Abstract
Coupling IP routers
with wavelength-selective optical cross-connects supports existing Internet
infrastructure in a wavelength division multiplexing (WDM) optical network.
Because optical wavelength routing is transparent to IP, packets can bypass
traditional forwarding and route directly through the optical cross-connect,
resulting in very high throughput and low delay. This approach shares features
with label switching, but wavelengths are a much more scarce resource than
labels. Because optical switches have
larger switching times than electronic switches, and wavelength conversion are
expensive, wavelength “label” swapping is not easily done. Wavelength “label”
assignments must consider these limitations to be practical in an optical
environment. This paper describes an optical label switching signaling
protocol, SWAP, which takes into account the optical device limitations, and is
designed to minimize wavelength swappings, utilize wavelengths with merging of
flows and reduce the reconfiguration of optical switches.
Introduction
Coupling Internet protocol (IP) routers with wavelength-selective optical cross-connects enables existing Internet infrastructure to operate in a wavelength division multiplexing (WDM) optical network. Because optical wavelength routing is transparent to IP, very high throughput and low delay can be achived when packets bypass the IP forwarding process by being routed directly through the optical cross-connect. While this approach is similar to label switching in general, there are several issues limiting its feasibility. Current label switching mechanisms assume a large label space, where label swapping is inexpensive (done in hardware at the ATM layer). Replacing 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 fine-grain network connections in the Internet today [NLANR]. Wavelength converters (corrolary to label swapping) are expensive [Yoo96]. Finally, practical active (data-dependent) optical switching elements are slow, and thus costly to reconfigure. Of the four types of optical switching elements: mechanical, electro-optical, thermo-optical, and SOA-based gate switched, only electro-optical has low switching time, but it suffers from high crosstalk and loss [RaSi98]. Thus reconfiguration of a practical optical wavelength cross-connect is generally slow.
These limitations warrant re-examination of IP-WDM switching approaches. The performance of an instance of this approach, called Packet over Wavelengths (POW) has been simulated and studied [BTWS99]. This study is based on a signaling protocol created to dynamically configure the lightpath for flows. This protocol is called the Simple Wavelength Assignment Protocol (SWAP). SWAP modeling and simulation is performed on the VINT/ns simulator [NS].
Prior and Related Work
POW forwarding is similar to label switching, i.e., they both replace complicated IP forwarding processing with a comparatively simple label lookup. IP processing relies on longest-prefix match in a large routing table, followed by updates of portions of the IP header, such as hopcount (TTL) and checksum. Label processing indexes the label in a small, fixed table, and performes a fixed label substitution. In POW, that lookup is done in the physical layer instead of in the link or network layer by the wavelength of the packet signal. This section includes further discussion of label switching approaches because of this similarity.
There are at least five known approaches to label switching: Toshiba Cell Switch Router (CSR), Ipsilon IP Switching, Cisco Tag Switching, IBM Aggregate Route-Based IP Switching (ARIS), and Multiprotocol Label Switching [DaDR98]. Among these, most notable are Tag Switching and IP Switching. Although MPLS is the IETF standard for label switching, its principal operations are significantly influenced by Tag Switching; this discussion thus regards these two approaches as the same. The primary difference between Tag Switching and IP Switching is the process of assigning labels to IP flows. Tag switching assigns flows based on routing protocol events; IP switching is based on the actual traffic seen on the network. The taxonomy described in [DaDR98] thus considers Tag Switching as control-driven and IP switching as data-driven.
POW avoids routing protocol interactions, making it more responsive to traffic trends. As a result, it is based on a data-driven approach similar to that of IP switching. Although IP switching has not performed well in real systems due to an overabundence of switchable flows, POW takes further steps, including coarsely aggregating flows to and merging flows, to reduce the number of flows handled by each switch. By merging, POW localizes the traffic dynamics to the nodes near the edges of the network.
Design
The POW Switch Architecture consists of four major components (Figure 1):
· Control Processor: an IP router equipped with gigabit data interface(s) and a control interface. Depending on the scalability-complexity trade-off, multiple data interfaces of the control processor could be implemented as a single electronic interface, where the optical path is terminated early but the optical interface ID must be stored in the form of internal headers (increased complexity).
· Optical Fabric: an interconnection network of optical switching elements. The elements are capable of switching optical signals from one input port to one of a set of output ports, where the output port is selected based on the wavelength of the signal at the input port.
· Wavelength Converters: these devices convert input signals with wavelength lin to output signals with wavelength lout.
· Wavelength MUX/DMUX: redirect a specific wavelength to a particular fabric, and merge the output of the fabrics to the external fibers. These are special cases of statically configured optical switching elements.
Although the details of the components are beyond the scope of this paper, the arrangement of the components follows share-with-local wavelength-convertible switch architecture [LeLi93] to reduce the number of wavelength converters per node.
Figure 1 - POW Switch Architecture
There are several high-level requirements for the POW signaling protocol. First, it must be kept as simple and as light (using few, brief messages) as possible. Second, the signaling should construct the continuous light path as far as possible. This second requirement is driven by the hardware limitation. Non-continuous light paths require wavelength conversion, which requires expensive hardware (this its use should be limited). Third, the protocol must support flow merging (grooming). These high-level requirements are the motivation for the following design decisions for Simple Wavelength Assignment Protocol (SWAP) signaling protocol[1].
To conform with the first requirement, i.e.: simplicity, SWAP is implemented on top of a reliable transport layer; this decouples the protocol from its reliable transmission. This is common practice for signaling protocols, e.g. the ATM signaling protocol is implemented on top of Service-Specific Connection-Oriented Protocol (SSCOP) which is a reliable transport above AAL5 [Kesh97]. The telephone Signaling System 7 (SS7) makes the same choice. SWAP components establishes per-neighbor TCP connnections, over which the signals are sent. Neighbor connections enable the switch to determine whether it is the first, the last or the intermediate hop for a particular flow, which is used to simplify the signaling, failure detection, and recovery. Neighbor relationships are maintained per link of the switches (one connection per link). The neighbor connections are not optimized because they are trivial and similar to any of the common routing protocols (IS-IS, OSPF, or BGP) [Perl92].
Figure 2 - Time sequence diagram of first-hop-initiated signaling
Constructing the contiguous light path as far as possible requires SWAP to pick a common free wavelength along the flow path, therefore SWAP must collect the list of free wavelengths for each hop. If there is one free wavelength common to all the hops, it will be picked. If not, SWAP may choose to construct a non-contiguous light path (if so configured) if there are sufficient wavelength converters available. This feature is unique to SWAP compared to label swapping techniques, because SWAP’s decision to pick a “label” (wavelength here functions as label) is a global decision. SWAP tries to minimize or eliminate swapping of “labels”. Therefore, SWAP incurs a round-trip time to select an wavelength. The resource must be also locked during signaling.
Next SWAP decides where to inititate the signalling. Either end of the path is appropriate, as they natural places where SWAP can efficiently gather complete path information. The first hop is good because a source SWAP can propagate its free wavelength set when it detects an active flow (SETUP). When the next hop receives that set, it intersects that set with its own free wavelength set, and forwards the result to the next hop. Assuming the final set is not empty, the last hop picks one free wavelength from the resulting set, configures its local switch, and send an acknowledgement (COMMIT) back to the previous hop with the chosen wavelength. Upon receiving an acknowledgement, the previous hop configures its local switch and passes the acknowledgement to its previous hop, until the packet is received by the first hop. The process is illustrated in Figure 2[2].
However, it is better to initiate the signaling from the last hop, for a number of reasons, notably in the presence of grooming (merging). In merging, there is a single last hop, but multiple first hops, which would complicate a source-initated 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. Third, last-hop-initiated signaling will ease the interaction with RSVP [ZDE+93] in the future. ARIS takes the same approach as it allows route aggregation.
The drawback of last hop initiated signaling is that it will take more time to complete the signaling. The first sequence is similar to the first-hop initiated, however the previous hop should not start sending the packets using the new wavelength unless the next hop has already setup the switch (to avoid losses). SWAP could do something similar to IP Switching, 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 SWAP, the first hop will wait one round-time for signaling propagation to complete. The process is shown in Figure 3[3].
Figure 3
– Time sequence diagram of last-hop-initiated signaling
Flow aggregation (grooming) also affects where to initiate the teardown mechanism. The last hop is undesirable, because drops in an aggregated flow are noticed at the sources first. 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. Figure 4 illustrates this effect on an aggregated flow.
Figure 4 – Flow Aggregation Scheme
|
Switch |
Link |
Free Wavelengths Set |
|
H |
GH |
lGH = {l1,l2,l3,l4} |
|
G |
FG |
lFG = {l2,l3,l4} |
|
F |
EF |
lEF = {l3,l4} |
|
DF |
lDF = {l1,l2} |
|
|
D |
CD |
lCD
= {l1, l2, l4} |
|
BD |
lDF
= {l3, l4} |
Table 1 – Free wavelengths at Figure 4 switches
Suppose SWAP was set to regard 20 pps[4] as a threshold to switch a flow. Switch D, F, G, and H all see an aggregate outgoing throughput of 20 pps or higher for flow F1 (* à Domain H, i.e., traffic going to H) and because switch H knows it is the last hop, it locks the free wavelength resource lGH and send SETUP(F1, lGH) to G. Upon receiving that SETUP, G does the wavelength set intersection lx = lGH Ç lFG = {l2,l3,l4}, lock lFG and sends SETUP(F1, lx) to F. F does the same intersection ly = lx Ç lDF = {l2} and lz = lx Ç lEF = {l3,l4}. Then it will send SETUP(F1, ly) to D and SETUP(F1, lz) to E. Both D and E know that they are the first hops for that path[5] and D will send COMMIT(F1, l2) and E will send COMMIT(F1, l3). Meanwhile, F waits for responses from both. Upon receiving the responses, F picks l3 because it is an element of lEF and E contributed more to the aggregate throughput than D. F removes l3 from the set lEF and l2 from the set lDF, and sends COMMIT(F1, l3) to G. G removes l3 from the set lFG, unlocks it and forwards COMMIT(F1, l3) to H. Then, H removes l3 from the set lGH, unlocks it, configuring its optical switch and flow converter to convert l3 back to the default wavelength l0, and send COMMIT_OK(F1) to G. Upon receiving COMMIT_OK(F1) from H, G sends COMMIT_OK(F1) to F, and switch F starts sending the flow through a wavelength converter that converts default wavelength l0 to l3.
Branch E 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 B to D. D realizes that the flow has been switched, so it sends SETUP(F1, lBD) to B[6]. B determines that it is the first hop and sends COMMIT(F1, l3). Upon receiving from B, D configures its optical switch and sends COMMIT_OK(F1) to B, and B sends the flow using l3 when it receives the message. If, subsequently, E detects that F1 throughput drop, it will send TEARDOWN(F1) to F. Switch F will not forward the message further because it still has a switched branch, i.e.: DF. As a result, F 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, SWAP 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 l0. However, if there is a switched previous hop, the switch will take the last hop role, i.e., converting the switched flow back to l0.
To summarize the SWAP design points:
1. Implemented on top of a reliable transport protocol.
2. First-hop is defined as the point where there is no upstream neighbor, or there are upstream neighbors but the incoming throughputs are not high enough. Last hop is defined as the point where there is no downstream neighbor.
3. Use last-hop-initiated setup if there is no switched path and aggregation-point-initiated setup if the aggregation point already has the flow switched (i.e., the aggregation point is the last hop of the augmentation to the existing switched path)
4. Use first-hop-initiated teardown. Teardown messages are terminated at the merging point if there is still a switched incoming branch.
5. Resources (free wavelengths) are maintained and locked independently for each incoming link.
6. Grooming points should maintain the flow status for each incoming branch and the flow packet count for each non-switched incoming branch.
7. Whether the switch is the first, the intermediate or the last hop is determined using te neighbor protocol.
8. The neighbor protocol is assumed to be available or be provided by routing protocols.
9. If the neighbor protocol reports that there are no upstream and downstream neighbors, SWAP will be disabled because SWAP implementation assumes there are at least two POW switches in sequence (actually, it is useful only if there are at least three switches in a row).
Implementation
There are 6 (six) 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 recepients 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 one of the wavelength sets along the path has been locked by others. Intermediate hops will pass this message to the next hop. The initiating hop will peform a “backoff” procedure.
· SETUP_FAIL(F). Inform the next hop that the connection cannot be made as there are no free resources[7]. Intermediate hops will pass this message to the next hop if they are not the initiating hop.
· COMMIT(F, l). Inform the next hop to use l 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.
In addition to the message types above, SWAP uses an X/Y classifier to detect a flow with the following default parameters:
·
FlowDetectionTimer = 20 secs
·
FlowActiveTimer
= 20 secs
·
HighThreshold
= 10 packets
·
LowThreshold
= 5 packets
·
BackoffPeriod
= 1 ms
·
BackoffLimit
= 10 retries
·
WavelengthConvertEnable
= true
for merging, false otherwise
·
PartialPathAllowed
= false
The messages and the parameters will initiate various state transitions. Each SWAP entity maintains two types of states for each flow entry: incoming and outgoing states. Each incoming branch has its own incoming state, but there is only one outgoing state. Some transitions in the incoming state will initiate further transitions in the outgoing one, and vice versa. This interaction is illustrated in Figure 5 and Figure 6. In these figures, identity (self-loop) transitions are not shown unless the transition causes a message to be sent. Such transitions are indicated by a transition number inside the state, as in case of transition O4 in Figure 6.
To understand SWAP state transitions, some definitions are provided for the further discussion. First, a switchable branch is a branch with PacketCount higher than HighThreshold and has an upstream neighbor. Second, consecutive switchable branches will potentially construct a lightpath. SWAP could construct several combination of lightpaths: contiguous or non-contiguous, and partial or full. A contiguous lightpath is a ligthpath that consist of several branches switched using the same wavelength. A non-contiguous lightpath is a lightpath that consists of several branches switched using different wavelengths. Non-contiguous lightpaths can be constructed if the SWAP parameter of WavelengthConvertEnable is true[8]. A partial lightpath is a path where not all the switchable branches are switched. A partial lightpath is constructed only if SWAP cannot acquire the necessary free resources along the entire path and the parameter PartialPathAllowed is True. Full lightpath is a path where all switchable branches are switched.
Figure 5 – Incoming State Transition[9]
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 flow-branch tuple (F, B), a SWAP process for that particular flow, S, will perform different roles based on the following criterias:
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 Figure 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, then S creates a state for F, passively monitors F, and waits for further signaling messages (path I2 in Figure 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.
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 lB and send SETUP(F1, lB) to its upstream neighbor, SP. If lB was locked by another on-going signaling for flow F2, SA will do a backoff procedure similar to that of CSMA/CD (path I13 and I14 on Figure 5). 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, lin) 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 ln, assign l = lin Ç ln and send SETUP(F, l) to the upstream neighbor. Here the outgoing state of F becomes WAIT_ALL_COMMIT (Figure 6). Alternately, if IC3 and IC5 are true, SP will pick lout Î ln and send COMMIT(F, lout) and upon receiving COMMIT_OK(F), it will take the active role (transition O5 and O6 in Figure 6, transition I4 in Figure 5).
Figure 6 - Outgoing State Transition
On the second phase of signaling, where SP is in the WAIT_ALL_COMMIT state, it will collect all COMMIT(F, l) received on any switchable branch Bn, n=0,1,2…,N and pick one wavelength lx, x=0,1,2,…,N such that Bx is the branch with the highest throughput. SP then will send COMMIT(F, lx) downstream. Meanwhile, SP also records the wavelength of preference of each branch[10].
This COMMIT(F, lx) 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 lx to l0 (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. This message will not be propagated by the downstream nodes 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 backoff procedure. Setup conflict messages must be propagated along the entire signaling path because every SWAP node Sn along the path has locked the wavelength set for the branch. By releasing those locks, SWAP creates opportunities to setup paths belonging to other flows.
The last part of SWAP signaling handles the teardown mechanism (transitions O7, O8, and O9 in Figure 6). If an active node S detects that the throughput of a swiched branch (F, B) drops below LowThreshold for a specific detection duration FlowActiveTimer, S will monitor that flow for another FlowActiveTimer 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.
Modeling and Simulation
SWAP was designed to enable the evaluation of the performance of the POW architecture. To ease SWAP development and verification and to evaluate specific POW topology, SWAP was implemented in VINT/ns, a popular network simulation package [NS].
Figure 7 – NS Model of POW switch
Essential NS components of the POW switch simulation model include (Figure 8):
· Optical Classifier: models the optical switching fabric. For simulation purposes wavelenghts are represented as an integer value in our specific.
· Router Links: model the bi-directional electrical data interfaces between the IP router and the fabric. The interface operates at OC-12 rates.
· Inter-switch Links: model the high speed optical links between POW switch. This interface operates at OC-48 rates.
· Wavelength Converters: simulate wavelength converter banks in the POW switches. For the simulation, a wavelength converter reassigns the integer value in the header representing the wavelength.
· Address Classifiers: model IP router forwarders. The classifier determines which output link a packet should be directed to. Locally-destined packets are delivered to the local application.
· Flow analyzer: an X/Y classifier, where X is HighThreshold (10 pkts) and Y is FlowDetectionTimer (20 secs). The analyzer may perform flow classification based on several tuples (see below).
· SWAP agent: models signaling between the POW switches.
The POW flow analyzer is equipped to classify several traffic flows based on the following types of tuples:
·
Fine Grain: classified based on the tuple:
{ingress node, src ip, src tcp port} à
{egress node, dst ip, dst tcp port}
·
Medium Grain:
{ingress node, src ip} à {egress node, dst ip}
·
Coarse Grain:
{ingress node} à
{egress node}.
·
Merged Fine Grain:
* à
{egress node, dst ip, dst tcp port}.
·
Merged Medium Grain:
* à
{egress node, dst ip}
·
Merged Coarse Grain:
* à
{egress node}
POW nodes are interconnected using NS Tcl script which encodes the vBNS backbone topology, shown in Figure 8. The vBNS network is a good example of the type of environment where POW would be useful.vBNS currently provides IP service on top of an ATM network. vBNS nodes are connected by a complete mesh of PVCs to each other. POW can replace this mesh with a more effective optical Internet.
Although
the main goal of POW is to evaluate the number of wavelengths required per
link, we also evaluated the overhead of SWAP protocol. LBL-PKT5 [ITA] was
chosen as the traffic model(.
Figure
8 - vBNS Backbone Topology
To
instrument the performance measure of SWAP in term of the overhead, the ratio
of SWAP messages is compared to the offered traffic. Five schemes are compared:
non-merged fine, medium and coarse grain flows and merged fine and coarse grain
flows with different numbers of available wavelengths: 0, 4, 8, 16, 32, and 64.
Default wavelength l0 is assumed and is not counted.
Figure 9 - SWAP overhead for non-merging flows
Figure 10 - SWAP overhead for merged flows
To instrument the performance measure of SWAP in
term of the overhead, the ratio of SWAP messages is compared to the offered
traffic. Five schemes are compared: non-merged fine, medium and coarse grain
flows and merged fine and coarse grain flows with different numbers of
available wavelengths: 0, 4, 8, 16, 32, and 64. Default wavelength l0 is assumed
and is not counted.
Figure 98
and Figure 109
indicate that for most cases, SWAP overhead is proportional to the number of
switchable flows. For non-merged fine grain case, there is an unexpected
increase in the overhead for 32 and 64 wavelengths. Figure 110 indicates that there is
no significant increase in the number of flows being switched. Simulation
traces show that there were a large amount of unsuccessful signaling due to the
high number of swichable branches. Primarily there were not enough common
wavelengths to construct the lightpaths, and SWAP incurred a high percentage
of unsuccessful signaling attempts. If
the flows persist after such a failure, this will significantly increase the
number of SWAP messages exchanged. Similar behavior is observed for
merged-flows.
Figure 11
- Percentage of Packet Swithed for non-merged flows
Figure 10 exhibits this behavior more clearly for merged-flows, where SWAP has three different impediments: insufficient wavelengths overall, sufficient wavelengths but insufficient common wavelength along the path, and insufficient switchable flows. Both Figures 9 and 10 show three phases in the graph. For the merging case, the first phase is between 0 and 8 available wavelengths. This phase shows the overhead of SWAP when there were not enough wavelengths: the number of signaling messages grows slowly because most of the messages are SETUP and SETUP_FAIL. In this phase, most SETUP attempts fail quickly because the free wavelength sets for the links between the last hop to the previous hop were empty.
The second phase between 8 and 32 available wavelengths corresponds to the case when there were enough wavelengths (more opportunity to switch the flows), and SWAP became more active performing its signaling. However, because there were insufficient common wavelengths along the enture path, SWAP kept trying unsuccessfully. In this case, the number of SWAP messages grows more rapidly in this case, and this could become a scalability issue if SWAP were used when the ratio between Common Wavelengths available per the number of Switched Flows (CW/SF) is low. In this case the competition for the wavelengths is not self-limiting, as in the case with too few switched flows. It may be necessary to consider preemption, where new large flows cause current small switched flows to be torn down, and wavelengths reassigned. Another alternative would be to consider limiting the propagation of setup messages, relative to the number of current flows, number of wavelengths, and likely length of the path. Either of these alternatives would require augmentation of the SWAP protocol.
Last phase is between 32 and 64 available wavelengths, and corresponds to sufficient common free wavelengths, where the number of SWAP messages during signaling was bounded. In this case, most of the SETUP attempts succeed. These three phases give an insight on the scalability of SWAP in term of the messaging overhead when CW/SF is very small (approaching zero), small, and large.
Conclusion
The SWAP protocol is an approach to assigning wavelengths dynamically based on traffic demand, and was created to facilitate the Packet over Wavelength (POW) architecture. SWAP’s wavelength assignment is similar to IP switching; it is data-driven, rather than control-driven. A notable difference is that POW introduces flow merging to reduce the number of wavelengths required. This requires SWAP 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 POW simulation, SWAP overhead was measured under three conditions: where the Common Wavelength Per Switchable Flows (CW/SF) ratio is zero, small, and large. SWAP scales well for zero and large CW/SF. For a small CW/SF ratio, SWAP scales poorly due to the breadth of wasted signaling. This may suggest that SWAP need to be augmented to support lightpath preemption, where new larger flows cause smaller existing flows to be torn down, and wavelengths reassigned. Preemption will enable SWAP to switch the N most significant switchable flows (for N free wavelengths), where the SWAP overhead would thus be bounded.
Future study is required to measure SWAP performance under higher traffic loads. In addition, further investigation is required to determine how SWAP parameters affect its steady-state and transient performance.
References
[Cal+97] R. Callon, et.al., “A Framework for Multiprotocol Label Switching”, (work in progress).
[DaDR98] B. Davie, P. Doolan, and Y. Rekhter, Switching in IP Networks: IP Switching, Tag Switching and Related Technologies, Morgan Kaufmann, San Francisco CA, 1998.
[Kesh97] S. Keshav, An Engineering Approach to Computer Networking: ATM Networks, the Internet and the Telephone Network, Addison-Wesley, Reading MA, 1997.
[ITA] http://ita.ee.lbl.gov/html/traces.html
[LeLi93] K.C. Lee and V. O. K. Li, “A Wavelength-Convertible Optical network,” Journal of Lightwave Technology, Vol. 11, No. 5/6, May 1993, pp. 962-970.
[LiMc97] S. Lin and N. McKeown, “A Simulation Study of
IP Switching,” Proceedings of ACM SIGCOMM 1997, pp 15-24. . PAGES?
[Mukh97] B. Mukherjee, Optical Communication Networks, McGraw-Hill, New York NY, 1997.
[NeML98] P. Newman, G. Minshall, and T. Lyon, “IP Switching - ATM Under IP,” IEEE/ACM Transactions on Networking, 6(2), April 1998, pp. 117-129.
[NLANR] http://www.nlanr.net/NA/
Learn/aggregation.html
[NS] http://netweb.usc.edu/vint
[Perl92] R. Perlman, Interconnections: Bridges and Routers, Addison-Wesley, Reading MA, 1992.
[RaSi98] R. Ramaswami
and K.N. Sivirajan, Optical Networks: A Practical Prespective, Morgan Kaufmann,
San Francisco CA, 1998.
[Yoo96] S.J.B. Yoo, “Wavelength Conversion Technologies for WDM Network Applications,” IEEE Journal of Lightwave Technology, Vol. 14 No. 6, June 1996, pp. 955-965.
[ZDE+93] L. Zhang, S. Deering, D. Estrin, et.al. “RSVP: A New Resource ReSerVation Protocol,” IEEE Network, Vol. 5, September 1993, pp. 8-18.
