FLOW PRE-EMPTIBLE WAVELENGTH ASSIGNMENT PROTOCOL FOR WDM NETWORKS S.Indira Gandhi, V.Vaidehi, T.Subash Department of Electronics Engineering Madras Institute of Technology, Anna University, Chennai 600044, INDIA ABSTRACT Existing Internet infrastructure is moving toward the Wavelength Division Multiplexing (WDM) optical network, where the wavelength converters are playing major role. The coupling of IP routers with wavelength-selective optical cross-connects supports the optical WDM networking. Because optical wavelength routing is transparent to IP, packets can bypass traditional forwarding and can be routed directly through the optical cross-connects. This results in very high throughput and low delay. This approach shares features with label switching, but wavelengths are a much more scarce resource than labels. Optical switches take larger time to switch than electronic switches, and wavelength conversions are expensive. So wavelength “label” swapping is not easily done. In an optical environment the Wavelength “label” assignments must consider the practical limitations of these devices. This paper describes an optical wavelength (label) switching signaling protocol, FP-WAP, which enables the POW architecture and is designed to minimize wavelength swappings, utilize wavelengths with flow merging, Pre-empt smaller flows with larger flows, reduce the reconfiguration of optical switches and takes into account the optical device limitations. Keywords: Flow Pre-Emptible Wavelength Assignment Protocol (FP-WAP), Packets over Wavelength (POW), Optical Cross Connects (OXCs)
1. INTRODUCTION Existing Internet infrastructure is moving toward the WDM optical network, where the wavelength converters [7] are playing major role. The coupling IP routers with wavelength-selective optical cross-connects support the optical network. Because optical wavelength routing is transparent to IP, packets can bypass traditional forwarding and route directly through the OXCs. This results in very high throughput and low delay. Current label switching mechanisms assume a large label space, where label swapping is inexpensive (done in hardware at the ATM layer) [1]. The replacement of labels with colors (wavelengths) in a WDM network raises several challenges. Current WDM technology is limited to a few (eight to 64) wavelengths per link, which is very small compared to the number of network connections in the Internet today. Wavelength converters (corollary to label swapping) are expensive [2]. Finally, practical active (data-dependent) optical switching elements are slow, and thus costly to reconfigure. These limitations warrant re-examination of IP-WDM switching approaches. Our work is based on a signaling protocol created to dynamically configure the lightpath for flows. This protocol is called the Flow Pre-emptible Wavelength Assignment Protocol (FP-WAP). This paper analyzes the performance of FPWAP, through the simulation. The POW architecture [3] is a performance instance of this approach [FP-WAP].
2. HIGH-LEVEL REQUIREMENTS There are several high-level requirements for the POW signaling protocol. First, it must be kept as simple and as light as possible. Second, the continuous light path should be constructed as far as possible. Non-continuous light paths should be limited, which require wavelength conversion. Third, flow merging (grooming) must be supported. FP-WAP is implemented on top of a reliable transport layer, for its reliable transmission. This confirms the first requirement the simplicity. This is common for signaling protocol [3]. The signals are sent over the per-neighbor TCP connections, which is unique in FP-WAP, constructed by FP-WAP components. The per-neighbor TCP connections simplify the signaling, failure detection, and recovery. The switches of each link maintain the neighbor relationships (one connection per link). The connections are trivial, not optimized, and similar to any of the common routing protocol [3]. Constructing the contiguous light path as far as possible requires FP-WAP to pick a common free wavelength along the flow path; therefore FP-WAP must collect the list of
1
free wavelengths for each hop. If there is one free wavelength common to all the hops, it will be picked. If not, FP-WAP may choose to construct a non-contiguous light path if there are sufficient wavelength converters [4] available. This feature is unique to FP-WAP compared to label swapping techniques, because FP-WAP’s decision to pick a “label” (wavelength) is a global decision. FP-WAP tries to minimize or eliminate swapping of “labels”. Therefore, FP-WAP incurs a round-trip time to select a wavelength. The resource must be also locked during signaling.
3. FP-WAP SIGNALING SCHEMES AND PROTOCOL DESIGN FP-WAP decides where to initiate the signaling. Either end of the path is appropriate, as they natural places where FP-WAP can efficiently gather complete path information. There are two signaling schemes namely First Hop initiated signaling [3]. Last Hop initiated signaling. 3.1 LAST-HOP INITIATED SIGNALING For a number of reasons, notably in the presence of grooming (merging), it is better to initiate the signaling from the last hop. In merging, there is a single last hop, but multiple first hops, which would complicate a source-initiated protocol. The last hop will also notice the flow earlier, as the traffic merges there. Second, it will simplify the protocol because there could be more than one outstanding setup request from upstream and the protocol must keep track of the upstream status so that it can selectively send the COMMITs back to it. The first sequence is similar to the first-hop initiated. Unless the next hop has already setup the switch the previous hop should not start sending the packets using the new wavelength (to avoid losses). Similar to IP Switching FP-WAP can do something, that the switch will send the packets using the slow path (through the IP router) while waiting for response from the next hop. However, because this technique requires temporary path termination and optical switches take a long time to setup, it is undesirable to do so. Instead in FP-WAP, the first hop will wait one round-trip time for signaling propagation to complete. The process is shown in Fig. 1. 3.2 FLOW AGGREGATION Flow aggregation means the aggregation of traffic flow in one particular node. Flow aggregation (grooming) also affects where to initiate the teardown mechanism. The drops in an aggregated flow are noticed at the sources first so the last hop is undesirable. Merging hops is also not desired because it requires the hop to monitor the optical signal. Therefore, it is the responsibility for the first hops to initiate the teardown. If the first hop switches of a switched flow detect a drop in the throughput, it will send TEARDOWN to the next hop and the next hop will pass it to further hops if there are no switched incoming branches. Fig. 2 illustrates an instance of Flow Aggregation. Suppose FP-WAP was set to regard 25 pps (packets per second) as a threshold to switch a flow. Switch D, E, F, G, and H all see an aggregate outgoing throughput of 25 pps or higher for flow F1 (* --> Domain G, i.e., traffic going to G) and F2 (* à Domain H, i.e., traffic going to H). Now we consider flow F1 i.e., traffic going to G for the detailed explanation.
Fig.1. Last Hop Initiated Signaling Scheme
When switch G detects the throughput of 25 pps or higher, and switch G knows it is the last hop, it locks the free wavelength resource ? FG and sends SETUP (F1 , ? FG ) to F. Upon receiving that SETUP, F does the wavelength set intersection ? x = ? FG n ? EF = {? 1 , ? 3 }, lock ? EF and sends SETUP (F1 , ? x) to E. E does the same intersection ? y = ? x n ? DE = {? 3 } and ? z = ? x n ? CE = {? 1 }. Then it will send SETUP (F1 , ? y ) to D and SETUP (F1 , ? z) to C. Both C and D know that they are the first hops for that path and C will send COMMIT (F1 , ? 3 ) and D will send COMMIT (F1 , ? 1 ). Meanwhile, E waits for responses from both. Upon receiving the responses, E picks ? 3 because it is an element of ? DE and
2
D contributed more to the aggregate throughput than C. E removes ? 3 from the set ? DE and ? 1 from the set ? CE , and sends COMMIT (F1 , ? 3 ) to F. F removes ? 3 from the set ? EF, unlocks it and forwards COMMIT (F1 , ? 3 ) to G. Then, G removes ? 3 from the set ? FG , unlocks it, configuring its optical switch and flow converter to convert ? 3 back to the default wavelength ? 0 , and send COMMIT_OK (F1 ) to F. Upon receiving COMMIT_OK (F1 ) from G, F sends COMMIT_OK (F1 ) to E, and switch E starts sending the flow through a wavelength converter that converts default wavelength ? 0 to ? 3 .
Fig.2. Flow Aggregation Scheme
Branch D is selected because the more a branch contributes to the aggregate throughput, the more likely it is expected to stay significant or become even more significant. If other branches become inactive, they will likely become a sub-path. The wavelengths of low flow branches are picked arbitrarily and merged to the target wavelength Now suppose there is an increase in the F1 throughput from C to E. E realizes that the flow has been switched, so it sends SETUP (F1 , ? CE ) to C. C determines that it is the first hop and sends COMMIT (F1 , ? 3 ). Upon receiving from C, E configures its optical switch and sends COMMIT_OK (F1 ) to C, and C sends the flow using ? 3 when it receives the message. Table 1 Free Wavelength at Fig.3 Switches SWITCH
LINK
FREE WAVELENGTHS SET
H G F E
FH FG EF DE CE
?FH={?1,?2 ,?3 } ?FG ={?1,?2 ,?3 } ?EF={?1 ,?3 } ?DE={?1 ,?3 } ?CE={?2 ,?3 }
If, subsequently, D detects that F1 throughput drop, it will send TEARDOWN (F1 ) to E. Switch E will not forward the message further because it still has a switched branch, i.e.: CE. As a result, E only frees the wavelength, removes the switched-path from its optical switch to its IP router, and its IP router will merge the incoming flow to the outgoing switched flow. Finally, FP-WAP requires that neighbor protocol emit periodic keep-alive messages so the POW switch will detect neighbor failures. The keep-alive message should incorporate a mechanism to detect the case of neighbor failure and subsequent recovery and up again prior to the timeout of its neighbor entry. This is handled by the routing protocol. If the failed neighbor is the previous hop, then the switch will do the same thing as if it received TEARDOWN (F1 ), TEARDOWN (F2 ), …, TEARDOWN (Fn ) where F1-n are the switched flows coming from the neighbor. If the failed neighbor is the next hop and there is no previous hop switched, then the switch just sends the flow using the default wavelength ? 0 . However, if there is a switched previous hop, the switch will take the last hop role. 3.3 FLOW PRE-EMPTION When compared to SWAP, this approach deals with the pre-emption of low traffic flow by larger traffic flows, which rise up the percentage packets switched and also increase efficiency. Once a SETUP_FAIL message is received, FP-WAP tries for pre-empting a flow in order to assign a wavelength for that flow. That is, a FP-SETUP (F, ?) is sent to the upstream neighbor to get a wavelength, which is to be pre-empted. Each node identifies the wavelength ? with the low traffic flow and augments it with the collected wavelength set. The traffic-initializing node then determines the wavelength ? from the collected set (wavelength that occur most number of times in the set). Upon selecting ? FP-COMMIT (F, ?) is sent back to the initializing node. Once COMMIT_OK is received the switches are re-configured accordingly.
3
4. FP-WAP MESSAGES AND PARAMETERS There are eight messages related to wavelength assignment (connection setup and teardown): SETUP (F, ?l ): Ask the previous hop to pick a wavelength to setup the Connection by sending the free wavelength set for the sub-path (also lock it). The recipients intersect this set with their own free wavelength set, and forward the result to the subsequent hop when they detect a high incoming throughput for a flow. SETUP_CONFLICT (F): Inform the next hop that others have locked one of the wavelength sets along the path. Intermediate hops will pass this message to the next hop. The initiating hops will perform a “back off” procedure. SETUP_FAIL (F): Inform the next hop that the connection cannot be made, as there are no free resources. Intermediate hops will pass this message to the next hop if they are not the initiating hop. COMMIT (F, λ): In form the next hop to use ? ?as the incoming wavelength. Intermediate hops will pass this message to the next hop if they are not the initiating hop. COMMIT_OK (F): Inform the previous hop that the optical switch has been setup. Intermediate hops will pass this message to the previous hop. This message will not be forwarded further if there are no branches waiting for it. TEARDOWN (F): Inform the next hop to tear down the connection. Intermediate hops will pass this message to the next hop. FP-SETUP (F,λ): Ask the previous hop to pick a wavelength to setup the Connection by picking up an allocated wavelength with low packet flow. The recipients augment this set with their own allocated wavelength with low packet flow, and forward the result to the subsequent hop when they detect a high incoming throughput for a flow. FP-COMMIT (F, λ): Once FP-SETUP is received select a wavelength ? which has large number of occurrence. In form the next hop to use ? ?as the incoming wavelength. Intermediate hops will pass this message to the next hop if they are not the initiating hop. Each FP-WAP entity maintains two types of states for each flow entry: incoming and outgoing states. Both the incoming and outgoing states are detailed in the next chapter. 4.1 INCOMING STATE TRANSITIONS The state of a traffic flow start as a new incoming flow F is received on a branch B (an interface of the switch). For a flowbranch tuple (F, B), a FP-WAP process for that particular flow, S, will perform different roles based on the following criteria: IC1: The output state of F has already been switched. IC2: S has a downstream neighbor in the direction of F. In the case of IC1 is true, then S will actively monitor the throughput of the flow (path I2 in Fig. 5). If IC1 is not true, IC2 will be used to make the decision. If IC2 is true, indicating S is not the egress node, and then S creates a state for F, passively monitors F, and waits for further signaling messages (path I2 in Fig. 5). However, if IC2 is not true, S is the egress node therefore it will actively monitor the flow F. Both ‘active S’ and ‘passive S’ need to monitor the flow so that the state for F eventually dies with the flow. A timer FlowDetectionTimer is used to facilitate this. The incoming state diagram is shown below in Fig. 3. Each node in a particular flow stays in any one of the states represented below.
Fig.3. Incoming State Transition Diagram 4.2 OUTGOING STATE TRANSITIONS
Once ‘active S’ (SA ) determines that the throughput of F1 on branch B is high, (F1 , B) becomes a switchable branch, and S will lock ? B and send SETUP (F1 , ? B) to its upstream neighbor, SP . If another on going signaling for flow F2 locked ?B, SA
4
will do a back off procedure similar to that of CSMA/CD (path I13 and I14 on Fig. 4). Once SA completes its backoff, it will retry the SETUP procedure. If after BackoffLimit tries, SA still cannot get the lock, it will give up the opportunity to switch F1 and monitor it for another detection window. The upstream neighbor SP (passive role) upon receiving SETUP (F, ?in) will use the following criteria: IC3: There is at least one switchable (F, B). IC4: Full lightpath construction is feasible, which means wavelength continuity is achievable or wavelength conversion is allowed and possible. IC5: Partial lightpath construction is allowed and feasible. If IC3 and IC4 are true, then for every switchable (F, Bn ), n=0, 1, 2… N, SP will spawn a task SPn , which locks ? n , assign ? = ? in ∩ ? n and send SETUP (F, λ) to the upstream neighbor. Here the outgoing state of F becomes WAIT_ALL_COMMIT (Fig. 4). Alternately, if IC3 and IC5 are true, SP will pick λουτ ∈ ? n and send COMMIT (F, λουτ) and upon receiving COMMIT_OK (F), it will take the active role (transition O5 and O6 in Fig. 6, transition I4 in Fig. 3). On the second phase of signaling, where SP is in the WAIT_ALL_COMMIT state, it will collect all COMMIT (F, ?) received on any switchable branch Bn, n=0,1,2…N and pick one wavelength ? x, x=0,1,2…N such that Bx is the branch with the highest throughput. SP then will send COMMIT (F, ? x) downstream. Meanwhile, SP also records the wavelength of preference of each branch. This COMMIT (F, ? x) will eventually be received by SA , the egress node. SA will configure the switch immediately, rather than (as SP ) waiting for all COMMIT messages to arrive before acting. The SA switch thus converts ? x to ? 0 (the slow path wavelength), terminating the lightpath. If the criteria IC3, IC4, and IC5 cannot be concurrently met, SPn will send a SETUP_FAIL (F) message downstream. The downstream nodes will not propagate this message if there are other signaling tasks persisting for a switchable branch Bm. However, if IC3, IC4, and IC5 can all be met, but the wavelength sets for the branch were locked, SPn will send SETUP_CONFLICT (F) toward SA so that SA can commence its back off procedure. Setup conflict messages must be propagated along the entire signaling path because every FP-WAP node Sn along the path has locked the wavelength set for the branch. By releasing those locks, FP-WAP creates opportunities to setup paths belonging to other flows.
Fig.4. Outgoing State Transition Diagram
The last part of FP-WAP signaling handles the teardown mechanism (transitions O7, O8, and O9 in Fig. 6). If an active node S detects that the throughput of a switched branch (F, B) drops below LowThreshold for a specific detection duration FlowActiveTimer, S will monitor that flow or another Flow Active Timer window (transition O7). If after the second window, the throughput remains below LowThreshold, S will send a TEARDOWN (F) message upstream (transition O8). S then becomes active or passive based on its position in flow tree. Fig. 4 represents the outgoing state diagram. Each node in a particular flow stays in any one of the states represented below. These states are determined either by incoming states or by the signals or by the parameters. Every transition is explained in detail below.
5. MODELLING AND SIMULATION FP-WAP is designed to enable the evaluation of the performance of the POW architecture. To ease FP-WAP development and verification and to evaluate POW topology FP-WAP is simulated with virtual socket programming in C under LINUX [5].
5
5.1 ESSENTIAL NODES There are three types of nodes in designing the network architecture for FP-WAP. They are i.Ingress node ii.Switches or OXCs iii.Egress node They are shown below in Fig. 5.
Fig.5. NETWORK MODEL
A traffic generator is designed which defines the traffic from the ingress nodes to the egress nodes. Each types of node is elaborated in the upcoming pages with their flow diagrams. Number of wavelengths that are going to be used is given in as input to the traffic generator. All the above-mentioned nodes have a Flow Analyzer and a link table, which helps in configuring switches and in routing packets. INGRESS NODE: This is the node, which would be present at the edge of the POW domain. They receive ip packets. These nodes do the encapsulation process (i.e.) they add the FP-WAP header to the incoming ip packets. These nodes actively monitor a flow and they originate signals namely COMMIT, SETUP, CONFLICT, SETUP_FAIL, TEARDOWN, FPSETUP and FP-COMMIT. SWITCHES: They switch the incoming packets with a particular wavelength to its destined link with the same or any other wavelengths. They usually have the parameters, PartialPathEnable and WavelengthConverterEnable. These two parameters decide the setting of partial light path. FP-WAP protocol minimizes the use of wavelength converters, as it is an expensive process. They usually generate signals namely, COMMIT, SETUP_FAIL, SETUP_CONFLICT, TEARDOWN, FP-SETUP and FP-COMMIT. EGRESS NODE: These nodes are another type of nodes, which are at the edges of the POW domain. They do the process called decapsulation, which removes the FP-WAP header from the ip-packet. These nodes usually monitor a particular flow in an active state. It checks for the packet count with High Threshold and sends the SETUP signal to its upstream neighbor. When it receives a SETUP_CONFLICT message it does the back-off procedure and once the BackoffTimer expire it sends back SETUP again. These nodes usually originate SETUP and COMMIT_OK messages. Flow Analyzer is a procedure found in all the nodes. It analyses the flow and updates all the FP-WAP parameters. They have the information about the total packets in a flow, wavelength converter etc. Fig. 6 is a plot between the number of wavelengths and the percentage of packets switched. Simulation traces show that there was a large amount of unsuccessful signaling due to the high number of switchable branches.
Fig.6. % of packets switched per link From Fig. 6, we find that the number of packets switched increases with the use of Flow Pre-Emption. Fig. 7 shows another plot between the no. of wavelengths and the percentage of the packets switched. This was plotted by having the parameters High Threshold = 20 packets / 5 secs. Here again we see a hike in percentage of packets being switched. Thus by varying the parameter HighThreshold,, the number of packets that will be switched can be varied. Table 3 gives the number of packets switched with HighThreshold=20 pkts/5secs.
6
Both High Threshold and Low Threshold should be optimally chosen to get high efficiency. Again with wavelength converters, the percentage packets switched goes higher but at the cost of conversion of wavelength (an expensive process) [2].
Fig.7. Plot 2 % of packets switched per link
CONCLUSION FP-WAP’s wavelength assignment is similar to IP switching; it is data-driven, rather than control-driven. The FP-WAP protocol is an approach to assigning wavelengths dynamically based on traffic demand, and was created to facilitate the Packet over Wavelength (POW) architecture. A notable difference is that POW introduces flow merging to reduce the number of wavelengths required. This requires FP-WAP nodes to perform both active and passive roles based on their positions in the flow tree. These roles must be performed at the point where electrical signals are still available: the first hop or last hop of the lightpaths. During FP-WAP simulation, the percentage of packets switched per link was measured. Fig. 6 & 7 indicates the reduction in the wavelength conversion. In order to avoid wasted signaling, FP-WAP augments SWAP to support light path preemption, where larger flows cause smaller flows to be torn down and wavelengths reassigned. Preemption of smaller flows by larger flows causes more percentage of packets to be switched Further study is required to measure FP-WAP performance under higher traffic loads. In addition, further investigation is required to determine the FP-WAP parameters affect its steady state and transient performance.
REFERENCES 1. 2. 3. 4. 5. 6. 7.
S. Keshav, “An Engineering Approach to Computer Networking: ATM Networks, The Internet and Telephone Network,” Reading MA, 11997. J. B. Yoo, “wavelength Conversion Technologies for WDM Network Applications,” IEEE Journal of Light wave Technology, Vol 14 no. 6, June 1996. Stephen Suryaputra, Joseph D. Touch, “Simple WavelengthAssignmentProtocol,”USC/Information Science Institute, October 26, 1996. K. C. Lee and V. O. K. Li, “A Wavelength-convertible Optical network,” Journal Of Light Wave Technology, Vol 11 May 1993. S. Lin and N. McKeown, “A Simulation Study of IP Switching,” proceedings of ACM SIGCOMM 1997. R. Ramaswami and K.N. Srirajan, “Optical Networks: A Practical Perspective, Morgan Kaufmann, San Francisco CA, 1998. Li-Wei chen, Eytan Modiano,” Efficient routing and wavelength assignment for reconfigurable WDM network with wavelength converters”, IEEE INFOCOM, 2003.
*
[email protected]; phone 91 44 22231963; fax 91 44 22232403; mitindia.edu
7
