Multi-constraint routing system and method

[0001] This application is a Continuation-in-part of the U.S. patent application to Lee et al. entitled “Multi-Constraint Routing System and Method”, Ser. No. 10/025,869 filed on Dec. 26, 2001 now U.S. Pat. No. 6,925,061, and claims priority of U.S. provisional application to Lee et al of the same title, Ser. No. 60/362,926 filed on Mar. 11, 2002.

[0002] The invention relates to systems and methods of identifying routes through a network in accordance with multiple constraints.

[0003] The problem of identifying routes through a network subject to a single constraint is well known. For example, Shortest Path First Algorithms by Dijkstra and Bellman-Ford are well known, and both use a single cost, single constraint approach. Dijkstra's method finds the shortest path between a source and any other node in the network.

[0004] The problem of routing while considering multiple constraints, while newer than the single constraint problem, has been addressed in a number of ways. See for example Kenji Ishida, Kitsutaro Amano, “A delay-constrained least-cost path routing protocol and the synthesis method”, IEEE conference, 1998; Ariel Orda, “Routing with End to End QoS guarantees in broadband networks”, IEEE conference, 1998; H. F. Salama, D. S. Reeves, Y. Viniotis, “A distributed algorithm for delay-constrained unicast routing”, IEEE conference, 1997; J. Song, H. K. Pung, L. Jacob, “A multi-constrained distributed QoS routing algorithm”, IEEE conference, 2000; S. Chen, K. Nahrstedt, “On finding multi-constrained paths”, IEEE conference, 1998; J. Zhou, “A new distributed routing algorithm for supporting delay-sensitive applications”, ICCT, 1998.

[0005] Most of these approaches use the Dijkstra or Bellman-Ford algorithms as a base algorithm to find a sub-optimal path.

[0006] Typical IGP (Internal Gateway Protocol) routing protocols such as OSPF (Open Shortest Path First) advertise a single link metric termed “administrative cost”. The default link metric in some cases for this “administrative cost” is the inverse of link bandwidth. Other implementations may use monetary cost, propagation delay etc.

[0007] IGP-TE (traffic engineering extension to IGP) adds one more metric which can be advertised within a network. In general, the TE-metric is used to carry any arbitrary metric for constraint-based routing of a set of LSPs (Label Switched paths) needing optimization. The current common practice is to use the IGP metric as a second TE metric, see e.g. IEFT draft “Use of IGP Metric as a second TE Metric” by Francois Le Faucheur et al (expires April 2002) which can be found at http://search.ietf.org/internet-drafts/draft-lefaucheur-te-metric-igp-01.txt

[0008] Proposals have also been made to extend the number of TE metrics which are to be advertised up to three, see e.g. IEFT draft “Multiple metrics for Traffic Engineering with IS-IS and OSPF” by Don Fedyk et al (expired May 2001) which can be found at http://search.ietf.org/internet-drafts/draft-fedyk-isis-ospf-te-metrics-01.txt

[0009] The cost associated with a path between a source and destination is the sum of the costs of the links that constitute the entire path. Existing proposals recommend not to strive for optimization of both the TE metric and the IGP metric during path computation for a given LSP. This is known to be an NP-complete (non-polynomial complete) problem.

[0010] Thus there is a need for heuristic strategies and systems for performing multi-constraint routing to minimize the cost for the entire path.

[0011] One broad aspect of the invention provides a method of performing multi-constraint routing. The method involves determining a composite cost, composite cost(link)=f(metric1, . . . ,metricN), for each of a plurality of links under consideration in a network topology, the composite cost being a function f of a plurality of metrics metric1, . . . metricN for each link, where N>=2; and performing routing through the network topology from a source to a destination based on the composite costs.

[0012] In some embodiments, the composite cost is determined according to:

[0013] $\mathrm{composite}\mathrm{cost}\left(\mathrm{link}\right)=\sum _{i=1}^N{c_i\left({\mathrm{metric}}_i\right)}^{n\left(i\right)}$
where ci is a constant for the ith metric, and n(i) is an exponent for the ith metric.

[0014] In some embodiments, the composite cost is determined according to:
composite cost(link)=a*(1/abw)+b*Admin cost+c*Delay;
where metric1=abw=available bandwidth, metric2=Admin cost and metric3=Delay=transmission delay, and wherein c1=a, c2=b, and c3=c, n(1)=−1, n(2)=1 and n(3)=1.

[0015] In some embodiments for each link, at least one network node advertises the link's metrics across the network such that any node in the network may implement the multi-constraint routing.

[0016] In some embodiments, wherein the composite cost increases with decreasing available bandwidth, routing is performed in a manner which promotes load balancing.

[0017] In some embodiments, the routing is performed in a manner which promotes load balancing with demand reservation on at least one link. This may involve for example defining a respective demand threshold (demand_threshold) for at least one link representing demand reservation for the at least one link.

[0018] This may involve adjusting the costs according to:
if (available bandwidth<=demand_threshold) cost(link)=composite cost(link)+MAX COST
else cost(link)=composite cost(link)
where demand_threshold is a respective demand reservation for the link, and MAX COST is a very large value.

[0019] Typically, the demand_threshold is set to a first value for at least one of the links, and is set to a second larger value for at least one of the links whereby bandwidth on the at least one of the links having the first value is reserved for larger bandwidth requests.

[0020] In some embodiments, for each of a plurality of different traffic types, a respective different composite cost equation is used for calculating the composite costs. The different traffic types might for example include video, voice and data, and might include different classes of service/qualities of service.

[0021] In another embodiment, a link cost is computed based on the composite cost for switching between a load balancing mode and a bin-packing mode based on available bandwidth of each of a plurality of links under consideration using a predetermined demand_threshold value in a network topology. In this method the cost for each said at least one link is determined as follows:
If (abw<=demand_threshold) cost(link)=a*Admin cost+b*(abw^K/PIR)+c*Delay+MAX COST
else
cost(link)=a*Admin cost+b*(PIR/abw^K)+c*Delay;
wherein PIR=Peak Information Rate, MAX COST is a large number, abw is the available bandwidth and K is a constant.

[0022] The constant K can be used to control a rate at which the cost changes with a variation in the available bandwidth. The large number (MAX COST) is added to the cost (link) to prevent the link from being selected, when there is at least one link with available bandwidth higher than the demand_threshold.

[0023] In still another embodiment, the cost for each said at least one link is determined as follows:
If (available bandwidth<=demand_threshold) cost(link)=a*Admin cost+b*(PIR/ abw^K)+c*Delay+MAX COST
else
cost(link)=a*Admin cost+b*(abw^K/PIR)+c*Delay;
wherein PIR=Peak Information Rate, MAX COST is a large number, abw is the available bandwidth and K is a constant.

[0024] The constant K can be used to control a rate at which the cost changes with a variation of the available bandwidth. In this embodiment, a different demand_threshold value is used for at least one link for achieving a different switch-over point between load balancing and bin-packing. In some embodiments, the method further involves providing at least one pruning constraint for each link, and pruning the network topology on the basis of the at least one pruning constraints to remove at least one link from the network topology. In this case, performing routing through the network from a source to a destination based on the composite costs is conducted using only the links not pruned from the network topology. The pruning of the network topology on the basis of the at least one pruning constraints to remove at least one link from the network topology is performed on an as-needed basis to remove one or more links from a set of candidate links.

[0025] Alternatively, pruning the network topology on the basis of the at least one pruning constraints to remove at least one link from the network topology may be performed on the entire network topology prior to performing the routing.

[0026] In some embodiments, at least one constraint is used both as a pruning constraint and as a constraint used in the composite costs.

[0027] Another broad aspect of the invention provides a routing system adapted to perform multi-constraint routing. The routing system has a topology repository adapted to store information identifying links, nodes and connections of a network. The routing system further comprises an input for receiving on an ongoing basis for each link, updated values for the metrics for the link. There is a multi-constraint router adapted to determine a composite cost; wherein
composite cost(link)=f(metric1, . . . , metricN)
for each of a plurality of the links in the network topology, the composite cost being a function f of a plurality of metrics metric1, . . . metricN for each link, where N>=2, and to perform routing through the network topology from a source to a destination based on the composite costs. The routing system may be implemented to provide any one or more of the above introduced methods.

[0028] In another embodiment the routing system is adapted to compute the cost for each said at least one link based on the composite cost for switching between a load balancing mode and a bin-packing mode based on available bandwidth of each of a plurality of links under consideration in a network topology. The switching between the load balancing and the bin-packing modes is performed based on a predetermined demand_threshold value. A different demand_threshold value is used for at least one link for achieving a different switch over point between load balancing and bin-packing.

[0042] FIG. 1 is a diagram of an example system within which an embodiment of the invention may be employed. The system has a CM (connection management) function 16 through which requests for new connections are received. An MPLS (multi-protocol label switching) protocol 14 is used to propagate requests to network components after a potential route has been identified. A CAC (call admission control) function 10 controls whether or not calls are to be admitted. An RT (resource tracker) 12 keeps track of resources. The CM 16 interacts with a routing system 19 which has an MCR (multi-constraint routing) function 18 which identifies routes through a network as a function of a network topology indicated at 20.

[0043] The process in setting up a connection involves first a customer 13 sending a service request to a network management system 11 which forwards the request to CM 16 which translates the request into a meaningful path request message (with source, destination, constraints). CM 16 then requests a path from routing system 19. The MCR function 18 within the routing system 19 calculates an optimal path between the source and destination satisfying the constraints using a new algorithm provided by an embodiment of the invention. Then, CM 16 requests MPLS 14 (or other route establishment system) to set up the path. MPLS 14 checks for resource availability before sending the request on to the next hop as indicated by arrow 15.

[0044] In one embodiment, given a network G=<N,E>, where N is the set of nodes and E the set of links, the task of the MCR function 18 is to find an optimal path between a given source and destination satisfying two sets of constraints. The first set of constraints is referred to herein as pruning constraints, and these constraints are constraints which allow the elimination of one or more routes from the set of possible routes. These are constraints that assist in narrowing the search to a reduced candidate path set. Examples of potential pruning constraints include bandwidth, pre-emption, path exclusion, Peak Information Rate (PIR), and link protection. The second set of constraints is referred to herein as cost constraints, and these are used in considering routes left over after application of the pruning constraints. These constraints are used in calculation of cost for every path in the reduced candidate set. Examples of potential cost constraints include administrative cost (administrative cost), available bandwidth (abw), and delay. Not every implementation necessarily will include pruning constraints, but all implementations will include cost constraints. Some constraints may be used both as pruning and cost constraints.

[0045] An example set of constraints is shown in the table of FIG. 4 where eleven different example constraints are given. Bandwidth per CoS (class of service) may be employed as a pruning or cost constraint. The pre-emption state of a link (indicating whether it is pre-emptable or not) may be used as a pruning constraint. Path exclusion (indicating that a particular path should not be used, e.g. maintenance) is also a pruning constraint. The source-destination fixed delay is a cost constraint. L1 logical link protection and L2 logical link protection are pruning constraints as are the PIR, group path bandwidth, and path diversity. Best effort load balancing and administrative cost are both cost constraints. It is to be understood that this is a non-limiting set of example constraints. Some or all and/or completely different constraints than these example constraints may be employed in an actual implementation. In some applications, it may be useful to classify the constraints differently than as outlined above.

[0046] A more detailed block diagram of the routing system 19 is shown in FIG. 2. The routing system 19 has a MCR control layer 30 which co-ordinates the receipt and response to requests from the connection management function CM 16. A pruning filter 32 performs pruning on the network topology 20 to produce a modified network topology 34. Cost function 38 performs a cost calculation on the modified network topology 34 and identifies a minimum cost path.

[0047] The function of the routing system in identifying a route from a source node to a destination node will be described in further detail with reference to the flowchart of FIG. 3. In identifying the route, two different lists are maintained on an ongoing basis. The first such list is referred to as the link_list, and is the list of links directly connected to a current node (the node for which a next link in the route is to be identified). The second such list is referred to as the candidate_list, and is the list of all possible routes/partial routes identified thus far. These lists are clarified below.

[0048] In step 3-1, the algorithm starts at each current node with the creation of a link_list consisting of all links connecting to the current node. Initially, the current node is the source node, and the initial link₁₃ list will be all links connecting directly to the source node. Also, initially, the candidate_list will be empty. It is noted that the whole algorithm is typically run by the same node, and the nodes referred to are logical nodes in the network topology.

[0049] In step 3-2, all links in the link_list that do not satisfy the pruning constraints are pruned, i.e. removed from the link_list.

[0050] In step 3-3, the candidate_list is updated as a function of the pruned link_list. For the first iteration, the candidate_list is set to equal the link_list of links leaving the source node. For subsequent iterations, the candidate_list is updated by adding to the candidate_list routes which make use of a previous route in the candidate_list plus a link in

[0051] In step 3-4, for all routes newly added to the candidate_list, a cost is calculated based on the cost constraints. This involves first calculating the cost for each link in the pruned link_list. Then, the cost for a particular route which makes use of a previous route in the candidate_list and a particular link from the pruned link_list is calculated by adding the cost of the previous route to the cost of the link from the pruned link_list.

[0052] Finally, in step 3-5, the route that has the least cost as determined by the cost constraints is selected for the current iteration. The next “current node” is the node terminating the selected route. When a particular route is selected, the previous route in the candidate_list which was used to build the particular route is removed from the candidate list. All other routes remain in the candidate list.

[0053] These steps are repeated until the destination is reached.

[0054] In another embodiment, the pruning of the entire network topology is done at the very beginning, and then a standard routing protocol run on the remaining topology.

[0055] Another broad aspect of the invention provides a cost function for use in performing routing functions. In one embodiment, the cost function is employed in conjunction with the above-described system with two constraint types. More generally, the approach can be used in any routing system. A new composite cost function having the following format is employed:
composite cost(link)=f(metric1, metric2, . . . ,metricN)
where f(.) is a multi-variable function, and metric1, . . . ,metricN are metrics advertised for each link. In all cases N>=2, and in a preferred embodiment, N=3.

[0056] In a specific example, the composite cost function has the following format:

[0057] $\mathrm{composite}\mathrm{cost}\left(\mathrm{link}\right)=\sum _{i=1}^N{c_i\left({\mathrm{metric}}_i\right)}^{n\left(i\right)}$
where ci is a scaling constant for the ith metric, and n(i) is an exponent. In a particular example of this format, consider the case where the three metrics are bandwidth, administrative cost and delay. A very specific preferred format of the composite cost is:
Cost(link)=a*(1/abw)+b*Admin cost+c*Delay
where a, b, c are constants that can be engineered depending upon goals defined for the routing system. In this equation, “abw” is the available bandwidth, and “a” is the bandwidth coefficient. Admin cost is the administrative cost provided by administrator, and “b” is the administrative cost coefficient. Delay is the transmission delay introduced by the link and “c” is the delay coefficient. In this case, metric1=available bandwidth, metric2=Admin cost and metric3=delay, and wherein c1=a, c2=b, and c3=c, n(1)=−1, n(2)=1 and n(3)=1.

[0058] It should be noted that this composite cost function approach is not an attempt to optimize multiple link metrics simultaneously, which can be degenerated to an NP-complete problem. Rather, it is an approach to take multiple link metrics into consideration in a flexible manner.

[0059] An example network topology to which the MCR function may be applied is shown in FIG. 5. FIG. 5 shows a topology consisting of eight nodes labeled A, B, C, D, E, F, G and H, and having twelve links AB, AD, AE, BC, BE, CF, CD, DG, EF, EH, GF, FH. This network topology will be used in example applications described below. It is assumed that link AE is a layer 2 logical link which would actually pass through node B.

[0060] An example set of metrics associated with the various links of FIG. 5 is presented in the table in FIG. 6. There is a set of parameters for each link, and in the event a given link has a protection path, there is a set of parameters for the protection path. In the illustrated example, it is assumed that links CD and AD have both working and protection paths. Each link has an available bandwidth field which indicates an available bandwidth metric for the link. This needs to be updated each time a new route is provisioned on a given link. Each link has a maximum bandwidth field indicating the maximum total bandwidth supported by the link. Each link also has an administrative cost metric and a delay metric. The available bandwidth, admin cost, and delay are preferably advertised network wide by whatever node in the system has first hand knowledge. For embodiments in which pruning is to be performed on the basis of the type of protection available for links, the table would include for each link an L1LL field when appropriate indicating for layer 1 links the type of protection for the link if any. Each link would also have an L2LL field when appropriate indicating for layer 2 logical links the type of protection for the link if any.

[0061] For this example, it is assumed that a customer wants to set up a link between nodes A and F. The connection management function 16 processes the request, and converts it into a request which makes sense to the routing system 19. In one example, the request is converted into the following set of parameters:

[0062] The minimum set of parameters are source, destination and bandwidth:

[0063] Source—identifies the source node for the request;

[0064] Destination—identifies the destination node for the request;

[0065] Bandwidth—the requested bandwidth for the connection.

[0066] In some implementations, one or more of the following may also be included as part of the request:

[0067] Pre-emptable state—identifies whether the link is pre-emptable—only required in embodiments contemplating pruning based on this state. For this example, it is assumed that pruning based on this state is not to be employed;

[0068] Delay—identifies a maximum delay tolerable for the requested route. Other parameters/constraints might also be included.

[0069] An example request might be as follows:
(source=A, destination=F, bandwidth=150)

[0070] These parameters would be passed to the routing system 19 which then identifies the best route through the network. For this example, the previously introduced cost function is used, namely:
composite cost (link)=a*(1/abw)+b*Admin cost+c*Delay
and for this example, we assume that a=1000, b=100, and c=100.

[0071] Referring to FIG. 7, shown is a version of FIG. 5 with the composite cost values filled in for each link. The working and protection paths for links AD and DC are also shown. It is noted that the composite costs need not necessarily be all computed at once, but rather in a stepwise fashion as required, and as described below. Next, any routing algorithm may be employed using the composite costs as a single cost per link.

[0072] The composite cost approach, which employs the inverse of the available bandwidth, will result in load balancing occurring across available links, at least to a certain degree depending upon the impact of the other metrics on the composite cost. Load balancing means that the occupancy of the various links will increase in a somewhat balanced way.

[0073] In another embodiment of the invention, a different cost function is employed for different connection requests. The particular cost function may for example be selected as a function of traffic type for the connection. For example, the traffic types might include voice, video, and data, and three different cost functions may be employed for these different traffic types. By way of a very specific example, for the three metric example given above, this can be summarized as follows:

[0074]

		a = bandwidth	b = admin	c = delay
	Traffic Type	coefficient	coefficient	coefficient
	Video	a1	b1	c1
	Voice	a2	b2	c2
	Data	a3	b3	c3

Then, if a request comes in for a video connection, the costs are computed using a1, b1 and c1. If a request comes in for a voice connection, the costs are computed using a2, b2 and c3. Finally, if a cost comes in for a data connection, the costs are computed using a3, b3 and c3.

[0075] More generally, a different cost equation may be used for different connection request types. This could be based on class of service/quality of service, or some other type of characteristics.

[0076] In another embodiment of the invention, the composite costs can be manipulated in such a way as to promote load balancing with demand reservation. Bin-packing involves preferentially reserving a portion of bandwidth on certain links until a sufficiently large bandwidth request is received, thereby ensuring that the large bandwidth request can in fact be serviced. Consider a simple example in which two possible links each have a maximum bandwidth of 100. If load balancing is employed until such time that the two links have an occupancy of 85 and 90 respectively, a new request for 20 will not be serviceable on either link. To decrease call blocking probabilities, bin-packing routes all traffic to one link until full then moves on to next link but this introduces congestion on particular links and does not follow the load balancing approach.

[0077] In this embodiment, the links are load balanced up to certain points. Demand reservation is employed to reserve some capacity for larger bandwidth connections. For this example, in order for load balancing to occur, there must be an available bandwidth term in the composite cost, and the composite cost must increase with decreasing available bandwidth such as is the case when a term a (1/available bandwidth) is included. An example of load balancing with demand reservation will be described with reference to FIG. 8. Shown is the bandwidth occupancy of two alternative links, link1 and link2. A demand threshold DT is indicated for the two links. In the example, the DT for link1 is 30 and the DT for link2 is 10, the demand threshold being a measure in per cent of a specified fraction of the maximum bandwidth, normalized to 100. Load balancing is employed up to and including when a connection is made which results in a watermark being exceeded by one or more of the links. Load balancing means the normal composite costs introduced earlier are used in performing the routing function. At that point, a new conditional composite cost is calculated as follows for links which are to have reserved bandwidth:
if (abw≦DT (demand_threshold) cost(link)=composite cost(link)+MAXCOST
else
cost(link)=composite cost(link)
where “MAXCOST” is just a large number, which will prevent the link from being selected. If two links both have this MAXCOST then they will cancel each other out and load balancing will prevail. Thus, in the simple two link example, after the watermark is passed, one of the links is selected to have reserved bandwidth for higher bandwidth requests. In the illustrated example, the demand threshold for link1 has been crossed and there is an available bandwidth of 27. The demand threshold for link2 has not been crossed and there is an available bandwidth of 15. Now, if a requested bandwidth of 25 is received, there is room for this on link1 because of the demand reservation made for that link. When both demand thresholds are crossed, the method reverts to load balancing again.

[0078] In another embodiment of the invention, routing is performed in a manner which promotes load balancing as well as bin-packing and the routing strategy switches between a load balancing mode and a bin-packing mode based on the available bandwidth of the links. For example, when the available bandwidths of the candidate links under consideration during path selection are higher than the demand_threshold, load balancing is performed, whereas for lower values of available bandwidth, bin-packing is used. In this embodiment of the routing system, the method for computing the composite cost for a link is shown in flow chart 9-0 of FIG. 9. Upon start (box 9-1), when the available bandwidth of a candidate link considered during a step in path selection is more than demand_threshold (exit “No” from box 9-2), step 9-4 of FIG. 9 is executed followed by termination of the method (box 9-5). The equation in step 9-4 indicates that composite cost(link) increases with a decrease in the available bandwidth (abw). K is a constant that can be used to control the rate at which the cost increases. Thus, the algorithm selects a link with a higher available bandwidth, i.e., a less congested link over a link with the same Admin cost and Delay parameters but a lesser available bandwidth. This tends to distribute the load evenly across the links giving rise to load balancing.

[0079] When the available bandwidth of a link under consideration is less than or equal to the demand_threshold (exit “Yes” from box 9-2), step 9-3 of FIG. 9 is executed followed by termination step 9-5. The equation in step 9-3 indicates that composite cost(link) decreases with a decrease in available bandwidth. Given two candidate links with the same Admin cost and Delay parameters, the routing algorithm will choose a link with a lower available bandwidth (lower cost) over a less congested link. This scenario corresponds to bin-packing: the congested link is used up first, leaving spare capacity on the less loaded links that can potentially accommodate a request with a high bandwidth demand in the future. The behavior of the algorithm is captured in FIG. 10. The addition of a large number MAX COST (see box 9-3), prevents a link from being selected when other candidate links are operating under the demand_threshold. When abw for all these links are less than or equal to demand_threshold, the system enters the bin-packing mode. Two values for the demand_threshold, 0 and PIR, may be considered further. By setting demand_threshold to 0, step 9-3 in FIG. 9 is never executed and the system always operates in a load balancing mode. Similarly by setting demand_threshold to any value higher than PIR, step 9-4 is never executed and the system always operates in a bin-packing mode.

[0080] In another embodiment, the method enables the system to perform load balancing as well as bin-packing, wherein the routing strategy switches between a load balancing mode and a bin-packing mode based on the available bandwidth of links in such a manner that when the available bandwidths of the candidate links considered during path selection are higher than demand_threshold, bin-packing is performed and load balancing is used for lower values of available bandwidths. The composite cost for a link is computed using the method shown in flow-chart 11-0 in FIG. 11. Upon start (box 11-1), and as long as the available bandwidth of a link is more than the demand_threshold (exit “No” from box 11-2), step 11-4 is executed followed by termination of the method (box 1-5). The second component of the right-hand-side (RHS) of the equation in 11-4 is directly proportional to available bandwidth and cost increases as available bandwidth increases. Given the same Admin cost and Delay parameters, the algorithm selects a more congested (lower available bandwidth) link prior to a link with a higher available bandwidth. This tends to pack more requests on the same link, giving rise to bin-packing. When the available bandwidth of a candidate link considered during a step in path selection becomes less than or equal to demand_threshold (exit “Yes” from box 11-2), step 11-3 of FIG. 11 is executed followed by termination of the method (box 11-5). Since the second term on the RHS of the equation in 11-3 increases with a decrease in available bandwidth, the algorithm tends to load lesser congested links first. Given two links with same Admin and Delay parameters the routing algorithm will choose the link with a higher available bandwidth (lower cost) over a congested link with lesser available bandwidth. This scenario corresponds to load balancing, where the traffic tends to be distributed evenly across the links. The behavior of the algorithm is captured in FIG. 12. The addition of a large number MAX COST (see step 11-3) prevents a link from being selected when other candidate links are operating under demand_threshold. When “abw” for all these links are less than or equal to demand_threshold, the system enters the load balancing mode. Two values for the demand_threshold may be considered further. By setting the demand_threshold to 0, step 11-3 in FIG. 11 is never executed and the system always operates in a bin-packing mode. With a demand_threshold higher than PIR, step 11-4 is never executed and the system always operates in a load balancing mode.

[0081] In another embodiment, the system enables the network operator to select a routing strategy depending upon the nature of request, through an appropriate software mechanism. For example, if most of the requests require small bandwidths, the method described in FIG. 8 may be selected. If a large proportion of the requests are of comparable bandwidths but there are some requests that require large bandwidths, the method described in FIG. 9 may be used. On the other hand if most requests demand large bandwidths of variable size, the method described in FIG. 11 may be preferable.

[0082] Numerous modifications and variations of the present invention are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims, the invention may be practiced otherwise than as specifically described herein.