channel to transmit this nonreal-time packet. Step 3. The latest information is transmitted to all access nodes and the token is released to the next access node. Each access node with a token-ring can update its SDT and CATT, based on the proposed PEM (PDS- EATS-MTD) algorithm. After the rescheduling algorithm is executed, the updated tables will be broadcasting to all other access nodes and the token is released to the next access node through the control channel. Thereafter, the scheduled/rescheduled packets can transmit according to the updated SDT tables. The PEM algorithm is described as follows: If there are one or more unscheduled packets, execute the following substeps. Otherwise, go to step 2. Step 1. If the priority of packet is high (PP = 0), execute the following substeps. Otherwise, go to b1. a1: Find reschedulable packets based on the ST and PP of SDT. a2: Preempt these packets. a3: Reserve the wavelengths of each link from source to destination using EAC algorithm. a4: Update SDT. a5: Select another unscheduled packets and go to step 1. Otherwise, go to step 2. b1: Select a wavelength which has the minimum difference according the channel s available time (AT) and the packet transmission time (ST). b2: Update SDT. b3: Select another unscheduled packets and go to step 1. Otherwise, go to step 2. Step 2. Broadcast the SDT and CATT to each access node. Step 3. Release token to next access node. Step 4. stop. Five access nodes are assumed in the WDM Ring Access Network to be connected by unidirectional fiber with three channels; one is the control channel (channel 0) and the others (channel 1, 2) are data channels. Nine packets in all access nodes are presumed to be scheduled/rescheduled by the PEM algorithm, as shown in Table 1(a). For instance, three packets P1, P2 and P3 have lengths of 3, 8, and 2 in access node 1. The real-time packet is presented by the bold and italic type and need to be assigned a lightpath from the source to the destination.

4 110 I-Shyan Hwang et al. The nonreal-time packet is only assigned a channel from the source to the next access node. P1 is a real-time packet with a packet priority of 0; P2 and P3 are nonrealtime packets with packet priorities of 1. The transmission of P1, P2 and P3, from the source node 1 to destination nodes 4, 5, and 3, begins at starting times 0, 1 and 4. Two nonreal-time packets, P4 and P5, with packet lengths of 2 and 4 are present in access node 2. The transmission of P4 and P5, from source node 2 to destination nodes 1 and 5, begins at starting times 1 and 6. Two nonreal-time packets, P6 and P7, with packet lengths of 3 and 2, are transmitted from source node 3 to destination nodes 4 and 2, and the starting times are 3 and 4, respectively. Two realtime packets, P8 and P9 have lengths 2 and 4, are destined to nodes 2 and 5 from source nodes 4, and both have starting times of 1. Access node 5 has no packet to transmit. The system operation is described in detail as follows. As access node 1 owns the token, it will schedule the packets, P1, P2 and P3. The real-time packet P1 with packet lengths 3 is transmitted from access node 1 to access node 4, and the SCs of SDT1, SDT2 and SDT3 are all 1. According to the EATS algorithm, the starting time and the ending time of SDT1 are 0 and 2, respectively; the starting time and the ending time of SDT2 are 3 and 5, and the starting time and ending time of SDT3 are 6 and 8. For the nonreal-time packet P2 with a packet length 8, the starting time is 1, and channel 1 from access node 1 to access node 2 is then occupied by P1 at that time; then P2 is transmitted through channel 2 from access node 1 to access node 2. The ending time of SDT1 is 7. Since the channel 1 is occupied by P1 and channel 2 is occupied by P2 at this moment, the minimum time difference of available channel 1 is 0 based on the MTD algorithm, and P3 is transmitted accordingly. Therefore, the SC and the ET

5 Real-Time Packet Rescheduling in WDM Ring Access Networks 111 of SDT1 is 1 and 5, respectively. Before the token is released to access node 2, the SDT1, STD2 and STD3 are broadcasting to update the tables restored in each access node, shown in Table 1 (b). Access node 2 has two nonreal-time packets P4 and P5 to be scheduled. Since the channel 1 is reserved for transmitting P1 from starting time 3 to ending time 5, P4 and P5 are scheduled to be transmitted through channel 1 using the MTD algorithm. The SC and ET of P4 are 1 and 2, and the SC and ET of P5 are 1 and 9 in the SDT2; then the new SDT2 will be broadcasting and update the old SDT2 restored in other access nodes. The same process is applied for access node 3 when it receives the token. The SC and ET of P6 are 1 and 5, SC and those of P7 are 2 and 5 in the SDT3. Table 1(c) summarizes the processes from access node 1 to access node 3. When access node 4 receives the token, it may schedule/reschedule two real-time packets P8 and P9. The status of P8, scheduled to be transmitted from access node 4 to access node 1, is that the SCs of SDT4 and SDT5 are both 1 and the ETs of SDT4 and SDT5 are 2 and 4. Then, P8 is rescheduled to preempt the prescheduled P3 by PDS-EATS algorithm from access node 1 to access node 2 and The ST and ET of P8 are 5 and 6 in SDT1. The ST and ET of P3 in SDT1 are delayed to 7 and 8, after which P9 is transmitted through channel 2 from access node 4 to access node 5 and the ET of SDT4 is 4. Table 1(d) shows the final results. In this example, the overall real-time packet delay with PEM is =19,whereas with EATS, it is = 21. Clearly, PEM has a lower overall real-time packet delay than EATS. 3. Simulation results and Discussion We simulate a dynamic network environment to evaluate the performance of OADM and its corresponding scheduling algorithms, using the proposed algorithms PEM and the other three different algorithms. The simulation environment is setup as follows: 1. Numbers of access nodes (N) are 32 and Numbers of channels are (C) 4, 8, 16, 32 and The bandwidth of each channel is 2.5 Gbits. 4. The packets are randomly generated. 5. The average size of real-time packet is smaller than that of nonreal-time packet. 6. Packet size is variable and the maximum packet size is smaller than 5 Mbits. 7. The ratio of the number of real-time packets to number of nonreal-time packets generated randomly is denoted as R_NR. 8. The ratio of the number of channels to number of nodes is denoted as C_N. 9. The delay time of the token is ignored. The traffic load is defined as the total throughput to the total bandwidth. The delay time of the packet transmission is defined as the duration from the time at which a packet is scheduled to be transmitted to the time at which the transmission of the packet is finished. The simulation compares the proposed PEM algorithm with three algorithms NPEM (with no PDS algorithm), PEE (PDS- EATS-EATS) and EATS (Earliest Available Time Scheduling algorithm) [5] in terms of average delay time for different traffic loads and the different number of channels. The major objectives of the simulations are: we demonstrate the superior performance of the PDS algorithms by showing that the average delay of the real-time traffic can be reduced significantly and demonstrate that the proposed MTD works much better than EATS in terms of average delay. 3.1 Average Delay Time vs. Data Arrival Rate Figure 4 depicts the average delay time vs. traffic loads for different C_Ns = {1:1, 1:2, 1:4, 1:8} when N = 32 and 64. The access network with PEM algorithm shows the best network performance. In Figure 4(a), we set N = 32 and have the following observations. (1) It is interesting to notice that average delay time for four different algorithms is PEM < PEE < NPEM < EATS for different C_Ns = {1:1, 1:2, 1:4, 1:8}. The reason is that the PEM and PEE algorithms can arrange the high-priority packets to be served faster than the other packets, therefore, resulting in delay reduction. (2) The PEM algorithm can reduce the average delay time significantly. For example, when arrival rate is 0.9 and C_N is 1:4, the average delay time with PEM is one half less than that with EATS. (3) Even without PDS, our proposed algorithm NPEM leads to a smaller average delay time than EATS. It is because that the MTD algorithm changes the priority of packets from low to high when the delay

6 112 I-Shyan Hwang et al. Figure 4. Average packet delay time with four channels vs. arrival rates for different C_Ns = {1:1, 1:2, 1:4, 1:8} and number of nodes, when (a)n=32and(b)n=64. time of low-priority packet is too long. In Figure 4(b), the number of access nodes is 64 and for different C_Ns are {1:1, 1:2, 1:4, 1:8}. We make the following observations. (1) The simulation results reveal that the average delay time associated with the PEM algorithm is less

7 Real-Time Packet Rescheduling in WDM Ring Access Networks 113 than those associated with other three algorithms, which are independent of the arrival rate because the PDS algorithms can arrange for real-time packets to be transmitted faster than the nonreal-time packets. PEM and PEE algorithms significantly outperform the other two algorithms when the traffic load is heavy. For example, the average delay time obtained using PEM is 26.5% lower than that obtained using NPEM, and is 50% lower than that obtained using EATS when the arrival rate is 0.9. (2) For arrival rate = 0.2, as we have already seen, the PEM algorithm (C_N is 1:8) outperforms the EATS algorithm (C_N is 1:4) by an average delay time of 35% or less. (3) The MTD algorithm achieves the best performance than the other three algorithms, NPEM, PEE, and EATS for nonreal-time packets. The average delay time obtained with PEM is 12% lower than that obtained with PEE when the arrival rate is 0.7. We can see that the average delay time is similar to EATS, in that the NPEM algorithm does not handle realtime packets. However, it achieves a better performance than EATS. Hence, the NPEM algorithm can be used for a system without the need to handle real-time traffic. Overall, the simulation results show that the PEM algorithm for real-time packets and nonreal-time packets outperforms the other three algorithms. 3.2 Average Channel Utilization vs. Number of Channels Figure 5 compares the average channel utilization for different R_NRs (a) for 3:7 and (b) for 5:5, respectively, (N=64, R_NR=3:7) Average channel Channel Utilization Average Channel channel Utilization % 80.00% 60.00% 40.00% 20.00% 0.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Channel channel number Number (a) (N=64, R_NR=5:5) channel Channelnumber Number (b) PEM NPEM PEE EATS PEM NPEM PEE EATS Figure 5. Average channel utilization for different R_NRs (a) for 3:7 and (b) for 5:5, when N = 64.

8 114 I-Shyan Hwang et al. when N = 64. Figure 5(a) shows the PEM algorithm has the best average channel utilization than the other three algorithms due to rescheduling the real-time packet preempted to non-real time packets. Another important simulation result is that when C is 64, the average channel utilization with EATS is at least less 30% than the other three algorithms. It is because that more non-real packets lead to the call blocking of real-time packet and the average channel utilization is decreased. Figure 5(b) shows the similar results that the average channel utilizations are decreased for all algorithms, but the PEM algorithm still performs the best. That is because PEM algorithm not only deals with real-time packet to preempt prescheduled nonreal-time packets, but also upgrades the priority of nonreal-time packet from low to high when the waiting time is too long. 4. Conclusion This paper proposes the PEM algorithm, using a token-passing protocol, to reach the QoS of real-time packets on the WDM ring access network. The system not only allows the real-time packets to preempt the nonrealtime packets using the PDS algorithm and chooses the proper channels to set up lightpaths of real-time packets using the EATS algorithm, but also uses the MTD algorithm to schedule non-real-time packets. Then, the updated information is broadcasting to all access nodes using the token-passing protocol. The simulation results show that the average delay time obtained with PEM algorithm is better than those obtained using the other three algorithms (PEE, NPEM, and EATS algorithm) for realtime packets when the traffic load is high. The MTD algorithm can increase the channel utilization. Therefore, under the same traffic load and same number of channels, the algorithms in order of increasing average delay from low to high is PEM, NPEM and EATS algorithm. Overall, the PEM algorithm outperforms the other three algorithms, NPEM, PEE, and EATS in terms of average delay time for different traffic loads and number of channels. References [1] Keepence, B., Quality of Service for Voice over IP, in Services Over the Internet What Does Quality Cost?, Ser. IEE Colloquium, No. 99, pp. 4/1 4/4 (1999). [2] Li, B. and Qin, Y., Traffic Scheduling with Per VC QoS Guarantee in WDM Networks, IEEE GLOBE- COM 98, pp (1998). [3] Jelger, C. S. and Elmirghani, J. M. H., A Simple MAC Protocol for WDM Metropolitan Access Ring Networks, IEEE CLOBECOM 2001, Vol. 3, pp (2001). [4] Chu, P. L. and Diao, J., Packet Rescheduling in WDM Star Networks with Real-time Service Differentiation, Journal of Lightwave Technology, Vol. 19, pp , (2001). [5] Jia, F., Mukherjee, B. and Iness, J., Scheduling Variable-length Messages in a Single-hop Multichannel Local Lightwave Network, IEEE/ACM Transactions Networking, Vol. 3, pp , (1995). [6] Bengi, K. and Van, H. R. CONRAD a Novel Medium Access Control Protocol for WDM Local Lightwave Networks Enabling Efficient Convergence of Realtime and Data Services, Proceedings on Local Computer Networks 2001, pp (2001). [7] Ganguly, B. and Chan, V., A Scheduled Approach to Optical Flow Switching in the ONRAMP Optical Access Network Testbed, OFC 2002, pp (2002). [8] Modiano, E. and Barry, R., A Novel Medium Access Control Protocol for WDM-based LAN s and Access Networks Using a Master Slave Scheduler, Journal of Lightwave Technology, Vol. 18, pp , (2000). [9] Hwang, I.-S., Lee, S.-N. and Lin, Y.-S., Stochastic Generative Model of Cost Effective OADM Using the Neural Network in a WDM Access Network, 1 st PNC/ 6 th JCIS 2002, pp (2002). [10] Lee, S.-N. and Hwang, I.-S., Stochastic Generative Model of Cost-effective OADM Using a Three-dimensional Neural Network in a WDM Access Network, The Fifth Pacific Rim Conference on Lasers and Electro-Optics (CLEO/PR 2003), pp. 391 (2003). [11] Takachio, N., Suzuki, H., Fujiwara, M., Jun-ichi Kani, K., Iwatsuki, K., Yamada, H., Shibata, T. and Kitoh, T., Wide Area Gigabit Access Network Based on 12.5 GHz Spaced 256 Channel Super-dense WDM Technologies, Electronics Letters, Vol. 37, pp (2001). [12] Mariconda, A. and Merli, S., An Optical Add-drop Multiplexer (OADM) Node Architecture in a Fully Transparent Self Healing Ring Network, 22 nd European Conference on Optical Communication ECOC 96 (1996). Manuscript Received: Apr. 18, 2005 Accepted: Nov. 2, 2005