Signaling schemes for distributed connection management in wavelength-routed optical mesh networks

Transcription

1 Signaling schemes for distributed connection management in wavelength-routed optical mesh networks Lu Shen, Xi Yang and Byrav Ramamurthy Department of Computer Science and Engineering University of Nebraska - Lincoln Lincoln, NE , U.S.A {lshen, xyang, byrav}@cse.unl.edu Abstract The next-generation optical transport network will evolve from point-to-point connectivity to mesh networking, which can provide fast and automatic provisioning with enhanced flexibility and survivability. Signaling is used to support connection setup, maintenance, and teardown in such a network. In this paper, we study the performance of two hop-by-hop and one parallel signaling schemes in wavelength-routed optical mesh networks. Based on the sequence between optical crossconnect (OXC) switching and signaling message processing, we classify hop-by-hop signaling into two types that comply with the requirements of GMPLS signaling protocols. These two types are forward before switching configuration (FBSC) and forward after switching configuration (FASC). Also, we propose a parallel signaling scheme that is different from the existing hop-by-hop GMPLS signaling protocols. Considering OXC architectures and traffic patterns, we compare the FBSC, FASC, and parallel signaling schemes using simulation experiments, in terms of network blocking probability and reservation time. The simulation data reveal that the performance of a signaling scheme depends on the nature of the signaling as well as the network setting (e.g., the OXC architecture and traffic pattern). We analyze reasons for this result and discuss tradeoffs between these signaling schemes. This work offers some insight into designing an efficient signaling protocol for wavelength-routed optical mesh networks. 1

2 1 Introduction The next-generation optical transport network will evolve from point-to-point connectivity to WDM mesh networking. In such a network, it is desirable to design a more flexible and efficient optical core using the emerging reconfigurable optical components, such as reconfigurable crossconnects (OXCs), reconfigurable optical add-drop multiplexers (OADMs), and tunable transmitters, receivers and filters. This requires more sophisticated network control and management (NC&M) functionalities to be incorporated into the optical layer. One network management functionality is connection management, which is referred to as the procedure of lightpath establishment (also called lightpath provisioning), maintenance, and teardown. The slow provisioning procedure in today s transport network cannot catch up with the speed of ever-increasing demands on data services. As a result, carriers are facing cost and scalability problems in their networks. Automatic connection management, in which lightpaths are established and torn down on demand, becomes a solution to these problems. In a wavelength-routed network deployed with reconfigurable OXCs at each network node, automatic provisioning can be implemented by dynamically configuring each OXC before the data transmission. There is an increasing consensus on leveraging the existing IP/Multiprotocol Label Switching (MPLS) routing and signaling protocols to realize automatic distributed connection management in the WDM optical transport network. The MPLS technology evolved from tag switching with the original aim to improve packet-forwarding efficiency of switching routers [1]. Generalized Multiprotocol Label Switching (GMPLS) was formulated later by extending MPLS to encompass time-division (e.g., SONET ADMs), wavelength-division (e.g., optical wavelength or lambda), and spatial switching (e.g., incoming port or fiber). In a GMPLS-based WDM network, enhanced IP interior gateway protocols (e.g., extended OSPF-TE) can be used to flood network state information, so that each node in the network can have a global knowledge of the network states. Based on the global information, a routing and wavelength assignment (RWA) algorithm can be implemented at the source node [2] [3] [4]. Then, the lightpath setup and teardown procedure can be implemented 2

3 by signaling protocols, such as the extensions to RSVP-TE and CR-LDP [5] [6] [7] [8] [9]. In this paper, we study two hop-by-hop signale schemes and our proposed parallel signaling scheme for distributed connection management, confining the signaling behavior to a single administrative domain with wavelength conversion capability. We organize the remainder of this paper as follows: Section 2 discusses the related work and our study. Section 3 gives a brief introduction to the logical OXC architecture and wavelength state transition diagram. Section 4 describes different signaling schemes to be compared in this paper and discusses the tradeoffs between them. Section 5 presents our simulation experiments and analyzes the results. Section 6 concludes this paper. 2 Related Work and Our Study 2.1 Related Work The authors in [10] proposed for the first time a distributed connection management protocol for wavelength-routed networks. The data structure, algorithm, and pseudo-code were described in detail. The procedure of lightpath establishment described in the paper can be divided into two phases: reservation and configuration setup. The reservation messages are sent out and processed at each controller in a parallel manner, while the setup messages proceed hop-by-hop to implement physical switch configuration after the successful reservation. Proofs of several properties, such as looplessness and successful lightpath setup and release, were also provided in the paper. Dynamic connection management has been demonstrated to be feasible in the prototype test-bed of the Multiwavelength Optical Networking (MONET) project [11], in which both centralized and distributed schemes were tested [12] [13]. The distributed connection management in MONET makes use of the so called Just-In-Time (JIT) signaling, which was originally proposed for WDM optical burst switching networks. One of the schemes for JIT, named as explicit setup and explicit release [14], can be used in circuit-switched networks. Other schemes set up or release wavelength resources at OXC nodes based on estimated times, which may potentially result in significant data loss in wavelength-routed networks. Earlier studies have used simulation to compare different signaling schemes for the distributed 3

4 connection management in wavelength-routed networks ( [15] [16] [17], [18], and [19]). The work in [15] compared two connection management protocols, link-state (proposed in [10]) and distributed routing, which employ different signaling schemes depending on the availability of global wavelength usage. Network blocking probability, bandwidth requirement, and connection setup time were used as the metrics for the comparison. The work in [16] evaluated network performance of different signaling schemes, such as dropping versus holding, source-initiated reservation (SIR) versus destination-initiated reservation (DIR), one-way reservation versus two-way reservation, simple control versus informed control, and spendthrift reservation versus frugal reservation. Assuming that each node does not have the information of the global wavelength usage, the paper compared different ways to probe and reserve available wavelengths in wavelength-continuous networks (the wavelength-routed networks without capability of wavelength conversion). The results in [16] showed that DIR yields the best performance, while introducing more complexity; one-way reservation, informed control, and frugal reservation improve network performance, but at the cost of greater complexity. 2.2 Our Study The work in [15] and [16] concentrated on how to collect/probe network resource information during the signaling procedure to implement the RWA computation for a better performance. This method has a relatively long reservation time for each connection request and is a good choice for a network where the knowledge of global network states is not available to each network node. By using extended Interior Gateway Protocol (IGP) routing in a wavelength-routed network, each node can have a global knowledge of the network states. In other words, there is enough information at a source node to execute an RWA. Therefore, our work focuses on how signaling schemes affect network performance by the nature of the signaling procedure, but not by the way of obtaining the network resource information. In the studies mentioned above, the simulations were conducted using traffic that follows Poisson distribution with a very short average service time, e.g., 100ms in [15] and 2.56ms in [16]. Such 4

5 service times are possible for some traffic flows with low granularity. However, in the context of optical transport network, traffic flows with low granularity merge into large aggregate streams at the core. It may not be suitable to simulate signaling effect by specifying the traffic service time according to the duration of a single low granularity flow injected from the edge. In the all-optical core, the demands on optical services, using one wavelength as the granularity unit, are more appropriate to be simulated as lightpath requests. The optical service is defined as the solution that a service provider sells to an end-user [20]. The service time may vary with different types of user applications, e.g., storage network, Ethernet inter-connectivity, and optical VPN. In our simulation, we use traffic following Poisson distribution, but vary optical service time in a range from hundreds of milliseconds to hours. All of the earlier work presented in Section 2.1, except for the work in [10], studied hop-byhop signaling schemes. Hop-by-hop signaling prevails because it is relatively easy to implement and complies with the current GMPLS signaling protocols. The hop-by-hop property of MPLS signaling is inherited from its predecessor (i.e., RSVP), which is proposed for supporting qualityof-service (QoS) in the Internet. Because the Internet is based on the best-effort packet switching and each IP router does not know the whole path of a packet, the RSVP reservation signaling is implemented in two main phases: one is to probe the path and available network resources; the other is to reserve the network resources on its reverse path. However, the optical transport network operates as a circuit-switched network, in which the source node may know the entire routing and wavelength assignment for a lightpath request with the aid of the extended IGP. Based on this fact, we propose a parallel signaling scheme, where signaling messages are sent out simultaneously. The parallel message forwarding is also used in the signaling procedure proposed in [10], where parallel signaling is used in the reservation phase while hop-by-hop signaling is used in all the other phases (the configuration and teardown phases). However, in our proposed parallel signaling scheme, the reservation and configuration phases are combined into one phase to reduce the total reservation time. Another difference of our study from previous work is that the OXC switching time and OXC 5

6 architecture are taken into consideration from an implementation perspective. In a distributed environment, the OXC switching time of various types of OXC architectures may be quite different. We compare the signaling schemes under different OXC architectures (see Section 3.1 for the classification of OXC architectures). We also address the signaling issues, arising from link management protocol (LMP) [26], which is introduced to maintain link connectivity status and may affect the performance of signaling in wavelength-routed optical networks (see Section 3.2). Taking into consideration the sequence between OXC switching and signaling message processing, we classify the hop-by-hop signaling schemes into two types: forward before switching configuration (FBSC) and forward after switching configuration (FASC). We conduct a series of simulation experiments to compare the effect of our proposed parallel signaling and the hop-by-hop signaling using FBSC and FASC under three types of OXC architectures (we will use FBSC and FASC to represent these two hop-by-hop signaling schemes in the rest of this paper). 3 OXC Architecture and Wavelength State Transition Diagram 3.1 Intelligent Reconfigurable OXCs Figure 1(a) illustrates the logical framework of control plane and data plane in the wavelengthrouted optical transport network, in which the OXC at each node is responsible for part of the network control and management. The user data are carried over the data plane, while the routing and signaling messages are exchanged over the control plane. UNI (user to network interface [21]) is the interface for the communication between clients and edge OXC nodes. A lightpath connection request is passed through UNI to one of the edge nodes, which is called a source node in this paper. The source node then starts the signaling procedure to reserve wavelengths and configure the OXCs for this connection. After the source node has confirmed that the reservation procedure is finished, it acknowledges the client, via UNI, to start data transmission over the data plane, where corresponding OXCs have been configured for the connection. Figure 1(b) depicts the logical architecture of the intelligent reconfigurable OXC. The switching fabric in the lower part of the figure is in charge of physically switching wavelengths from input ports 6

7 controller control plane Controller switching-request-waiting queue routing table UNI controller controller controller UNI network resources status message processor processor interface information from OSC Switching Request Switching Confirmation Clients IP ATM SONET OXC OXC OXC OXC data plane Clients IP ATM SONET fibers input ports Interface switching fabric output ports fibers (a) Control plane and data plane (b) OXC logical architecture Figure 1: Logical architecture of OXC and control plane. to output ports under the control of the controller shown in the upper part. The controller, which may be either embedded into the OXC or a separate part attached to the OXC, is responsible for the network management functionalities, such as routing, signaling, RWA, link connectivity management, etc. The switching-request-waiting queue in the controller stores the requests of OXC switching configuration submitted by signaling messages. A database maintains the information of both network resources and local resources. The routing table maintains the routing information for both the control plane and the data plane (Note that control plane and data plane may use different routing algorithms, see Section 4.1). Signaling messages can be carried on an optical supervisory channel (OSC), which may be a dedicated wavelength in the network. For instance, the OSC in MONET is an out-of-band wavelength that can be dropped into an ATM switching fabric at each OXC node and can also be multiplexed into output fibers [11] [12]. The message processor processes the messages carried on the OSC and then passes the useful information to the controller processor. According to the information passed from the message processor, the controller processor updates the network resources status, inserts OXC switching requests into the switching-request-waiting queue, and updates the routing table. The processor is also responsible for communicating with the OXC internal switching fabric to implement the physical switching. The switching requests in the switching-request-waiting queue are scheduled by the processor and passed into the switching fabric. After the physical switch- 7

8 ing configuration for a request is done, a switching confirmation message will be sent back to the processor. Therefore, the time for switching an OXC for one switching request is the sum of the waiting time in the switching-request-waiting queue and the physical switching configuration time of the OXC switching fabric. For lightpath establishment, the purpose of signaling is to reserve wavelengths and physically configure the OXCs on the lightpath. We consider the OXC switching time when comparing different signaling schemes, since it has impact on the reservation time. The switching time of an OXC, defined as the time for physically configuring the OXC, varies by several orders of magnitude from nanoseconds to milliseconds depending on its technology. For example, an OXC using the popular MEMS (Micro-Electro-Mechanical Systems) technology involves physical mirror movement (tilt) and may take tens of milliseconds for switching [23], while an OXC using hologram technology is said to have switching time of nanoseconds [24]. In this paper, we study networks using large-scale optical switching fabrics, whose switching time is normally on the order of milliseconds [23]. In an optical mesh network with distributed control management, an OXC may receive many switching requests at a time. However, not every OXC is able to handle all the switching requests simultaneously. As a result, a signaling messages may experience significant waiting time in the switching-request-waiting queue at each intermediate node. The OXC architecture determines whether an OXC has the capability of switching several wavelengths simultaneously We follow the classification of OXC architectures in [25] to categorize OXCs into three types: 1. Sequential cross-connect (or Sequential OXC), where physical switching of an OXC cannot begin until the previous one has completed; 2. Parallel cross-connect (or Parallel OXC), where each switching request is processed to implement physical switching immediately, without any waiting; 3. Batch cross-connect (or Batch OXC), where a number of switching requests are batched together to be processed simultaneously, but where each batch of commands must wait until the previous batch has been completed. 8

9 3.2 LMP and Wavelength State Machine Link management protocol (LMP) has been proposed to maintain the link connectivity information between neighboring nodes in GMPLS-based networks [26]. If out-of-band signaling is used, the control channel and data channels are separate. Thus, the status of the control channel may not be consistent with the status of the data channels. Furthermore, when wavelength channels are not in use, their connectivity status are not available if they are not powered. One of the main features of LMP is to verify the availability of each data channel between neighboring nodes. LMP at a node sends Test messages over all free outgoing data channels on a periodic basis. When a neighboring node receives the Test message, it replies with a TestStatusSuccess or TestStatusFailure message over the control channel. The detailed specification of LMP is beyond the scope of this paper. We are interested in how the employment of LMP affects signaling behavior in optical transport networks. When a signaling message arrives at a node to reserve a wavelength on one of its outgoing links, the wavelength may be under LMP verification. Two signaling behaviors may exist for this scenario: 1. the node holds the signaling message until the verification is done; 2. the node sends an EndVerify message to terminate the verification and immediately reserves the wavelength for the connection; In case (1), the reservation time may be longer due to the waiting time for the verification procedure. In case (2), the reservation time is the same as that of the signaling scheme without considering LMP. But the channel connectivity status may be out-of-date. The wavelength state transition diagram in the network with LMP is different from that in the network without LMP. Figure 2 (a) shows the wavelength state transition diagram without considering LMP. The state of a wavelength changes from Free to Reserved, when the node receives a PATH signaling message to reserve the wavelength (the PATH message will be explained in the next section). When the switching confirmation of the wavelength is received from the OXC 9

10 Bring up network Free Down Receive PATH signaling message Reserved Receive Release signaling message Detect Fault Switched Receive Switching Confirmation Bring up network Start LMP link connectivity verification Free Under Verification Receive TestStatusSuccess message Receive PATH signaling message Receive PATH signaling message Receive TestStatusFailure message Start LMP link connectivity verification Receive Release signaling message Down Detect Fault Reserved Receive Switching Confirmation Switched (a) wavelength state transition diagram without considering LMP (b) wavelength state transition diagram with consideration of LMP Figure 2: Wavelength state transition diagram. switching fabric, the state changes from Reserved to Switched, implying that the physical configuration of this channel is finished and the wavelength is ready for use by this connection. When a Release signaling message is received, the wavelength state changes from Switched to Free, implying that now it can be used by other connections. Without the consideration of LMP, a fault on a channel can be detected only when the wavelength is in use. After a fault is detected on a wavelength, its state will become Down, implying that this channel is in an abnormal condition. Figure (b) shows the wavelength state transition diagram with the consideration of LMP. Under Verification is a new state that indicates the wavelength is under the LMP verification. LMP can detect the failure of wavelength channels when they are in Free or Switched state. The dashed arrow line in Figure 2(b) is added for case (2), which as mentioned above deals with the scenario that a PATH signaling message arrives to reserve a wavelength in Under Verification state. The wavelength state transition diagram for case (1) is obtained by removing the dashed arrow line. 4 Description of Different Signaling Schemes This section describes the signaling procedure for different signaling schemes and discusses tradeoffs between them. We use PATH and RESV to represent the downstream and upstream signaling messages respectively. Although the signaling messages of RSVP protocol also use these names, the reservation procedure and actions taken on each switching node for the signaling schemes discussed in this paper is different from those of RSVP (especially for parallel signaling scheme). 10

11 4.1 Hop-by-hop Signaling and Parallel Signaling In hop-by-hop signaling, when a lightpath establishment request arrives at a source node, the source node selects a route and available wavelengths along the route, according to the RWA algorithm and the global network resource information stored in the database. Then a PATH message is sent out along the selected route in a hop-by-hop manner to reserve the wavelength on each link and configure the OXC at each node. If there is no wavelength available on the outgoing link at one of the intermediate nodes, the connection is blocked and a Release message is sent back to the source node along the reverse route to release all the wavelengths reserved previously. If the PATH message reaches the destination node, that implies that the success of the reservation, a RESV message is sent back along the reverse route. When the source node receives the RESV message, it can start sending data over the wavelengths configured on each link of the route. When the data transmission is finished, the source node sends out a Release message to each node along the route, to release the wavelengths occupied by this connection. All of the above messages proceed in a hop-by-hop manner. Corresponding actions for wavelength reservation and OXC configuration will be performed at each intermediate node. 10 J 9 I 1 2 K G 3 A 8 C D PATH message H B F 6 E L M RESV message O N reservatuin time I J K A F O PATH 1 PATH 2 PATH 3 PATH 4 PATH 5 RESV 10 RESV 9 RESV 8 RESV 7 RESV 6 Data Transmission Release 1 Release 2 Release 3 Release 4 Release 5 Release ACK PATH message RESV message (a) Topology illustration (b) Timescale illustration Figure 3: Illustration of the hop-by-hop signaling scheme. In parallel signaling, when a lightpath establishment request arrives at a source node, the source node simultaneously sends PATH messages to each node involved in the connection. A RESV 11

12 message is sent back to the source node from each node after the success of the reservation and OXC switching. If the source node receives all the RESV messages from every node on the route, it can initiate the transfer of data. After the data transmission is done, the source node simultaneously sends out Release messages to all the nodes involved in the connection to release the network resources. If there is no wavelength available at any of the nodes during the reservation, a NACK message will be sent back to the source node, indicating the connection establishment is blocked due to lack of available wavelengths. While receiving the NACK message, the source node sends out Release messages to each node on the route to release the wavelength reserved previously. All the messages for signaling scheme are sent out in parallel. Examples of hop-by-hop signaling and parallel signaling are illustrated in Figure 3 and Figure 4 respectively, in which the signaling procedure is illustrated using both topology and timescale diagrams. The solid arrow lines represent PATH or Release messages and the dashed arrow lines represent RESV messages. J I G B H F O reservatuin time I J K A F O PATH PATH PATH RESV RESV RESV PATH RESV PATH RESV K A E N Data Transimission C M L Release Release Release Release Release D Rlease ACK Rlease ACK Rlease ACK Rlease ACK Rlease ACK PATH message RESV message PATH message RESV message (a) Topology illustration (b) Timescale illustration Figure 4: Illustration of the parallel signaling scheme. As discussed in Section 3.1, the control plane and the data plane for optical transport networks can be separate. Assume that an IP-based out-of-band signaling is used for the control plane. In hop-by-hop signaling, all signaling messages traverse the route selected by the RWA algorithm, while in parallel signaling, the signaling messages follow the route decided by the IP routing algorithm. These two routes may be different. For example, if the RWA algorithm employs the least-congested 12

13 routing algorithm and the IP routing algorithm uses shortest path first algorithm, the route decided by the IP routing algorithm may be shorter than the one decided by the RWA algorithm. In this case, one round-trip of hop-by-hop signaling may be subject to more nodal processing time and propagation time. As a result, the performance of these two schemes may be different. Tradeoffs exist between these two signaling schemes. Compared to hop-by-hop signaling, parallel signaling has the following advantages: Lower probability of resource contention. Resource contention is well known in a distributed environment. The signaling messages of several connections may be proceeding to reserve network resources needed by each of them. As a result, all of them may be blocked, even though there are sufficient resources for some of them [22]. By sending out signaling messages in parallel along the shortest path, parallel signaling can finish the reservation procedure in a shorter time. This reduces the probability of resource contention. Faster network resource release speed. In the lightpath teardown stage, Release signaling messages are also sent out in parallel to release network resources as quickly as possible. This fast release procedure enhances the network utilization, because the network resources can be used by other connections after being released. Nevertheless, this effect is only obvious when the network load is high and the connection service time is short as shown in our simulation results. However, the network performance improvement of parallel signaling caused by the above factors depends, to a great extent, on the traffic pattern and OXC architectures. Our simulation data show that the performance of parallel signaling may become undesirable in some situations (see Section 5.3.2). One drawback of parallel signaling is that it involves more overhead in terms of the number of signaling messages. For example, in order to establish a lightpath spanning five nodes (one source node, three intermediate nodes, and one destination node), four PATH messages will be sent 13

14 out for parallel signaling by the source node, while only one message is required for hop-by-hop signaling. Even worse, when the reservation fails, the source node needs to send Release messages to all nodes involved in the connection. Therefore, for parallel signaling, each node has to consume more computing resources and occupy more storage room to deal with the increased number of signaling messages (e.g., to keep extra states for each signaling message). In addition, two routing tables (one for the control plane and one for the data plane) need to be maintained at each node, resulting in an increased complexity of implementation. As a result, a scalability problem with parallel signaling seems to exist in large-scale networks. Nevertheless, the above drawbacks are not severe problems in optical networks. The bandwidth overhead arising from the increased number of the parallel signaling messages is negligible if a dedicated wavelength is used as the control channel. In addition, a signaling message may be smaller for parallel signaling than for hop-by-hop signaling, since the parallel signaling message need not to contain the address and reservation information of all the nodes on the lightpath. The scalability problem is also not as severe as imagined. Becasue the number of parallel signaling messages is linearly proportional to the length of a lightpath, which is n 1 in the worst case (n is the total number of the network nodes). A realistic concern about the parallel signaling scheme is that it is not compatible with the existing hop-by-hop GMPLS signaling protocols. This may complicate the implementation of parallel signaling, since the GMPLS protocols have received more interest and have been well studied for their implementation. However, for better performance it is possible for a carrier to choose a parallel signaling scheme in its local administrative domain, which may use network-to-network interface (NNI) to communicate with other domains that employ GMPLS signaling protocols. 4.2 Forward Before Switching Configuration and Forward After Switching Configuration Forward before switching configuration (FBSC) and forward after switching configuration (FASC) are two hop-by-hop signaling schemes. In [25], they are referred to as Forward Before XC and 14

15 Forward After XC respectively. In FBSC, when an intermediate node receives a PATH message, it reserves a wavelength on the outgoing link, inserts the switching request in the switching-requestwaiting queue, and forwards the PATH message to the next hop without waiting for the switching confirmation. At the destination node, a RESV message is sent out in the reverse direction to the source node after receiving the switching confirmation. At each intermediate node, the RESV message waits for the switching confirmation of the switching request submitted by the previous PATH message before it proceeds to the next node. The RESV message proceeds to the next node only after the wavelength status becomes Switched. In FASC, each node waits for the OXC switching confirmation, before it forwards the PATH message to the next node. So, FASC has a longer reservation time, that may potentially increase the chance of race conditions. Especially when OXCs with the sequential OXC architecture [25] are employed in a network, the time for FASC to wait for the switching confirmation at each intermediate node can be significant due to the increased waiting time in the switching-requestwaiting queue. Since the arrival of the PATH message of FASC at the destination implies the success of the reservation and configuration procedure, one of the improvements for FASC is to send RESV message from the destination to the source node directly following the shortest path, but not the reverse path of the lightpath. Then, two routing table needs to be maintained in this case. Figure 5 (a) and (b) show two examples of the signaling procedure for the lightpath establishment of FBSC and FASC respectively. Note that the signaling behavior of PATH and RESV signaling messages is different from the traditional RSVP protocol. To save reservation time, the resource reservation may begin during the forwarding of PATH messages, since the source node can have knowledge of the network resources (wavelengths) to be reserved. In [6], this is discussed as the suggested label. However, forwarding before or after OXC switching confirmation, which is not addressed in [6], remains an implementation choice. 15

16 I J K A F O I J K A F O reservatuin time reservatuin time PATH 1 PATH 2 PATH 3 PATH 4 RESV 10 RESV 9 RESV 8 RESV 7 PATH 5 RESV 6 PATH 1 PATH 2 PATH 3 PATH 4 PATH 5 RESV 10 RESV 9 RESV 8 RESV 7 RESV 6 PATH message RESV message OXC Switching Time PATH message RESV message OXC Switching Time (b) FBSC (b) FASC 5 Simulation Figure 5: Illustration of the hop-by-hop signaling using FBSC and FASC. We conducted a series of simulation experiments, using the discrete event driven SIMulator for Optical Networks (SIMON) [27], to study the performance of the signaling schemes discussed in the previous section in wavelength-routed mesh networks. In this section, we first present the metrics for the comparison and the parameters used in the simulation. Then we analyze the simulation results. 5.1 Metrics We use network blocking probability and reservation time as metrics of network performance. The network blocking probability is the ratio of the number of failed connections to the total number of connection requests. The reservation time of a connection, also referred to as response time in some studies [15], is the interval between the arrival time of a connection request and the starting time of data transmission. Factors that may affect the reservation time include: Nodal processing time: the time required for a node to process a signaling message ( e.g. the time needed for looking up the database and the routing table). Transmission time: the time needed for transmitting a signaling message. 16

17 Propagation time: the ratio of the length of a link to the speed of light. OXC switching time and OXC architecture: discussed in 3.1. For example, if FASC and parallel OXCs are used, the reservation time of the lightpath in Figure 3 can be computed as the round trip time of the signaling message: 10 propagation time+ 11 processing time + 10 transmission time + 6 switching time. For the parallel signaling scheme shown in Figure 4 and parallel OXCs, the reservation time is the maximum round trip time of the signaling messages (i.e., the round trip between node I and node O): 8 propagation time + 9 processing time+8 transmission time+switching time. For FBSC in the network equipped with parallel OXCs, and for all the signaling schemes in the network equipped with sequential or batch OXCs, the reservation time must be obtained using simulation. 5.2 Parameters for the simulation Following are the parameters set for our simulation: Network topology. The simulation is conducted on the Pacific Bell network topology (shown in Figure 3), which has 15 nodes, 21 bi-directional links with lengths of 100-km each, and 8 wavelengths in each direction. OXC switching time. We use 20ms as the OXC switching time, which fit well with today s popular MEMS switches, whose switching time is on the order of milliseconds. Furthermore, switching time of largescale switching fabrics is normally on the order of milliseconds [23]. Transmission time of signaling messages. The transmission time is set to be 0.02ms,which is the time needed to send a 256 Byte message over a 100Mbps IP switching fabric. The message size, 256 Bytes, is set to ensure that all the information required for any type of signaling message can be filled [6]. We assume all signaling messages have the same size. 17

18 Connection requests. The simulation offers 100,000 connection requests that are uniformly distributed over all source-destination pairs. Connection requests arrive following a Poisson distribution (the service time also follows the exponential distribution). Traffic pattern. To study the effect of signaling schemes under different traffic patterns, we simulate four types of traffic with average service time of 0.3 second, 3 seconds, 300 seconds and 1 hour. These numbers represent optical services of each magnitude in the time range from seconds to hours. For each of these traffic patterns, we change the average connection request inter-arrival time to obtain the network load of 10, 20, 30, 40, 50 and 60 Erlang. RWA algorithm. For hop-by-hop signaling, we use an adaptive routing algorithm, in which each link is assigned a weight of W and a shortest weighted path is selected. The weight W is defined as follows: W = { 1 NumW av if NumW av > 0 if NumW av = 0 where N umw av denotes the number of available wavelengths on a link. This algorithm dynamically chooses a path with available wavelengths. Since we assume that each OXC node has the capability of wavelength conversion, the wavelength assignment is implemented by picking up the first available wavelength at each node. This RWA algorithm facilitates our study of the signaling effect by trying to minimize the network blocking probability caused by RWA algorithm. For parallel signaling, the wavelength assignment algorithm remains the same as described above. But, we use the shortest path first as the routing algorithm where the weight of each link is set to be its length, to ensure that signaling messages traverse along the shortest distance between a pair of nodes. 18

19 5.3 Analysis of Simulation Results In this section, we analyze the simulation results. We present the data obtained under the sequential OXC as well as parallel OXC. Note that the results of the signaling schemes under the batch OXC, which are not presented in this section, are almost the same as those of signaling under the sequential OXC FBSC versus FASC Figure 6 plots the blocking probability versus network load for the signaling schemes under traffic with average service time of 0.3 second and 3 seconds. The results for FBSC, FASC, and parallel signaling under parallel and sequential OXCs are presented. Blocking Probability FBSC Signaling & Parallel OXC FBSC Signaling & Sequential OXC FASC Signaling & Parallel OXC FASC Signaling & Sequential OXC Parallel Signaling & Parallel OXC Parallel Signaling & Sequential OXC Blocking Probability FBSC Signaling & Parallel OXC FBSC Signaling & Sequential OXC FASC Signaling & Parallel OXC FASC Signaling & SequentialOXC Parallel Signaling & Parallel OXC Parallel Signaling & Sequential OXC Load(Erlang) (a) average service time = 0.3 second Load(Erlang) (b) average service time = 3 second Figure 6: Blocking probability versus network load in the Pacific Bell network with 8 wavelengths. In Figures 6(a) and (b), FBSC shows a better performance than FASC, under both the parallel and sequential OXC architectures. That is because FBSC saves a unit of switching time (waiting time in the switching-request-waiting queue plus physical switching time) for the PATH message at each intermediate node. Since race conditions [22] happen in the downstream direction from source to destination, the reduction of the reservation time reduces the chance of race conditions. Hence, FBSC has a lower blocking probability than FASC. The improvement of FBSC over FASC becomes more significant under a higher network load. Because for a fixed traffic pattern (average service time is fixed for a traffic pattern in this paper), a higher network load implies a shorter inter-arrival time, which may cause more wavelength competition. As a result, the performance improvement of FBSC over FASC becomes more obvious as the network load increases. 19

20 5.3.2 Hop-by-hop Signaling versus Parallel Signaling The data in Figures 6(a) and (b) also show that the parallel signaling under parallel OXCs has the best performance in terms of network blocking probability. For parallel signaling under sequential OXCs and traffic with average service time of 3 seconds, its blocking probability is lower than the other two hop-by-hop signaling schemes. However, under traffic with average service time of 0.3 second and the sequential OXC architecture, parallel signaling shows undesirable performance. In this setting, when the network load is below 30 Erlang, it still has a better performance than the other two hop-by-hop signaling methods. As the network load increases above 30 Erlang, it yields the worst performance with the blocking probability increasing dramatically. Three factors contribute to the sharp increase in the blocking probability: the smaller inter-arrival time under the higher network load, the accumulated waiting time in the switching-request-waiting queue of sequential OXCs, and the nature of the parallel signaling scheme Reservation Time Reservation Time(ms) FBSC Signaling & Parallel OXC FBSC Signaling & Sequential OXC FASC Signaling & Parallel OXC FASCSignaling & Sequential OXC Parallel Signaling & Parallel OXC Parallel Signaling & Sequential OXC Load(Erlang) (a) average Service time = 0.3 seconds Reservation Time(ms) FBSC Signaling & Parallel OXC FBSC Signaling & Sequential OXC FASC Signaling & Parallel OXC FASC Signaling & Sequential OXC Parallel Signaling & Parallel OXC Parallel Signaling & Sequential OXC Load(Erlang) (b) average service time = 3 seconds Figure 7: Reservation time versus network load in the Pacific Bell network with 8 wavelengths. Figure 7 presents the average reservation time of signaling schemes under traffic with average service time of 3 and 0.3 seconds. In general, the reservation times of FBSC and parallel signaling are smaller than those of FASC. FBSC yields a stable value of reservation time (around 28ms) for both traffic patterns. There is no significant difference in the reservation times between FBSC under parallel OXCs and FBSC under sequential OXCs. That is because the accumulated waiting time of a PATH message in the switching-request-waiting queue is ameliorated by forwarding the 20

21 PATH message before obtaining switching confirmation at each intermediate node. The FASC scheme yields an average reservation time above 100ms, which is at least 3 times longer than the reservation time of FBSC. The accumulated waiting time for switching confirmation is the primary contributor to the longer reservation time of FASC. Parallel signaling achieves an average reservation time usually close to that of FBSC. However, the reservation time of parallel signaling under sequential OXCs for traffic with the service time of 0.3 second is much greater than that of FBSC. The reservation time of parallel signaling under sequential OXCs for traffic with the average service time of 0.3s second has a sharp increase when the network load increases above 20 Erlang (This effect also happens to FASC when the network load increases above 30 Erlang in the same setting). This is because the decreased connection inter-arrival time causes more switching requests at the OXC controller, resulting in a longer waiting time in the switching-request-waiting for sequential OXC. Since the network resources are not ready for use during the reservation period, a long reservation time implies a low network utilization. For this reason, shorter reservation time is desirable for signaling in optical transport networks The Impact of Traffic Pattern on Blocking Probability Figure 8 presents the data of network blocking probability versus network load for FBSC, FASC, and parallel signaling under sequential OXC for four different traffic patterns (average service time of 0.3 seconds, 3 seconds, 300 seconds, and one hour). We notice from the figure that the impact of traffic patterns on the signaling performance is twofold: First, parallel signaling shows the worst performance under the traffic with average service time of 0.3 seconds. However, it has the best performance in other traffic patterns. We have explained the reason for this effect in Section Second, in terms of blocking probability, the effect of performance improvement of FBSC over FASC becomes more significant as the average service time decreases. For instance, under the traffic with the average service time of 0.3 seconds and the network load of 60 Erlang, the relative improvement of FBSC over FASC with Sequential OXC is 70.3%. As the average service time increases to 3 seconds, 300 seconds and one hour, it decreases to 11.5%, 4.7% and 0% respectively. 21

22 Blocking Probability (%) Blocking Probability (%) FBSC Signaling & Sequential OXC FASC Signaling & Sequential OXC Parallel Signaling & Sequential OXC Load(Erlang) (a) average Service time = 0.3 seconds FBSC Signaling & Sequential OXC FASC Signaling & Sequential OXC Parallel Signaling & Sequential OXC Load(Erlang) (c) average Service time = 300 seconds Blocking Probability (%) Blocking Probability (%) FBSC Signaling & Sequential OXC FASC Signaling & Sequential OXC Parallel Signaling & Sequential OXC Load(Erlang) (b)average Service time = 3 seconds FBSC Signaling & Sequential OXC FASC Signaling & Sequential OXC Parallel Signaling & Sequential OXC Load(Erlang) (d) average Service time = 1 hour Figure 8: Blocking probability versus network load under different traffic patterns. Difference of P service time = 0.3s service time =3s service time =300s service time = one hour Load(Erlang) Relative improvement x100 (%) service time = 0.3s service time =3s service time =300s service time = one hour Load(Erlang) Figure 9: P between FBSC and FASC. Figure 10: Relative improvement of FBSC over FASC in terms of blocking probability. We define the number of arriving connection requests per reservation, denoted by P, as the ratio of the average reservation time to the average connection inter-arrival time. P represents the average number of connection requests, that arrive at the network during the period of each reservation. Intuitively, the performance improvement between two signaling schemes,in terms of blocking probability, is reflected by the difference in P. Let P be the difference of P between two signaling schemes. In a network setting with a greater P, the performance difference between two signaling schemes becomes more obvious. Figure 9 shows the P of FBSC and FASC under 22

23 sequential OXC for four different traffic patterns and Figure 10 shows the relative improvement of FBSC over FASC in the same setting. The larger P in Figure 9 reflects a more significant network performance improvement of FBSC over FASC in Figure LMP Blocking Probability (%) Non LMP & Service Time = 3 s LMP & Service Time = 3 s Non LMP & Service Time = 0.3 s LMP & Service Time = 0.3 s Load(Erlang) Figure 11: Blocking probability versus network load for signaling with and without LMP. As discussed in Section 3.2, the introduction of LMP connectivity verification may increase the total reservation time, if a signaling message waits for the end of verification on the way down to the destination. We compare the FBSC scheme waiting for LMP verification (LMP in Figure 11) and the FBSC interrupting LMP verification (Non-LMP in Figure 11). The latter one yields a lower blocking probability as shown in Figure 11. However, the difference between them is small for the traffic with the average service time greater than 3 seconds. The impact of traffic patterns on blocking probability remains the same in this case. 6 Conclusion In this paper we studied signaling schemes for distributed connection management in wavelengthrouted optical mesh networks. From an implementation perspective, we investigated extensive factors (including the OXC architecture and switching time, operating sequence between the OXC switching configuration and signaling message processing, traffic pattern, and LMP) that affect the performance of these signaling schemes. We discussed why the performance of these signaling schemes differs under various network settings. Our simulation results show that, in general, the 23

24 hop-by-hop signaling using FBSC outperforms the hop-by-hop signaling using FASC in terms of blocking probability and reservation time. The simulation results also show that the performance of a signaling scheme may vary with changes in the network setting, such as the OXC architecture and traffic patterns. For the first time, a complete parallel signaling scheme is simulated in our study. It shows a better performance than the hop-by-hop signaling schemes in some situations. Given a network configured with parallel OXCs or having traffic with relatively long service time, the parallel signaling scheme can chieve better performance. However, the parallel signaling scheme has undesirable performance when sequential OXCs are used and the average connection service time decreases to a certain value. Considering both network scalability and performance, FBSC becomes a good choice, since the performance of FBSC is not affected by the network setting as significantly as FASC and parallel signaling. Although hop-by-hop signaling is commercially more mature, using parallel signaling still has benefits in some situations as we discussed. Tradeoffs exist among different signaling schemes. The choice of a signaling scheme depends on the actual design of a network. Due to the existence of diverse settings in real-world networks, individual carriers need to choose signaling schemes based on the knowledge of their own networks. References [1] B. Davie and Y. Rekhter, MPLS Technology and Applications, Academic Press, May [2] H. Zang, J. Jue and B. Mukherjee, A review of routing and wavelength assignment approaches for wavelength routed optical WDM networks, Optical Networks Magazine, vol 1, no. 1, January [3] N. M. Bhide, K. M. Sivalingam and T. Fabry-Asztalos, Routing mechanisms employing adaptive weight functions for shortest path rounting in optical WDM networks, Journal of Photonic Network Communications, July [4] A. Banerjee, et al., Generalized Multiprotocol Label Switching: an overview of routing and management enhancements, IEEE Communications Magazine, January [5] A. Banerjee, et al., Generalized Multiprotocol Label Switching: an overview of signaling enhancements and recovery techniques, IEEE Communications magazine, July [6] L. Berger, Generalized Multi-Protocol Label Switching (GMPLS) - signaling functional description, IETF RFC 3471, January