WO1990001237A1

WO1990001237A1 - Adaptive routing of network traffic

Info

Publication number: WO1990001237A1
Application number: PCT/US1989/003185
Authority: WO
Inventors: Komandur R. Krishnan
Original assignee: Bell Communications Research, Inc.
Priority date: 1988-07-25
Filing date: 1989-07-24
Publication date: 1990-02-08
Also published as: DE68915281T2; EP0426737A1; DE68915281D1; JP2743103B2; JPH04502239A; EP0426737B1; CA1310726C; US4931941A

Abstract

A process is disclosed for adaptively routing service requests through a network (100) comprising nodes (101-104) interconnected with links (201-206). The network is also arranged so that each node may communicate with each of the other nodes. At preselected time intervals, sets of routes through the network are generated in response to the network configuration and traffic information. Over predetermined periods, traffic data is measured to determine carried load and blocking for each link. Also, at predetermined time intervals, occupancy factors are computed as determined by the network configuration and in response to traffic information including measured traffic data. Each of the occupancy factors is derived from a nominal routing scheme in which a call blocked on a single route is treated as a lost call as in separable routing. Upon a request for service, the occupancy factors corresponding to the busy-idle status of the links are used to compute an occupancy value associated with each of the routes. Each route is converted to a candidate route based on traffic load at the initiation of the service request. The minimum occupancy value for each set of routes is selected as the candidate route for bridging a given node pair. If this minimum value is less than a preselected threshold, the traffic is routed over this candidate route, thereby satisfying the service request. Otherwise, the request is denied.

Description

ADAPTIVE ROUTING OF NETWORK TRAFFIC

Technical Field of the Invention

This invention relates generally to routing of traffic through a complex network such as a telecommunications system and, more particularly, to a methodology for the state-dependent, adaptive routing of traffic in a stored-program-controlled environment. Background of the Invention

A basic function required of a typical communication network is one of routing a service request through the network and then establishing a connection between an originating point and a destination point according to the route selected. In the terminology of telephony, this activity is often described as a calling party placing a call to a called party. The call is established on a path through the network designated a route. In turn, a route may be comprised of one or more intermediate points or nodes interconnected to other nodes by one or more links or trunk groups comprising servers or trunks.

Examples of such networks are the intra-LATA (Local Access and Transport Area) telephone network and the inter-LATA (that is, long distance) telephone network.

It occasionally happens that the traffic, e.g., the call, is blocked either involuntarily or voluntarily. An involuntary blocking occurs when no route through the network can be found which has at least one free trunk in all the trunk groups. A voluntary blocking occurs when, even though an available route can be found, the call is blocked to protect future call attempts. A voluntary blocking may be indicated when the only available routes all go over two or more trunk groups, with a sizable risk that carrying one more call now will lead to multiple call blockings in the near future.

Traffic routing is evolving from time-invariant, hierarchical schemes to time-dependent, non-hierarchical schemes. In the first scheme, a routing plan with a hierarchical ranking of nodes is employed. In the telephone environment, these nodes may also be considered as switching centers since the nodes arc implemented with switches. The highest ranking center has been designated a regional center (RC) followed in rank order by a sectional center (SC), a primary center (PC) and a toll center (TC). Each RC has one or more SCs homing on it; each PC homes on an RC or SC and has one or more TCs homing on it. Under this plan, any center homes on another center of higher rank within the same region but not on a center of lower rank. Time-invariant, hierarchical routing through such a network is effected by progressive routing (PR).

To describe briefly progressive routing and, shortly, other conventional schemes, a simplified hierarchical network is considered. In this illustrative network, a first PC (PC1) and a second PC (PC2) are directly connected; similarly, a first TC (TC1) connects to a second TC (TC2). Moreover, TC1 homes on PC1 and TC2 homes on PC2; TC1 is also connected to PC2 by a direct trunk group and, similarly, TC2 is connected to PCI by a direct trunk group. The trunk groups (TG) interconnecting the various centers are designated by: TG1 links TC1 and TC2; TG2 links TC1 and PC2; TG3 links TC2 and PC2; TG4 links TC1 and PC1; TG5 links TC2 and PC1; and TG6 links PC1 and PC2. The call under consideration is to be routed from TC1 to TC2.

There are four possible routes between TC1 and TC2 defined as follows: the first route (R1) is composed of TG1; R2 includes TG2 and TG3; R3 comprises TG4 and TG5; and R4 comprises TG4, TG6 and TG3. In this progressive routing example, the first route considered, as required by the hierarchical ranking, is R1. If TG1 has a nonblocking status, the call is established over TG1. However, if TG1 is blocked, then R2 is considered next. If TG2 is free, routing control is passed from TCI to PC2, without regard to the blocking status of TG3, the next link in R2. If TG3 is blocked, the calling party is given a network congestion signal indicative of a blocked route. With progressive routing, R3 or R4 is never tested if TG2 of R2 is free. In this example, it may have been possible to route the call over R3 if the status of each trunk group in R2 was known to TC1. With progressive routing, the routes between originating and destination points are not considered globally, but rather are treated on a local, step-by-step basis. Consideration on a local basis has, in part, been dictated by communication and signaling limitations imposed on the nodes.

With the present availability of stored program control (SPC) and socalled Common Channel Signaling (CCS) systems, communication among the various centers may now be effected without regard to hierarchy. One such routing method which advantageously utilizes the properties of SPC and CCS is disclosed in U. S. Patent

No. 4,345,116, issued August 17, 1982. The subject matter of this reference, described as a dynamic, non-hierarchical routing (DNHR) scheme, is one example of time-dependent, non-hierarchical techniques.

As disclosed in the reference, the methodology controls the generation of a sequence of routes between a source node or switching point and a terminating node or switching point, each route choice being generated without regard to any network hierarchy. Each route, however, is generated in response to traffic demands during predetermined intervals and is subject to a grade of service constraint. That is, the selected sequence is one which is time sensitive and which is so chosen that its use during the specified time interval tends to mitigate the cost of the network that must be provided to meet grade of service. In the notation of the foregoing example, a sequence of routes between TC1 and PC2 for a first time period may be {TG1 + TG3; TG2; TG4 + TG6}, whereas in another time period the sequence may be {TG4 + TG6; TG1 + TG3}. If it is required to establish a call from TCI to PC2 during the first time period, then the sequence associated with the source (TC1) and termination PC2) nodes for this time period is accessed and processed sequentially by TC1. Thus, the route comprising TG1 and TG3 is selected initially and TC1 requests information, via the CCS network, from node TC2 regarding the blocking status of the pertinent trunk group TG3 terminating on intermediate node TC2. If it is supposed that TG1 is free but TG3 is blocked, then a so-called crankback signal is transmitted to the source node TC1 so the next route in the sequence, namely TG2, may be selected. If TG2 is not blocked, the. call would be routed accordingly. This result should be contrasted to the progressive routing scheme; if progressive routing had been used with the sequence of routes under consideration, the call would have been blocked.

In a broad sense, it may be said that the foregoing technique for generating a sequence of routes between node pairs is dynamic in that routing decisions depend on the time of day or week. However, this technique is not truly state dependent. A state dependent routing scheme attempts to minimize call blocking by making, every time a call is attempted, a sophisticated routing decision based not only on the time of the call but also on the numbers of busy and idle trunks in the various trunk groups.

One example of such a state-dependent, dynamic, non-hierarchical routing (SDD) scheme is disclosed in a paper entitled "Dynamic Routing For Intercity Telephone Networks", published in the Proceedings of the Tenth International Teletraffic Congress, June, 1983 and authored by W. H. Cameron, J. Regnier, P. Galloy and A. M. Savoic. An extension to this basic SDD technique is described in another paper entitled "Multibour Dimensioning For A Dynamically Routed Network", published in the

Proceedings for the Eleventh International Teletraffic Congress, September, 1985 as authored by R. Huberman, S. Hurtubise, A. LeNir and T. Drwiega. Another example of a SDD scheme is described in a paper entitled "Use Of A Trunk Status Map For Real-Time

DNHR", also published in the Proceedings for the Eleventh International Teletraffic Congress, September, 1985 as authored by G. R. Ash.

To contrast SDD techniques with non-state dependent DNHR networks, the previous example is considered in the SDD context. It is recalled that during the first time period, the sequence {TG1 + TG3; TG2; TG4 + TG6} was generated. Now, rather than selecting the routes in sequential order, the route which may be selected from the set is the one that contains the largest number of free trunks at the initiation of the call attempt. For instance, if both TG2 and TG3 have, at the same instant, the fewest number of free trunks, then the route selected comprises TG4 and TG6. Other variations on the basic technique are possible. In essence, however, SDD only accounts for the past and present state of the network at the initiation of a call attempt.

All of the previously discussed techniques represent tractable solution to a very complex optimal control problem with a huge state space. Each solution is based on certain information about the network and certain approximations and, accordingly, departs from the optimal solution. As the solutions progress from PR to DNHR to SDD, more network information is being used. No routing solution, however, accounted for the future state of the network given the past and present states.

The first routing scheme in which the future effect of call-routing is explicitly considered is the state-dependent "separable" routing technique. This technique is the subject matter of U. S. patent No. 4,704,724 issued to T. J. Ott and me on November 3, 1987. As disclosed in this patent, a "cost" of adding the (j+ 1)st call to a trunk group when calls are already in progress is determined for each trunk group. This cost is a measure of the probable effect of admitting the present call on the blocking of future calls offered to the trunk group. The cost of a multi-link route is the sum of the costs of the trunk groups on the route. When a call arrives, the cost of each potential route is calculated for the current network state, and the call is either carried on the least-cost route, or it is rejected if the cost exceeds the cost of blocking the present call.

The version of the separable routing technique disclosed in our patent makes several approximations in its mathematical model to foster a tractable solution. In particular, the technique utilizes a procedure known as "policy iteration" in the theory of Markov Decision Processes. This name is associated with the following procedure: (i) begin with a "nominal" routing scheme (policy) for which it is feasible to determine the cost associated with different route selections in the network; and (ii) let these nominal costs serve as a guide to the actual selection of routes for each arriving call, thereby producing a new routing scheme or a new policy, which is the result of a "policy iteration" applied to the nominal scheme. The vcrsio'n of separable routing disclosed in our patent was derived from a nominal scheme in which a call blocked on a single route was treated as a lost call. Such a starting point provided excellent call routing capabilities.

In the version of separable routing disclosed in the above patent, the aforementioned costs are determined from the off-line solution of a large non-linear program which requires the prior knowledge of all the traffic demands on the network on a global basis. Since such information is in the nature of a forecast, the solution may be sensitive to forecast errors. Although our separable routing technique possesses a considerable degree of robustness to network load variations, it would be beneficial to eliminate the need for prior load information and depend, instead, on data collected in the course of normal traffic measurements in the network. Summary of the Invention

The limitations and shortcomings of these and other techniques, are obviated in accordance with the present invention, by a method that controls the selection of traffic routes through a network as determined by the network configuration and traffic measurements on the individual links forming the network. Each link of the network uses standard traffic measurement data, obtained at regular intervals for each individual link, to update its cost component. Thus, although the network routing scheme may be considered as being separable in the sense of the prior separable routing patent, the actual data used in computations originates from local traffic measurements rather than being derived from the solution of a non-linear problem requiring prior global knowledge. In this way the routing scheme is freed from reliance on prior load estimates of uncertain accuracy, which allows the routing to react better to shifts in load patterns.

Detailed Description

In describing an illustrative embodiment, it is helpful in conveying the principles of the present invention to consider the specific network depicted in FIG. 1.

Although this network is less complex than those encountered in practice, the small size of the network allows for brevity of description without loss of generality.

With reference to FIG. 1, network 100 comprises four nodes 101- 104 (N1-N4) interconnected by six links 201-206. In the network under consideration, each node is thought of as being interconnected to the other nodes via links. However, some links may not actually be implemented, as exemplified by link 202 (shown dashed) . In general, a network of N nodes is interconnected by (N)(N-1)/2 links. Besides the node-link configuration, network processor 150 is coupled to nodes 101-104 via channels 151-154 , respectively. Oftentimes, channel 151-154 are realized with dedicated, high-speed data lines.

In the telecommunications art, each node 101,102,103 or 104 may be realized as a control switching point comprising an electronic stored program controlled office such as the No. 4 ESS disclosed in the publication of The Bell System Technical Journal, Vol. 56, No. 7, September 1977, pp. 1015-1336. A portion of network

processor 150 and channels 151-154 may be provided by a Common Channel Interoffice Signaling system as disclosed in the publication of The Bell System Technical Journal,

Vol. 57, No. 2, February, 1978, pp. 221-447. Thus, these references may be consulted for a detailed understanding of control switching points as well as the signaling among these points. Such arrangements are in public use and, as will be described in what follows, such arrangements may be modified for embodying the principles of the present invention. Moreover, in the art of telephony, certain general terms like node, link and server (a link comprises a plurality of servers) are referred to by different terms. Thus, a node is a switching point, a link is a trunk group and a server is a trunk. These terms may be used interchangeably when telephony-related subject matter is discussed.

To describe succinctly the routing process of the present invention, a notation is introduced. The process is then described with reference to this notation and the parameters exhibited by FIG. 1. The network under consideration has N nodes and K= (N)(N-1)/2 links. Each link k, k= l to K, comprises s. servers where s_k≥O. Occupancy factors Δ(k,j), with 1≤k≤K, O≤ j:≤ s_k,

0≤ Δ(k,j) ≤1 and Δ(k,s_k) = 1, may be computed based on N,K,s_k (k= 1 to K), the time of day, the day of the week and traffic measurements. These factors have the following interpretation: each Δ(k,j) approximates the "cost", in terms of future call blockings, of increasing the number of busy servers in link k from j to j+ 1. Procedures for computing the factors are discussed later; it is to be noted that the routing process of the present invention differs from the separable routing scheme described in our U. S. Patent No. 4,704,724 by the manner in which the routing factors are obtained, and not in their use.

With respect to FIG. 1, N = 4 and K= 6. Link 1 interconnects nodes 101 and 102 and is identified with reference numeral 201; link 1 is also referred to as trunk group 1 (TG1). As depicted, s₁ = 2. Link 2, shown as interconnecting nodes 101 and 103, is not actually available to route a call since S2= 0. The remaining links may be described similarly.

A summary of complete link and server information for network 100 is given by the first and second columns, respectively, of TABLE I.

In addition, the next column shows the various states j, j= 0 to s_k, for each link and the last column lists the corresponding factors Δ(k,j) for each link and state.

The procedure also requires that a set of routes between each pair of nodes be available. Typically, these routes are generated at predetermined intervals and stored until updated. Generation is usually effected off-line with processor 150. For sake of example, routes of length one or two links in network 100 are considered. Then, for instance, nodes 101 and 102 have two possible interconnect routes, namely: link 201 (TG1) ; and the cascade of links 203 and 205 (TG3 and TG5). The set of routes between nodes 101 and 102, designated S₁, has elements {R11,R₁₂ } where R₁₁ = TG1 and R₁₂ = TG3 + TG5, with the plus operator designating series interconnection. In general, S. = {R. , m= 1 to M}, M being the number of one or two-link routes between the node pair associated with S_k . A summary of the one or two-link routes for network 100 is given by TABLE II (R(k,m) is equivalent to R_km).

When a call is offered to the network which requires routing from, say, node I to node J (I= N1,N2,N3 or N4, J= N1,N2,N3 or N4, I≠ J, the set associated with the node pair is accessed. For each route, designated R, in the accessed set, occupancy values V(R) are computed as follows:

where the sum is over the k lipks in route R, and X_k is the number of busy trunks in link k at the time of the call attempt.

As a final step, the route between nodes I and J exhibiting the minimum value is selected for further scrutiny. If the minimum value is less than a threshold, the call is routed over the route associated with the minimum value; otherwise, the call is blocked.

To elucidate the computations of occupancy values, the network of

FIG. 1 is utilized and, again for illustrative purposes, there is the restriction that routes of, at most, two links per permitted. At the instant a call is to be routed through the network, the state of the network is as summarized in TABLE III.

As exhibited by the middle column ("Busy") of TABLE III, all links except link 6 have all but one server busy; link 6 is idle. The appropriate factors from TABLE I for this instantaneous network state are repeated, for each link, in the third column of TABLE III. With the data shown in TABLE III and the route sets summarized in TABLE II, the occupancy values V(R) for each route in each set may now be computed. For instance, for the first set,

V(R₁₁) = Δζ1,X₁) = Δ(1,1) = 0.485, and (1)

V(R₁₂) = Δ(3,X₃) + Δ(5,X₅) = Δ(3,2)+ Δ(5,3) 061. (2)

By way of another example, the second route in the fifth set becomes

V(R₅₂) = Δ(1,X₁) + Δ(3,X₃) = Δ(1,1) + Δ(3,2) = 1.044. (3)

A summary of values V(R) for each route in each set is presented In TABLE IV.

If it is supposed that an incoming call is to be routed from N1 to N2, then the set of routes associated with this call is S₁ = {R₁₁,R₁₂}_. There is now a corresponding occupancy value set {0.485,1.061}. The route with the minimum occupancy value at this instant is R _{11 ,} and this is to be compared to a preselected threshold. This threshold is generally 1.0 for unnormalized Δ(k,j) factors. Since the minimum occupancy value, namely, 0.485, is less than the threshold, the call is routed over the route corresponding to this minimum occupancy value. Here, the call from N1 to N2 is routed over link 1 or TG1 since R₁₁ = TG1. (To illustrate blocking, if X₁=2, then the set would be {1.000,1.061}, and the call would be blocked).

By way of terminology, the route in a given set having the minimum occupancy value is called the candidate route in that set. Embedded in the notion of candidate route is the continually changing nature of the routes in response to busy-idle conditions. A set of routes is somewhat static, whereas a candidate route is dynamic. Of course, the set of routes may be updated, but this generally occurs off-line at a relatively slow rate compared to network dynamics. FIG. 2 is a flow diagram illustrative of the steps described above for controlling the routing of a request for service through a given network. Block 300 indicates that certain information about the network configuration, such as the number of nodes, the links, the servers, and traffic information, such as the time of day, day of the week and past and present traffic, is available for processing. From the configuration and traffic information, the sets of routes are generated, as indicated by block 310. These sets are valid for preselected time intervals, which may be measured in weeks or months. Block 320 indicates that, for predetermined time intervals, the Δ(k,j) factors are computed with reference to established and measured traffic. Upon a service request, the processing of block 330 is invoked. The state of the network at the initiation of the service request is employed to provide the Δ(k,X_k) factors and, in turn, the candidate route from the evaluation of the occupancy values for the corresponding routes. Finally, as depicted by block 340, traffic is routed on the candidate route between pairs of nodes, provided the minimum occupancy value is less than a preselected threshold. If the threshold is exceeded, the service request is blocked.

To this point in the description, the discussion has focussed on the methodology in an overview fashion, even though a specific network has been employed to exemplify basic principles. The focus of the description now shifts to allocation of the process steps among the various network components. Two cases are distinguished. Both cases, however, require that all nodes are comprised of stored program control type devices and that updates to the coefficients Δ(k,j) occur, generally in real-time but relatively infrequently, say every half hour. More specifically, the coefficients are updated every six holding times; since an "average" holding time is about five minutes, this yields updates every half hour. Case I

In the first case, only routes of length one or two links are permitted and all nodes are able to send each other messages via the network processor. The network of FIG. 1 may again be employed to depict this situation. Each node 101-104 maintains a table of factors Δ(k,j), but only for those links terminating on the given node. These factors are updated infrequently by network processor 150 and downloaded to nodes 101-104 via channels 151-154, respectively. Each node 101,102,103 or 104 has stored, for every link k terminating on that node, the current value of X. and therefore can provide the factors Δ(k,X_k ) for its portion of the routes.

When a call arrives at the originating node, say node I, node I effects a called number translation so that the destination node, say node J, may be identified. Node I sends a message to node J via processor 150 and the corresponding channels. The message identifies node I and reports the call attempt. The standard response from node J is a message which contains, for all routes between nodes I and J, the Δ(k,X_k ) factors of the links terminating on node J. (Δ(k,X_k) for one-link routes need not be sent since node I already has that information).

Since only one or two links are allowed in any route, node I now has all the information to complete the necessary additions, find the minimum occupancy value to determine the candidate route and then make a routing decision. If the call is accepted, it may be set up by any of the known signaling techniques utilizing processor 150 and channels 151-154.

In the earlier description, a call was to be routed from node 101 to node 102 given the network state summarized by TABLE III. To illustrate the methodology of this case, it is required that node 101 evaluate the occupancy values as depicted by equations 1 and 2. Node 101 has available information about links 201 and 203 since these links emanate from node 101; to complete the computations, node 101 requires the information with respect to link 205 which terminates on node 102. Upon receipt of a request message, node 102 responds with a message containing the factor Δ(5,X₅). The candidate route may now be computed and decision rendered accordingly.

The efficacy of this method lies in the fact that only one message transaction per call attempt is generated, and this transaction impacts only processor 150 and the originating and destination nodes. As such, the load on processor 150 is mitigated.

FIG. 3 is a flow diagram illustrative of the steps described with respect to Case 1 for controlling the routing of a service request through a given network. Block 400, which is essentially the same as block 300 of FIG. 2, indicates that certain information such as number of nodes, links and servers as well as time of day and week and past and present traffic is available to access and process, generally in network processor 150. Block 410 depicts that the generation of sets of routes at predefined intervals is effected in processor 150. As indicated by block 420, the sets associated with particular nodes are downloaded to the respective nodes, generally on an infrequent basis. Block 430 indicates that, within each node, the Δ(k,j) factors are computed for predetermined time periods in response to past and present traffic and measurements on each link k. Blocks 440 and 450 represent the message interchange steps between nodes I and J upon a request for service. The primary purpose of the message interchange is to reveal to node I the Δ(k,X_k ) factors that it does not have within its stored tables. Based on the revealed factors Δ(k,X_k ), the occupancy values of the routes may be evaluated to produce the candidate route, as block 460 depicts. Finally, block 470, which performs basically the same function as block 340 of FIG. 2, compares the candidate value to the threshold and handles the service request according to the result of the comparison. Many variations on the basic technique of Case 1 may be anticipated by those skilled in the art. For instance, the request message sent from node I to node J may also contain the information about the service request, e.g., the called party digits. In this manner, node J may check the outgoing telecommunication path to determine whether it is available. If it is not available, the response message may indicate this additional information and a busy tone may be returned to the calling party by node I.

Also, it may be advantageous for node I to reveal the Δ(k,X_k ) factors in the message it transmits to node J. In this approach, node J does the necessary evaluations and makes the routing decision. With this approach, the call set-up may be initiated by node J at the same time the response message is returned to node I, thereby decreasing call set-up delay.

Case 2

In this second case, routes of arbitrary length are allowed. The Δ(k,j) factors are updated infrequently (say every half hour) based on measured link data transmitted to processor 150, and the factors remain in network processor 150, that is, no downloading occurs. Frequently, say every ten seconds, nodes 101-104 report to processor 150 the status of all links terminating at the individual nodes. (In actuality, only links where the server activity changes need to be reported). Network processor 150 then computes, for every pair of nodes, the occupancy values of the sets of routes to arrive at the candidate routes. These routes with minimum occupancy values are stored and any node pair requiring an interconnection path may access the occupancy value information. Again, a route is selected only if the minimum occupancy value associated with a route is less than a threshold. The flow diagram of FIG. 2 represents the steps of this case. Basically, for this case, the computations are completed in network processor 150 in contrast to allocating some computations to the nodes, as in the former case.

Variations on this basic technique may also be appreciated by those skilled in the art. For instance, after the occupancy values are computed, those routes with values exceeding the threshold may be eliminated. Then, for each pair of nodes, processor 150 assigns a probability to each of the remaining routes, where lower-occupancy value routes are given higher probabilities. The probabilities arc downloaded to the respective nodes.

In the interval between computations of occupancy values, if a call request is made from node I to node J, node I will randomly select a route from the available list based on the assigned probabilities. A call set-up is initiated and if the chosen route is available, the call path is established. Otherwise, a new route is chosen randomly, using the same probabilities, but excluding the old route. This procedure continues until the list it exhausted.

Determination of Δ(k,j) Factors

Initially, it is assumed that in the near future (the next half hour) traffic on link k will behave as if it comes from an independent offered load which behaves like a Poisson process. Given this assumption, the Δ(k,j) factors ascribed to the behavior of link k are given by

where B is the Erlang-B function.

The Δ(k,j) factors arise from the following consideration: at time t=0, there are j busy servers and then one customer is added so that the number of busy servers increases to j+ 1; Δ(k,j) is the probability that at least one call blocking will occur during the life length of the customer just added.

Even for the situation wherein different links are dependent and "offered loads" are not Poisson nor clearly defined, equation (4) is still used as an approximation to cover this situation. The problem of what value to use for y_k in equation (4) remains to be addressed.

In our earlier patent on separable routing (U. S. Patent No.

4,704,724), it was proposed that the y_k 's be computed by the off-line solution of a large nonlinear program, which required prior global information about all the traffic demands on the network. Now, instead, the required y_k values are derived from the standard trunkgroup traffic measurements that are performed in the network.

To determine the appropriate y_k values, the following relationship is used: y_k = ɸ/(1-α), (5) where ɸ is the carried load on link k and α is the blocking (proportion of the time that all trunks in the group are busy) on link k, as measured over predetermined periods. The value of a is obtained from the same set of sampled-trunk-usage scans that are used to measure ɸ on a trunk group. The carried load is measured by scanning the state of the link or trunk group over regular periods and recording the number of busy trunks or servers observed; to obtain blocking, it is merely required to keep track of the number of occasions when the scan found all trunks or servers to be busy. With the capability of obtaining y_k from traffic measurements, the following procedure may be used to generate the Δ(k,j) factors and thereby route offered traffic:

(1) assume an initial set of equivalent offered loads y. for each link k for use in equation (4), and implement the separable state-dependent routing methodology

(one initial choice is y_k = s_k, the number of links in link k);

(2) over regular periods, of sufficient duration for the network to obtain statistical equilibrium (typically six holding times), measure link loads to obtain ɸ and α and compute a new y_k for each link k according to equation (5); and

(3) use the new y_k in equation (4) to determine new factor Δ(k,j), and implement the separable state-dependent routing methodology until new values of y_k for each link k are computed from measurements taken over the next period.

It is to be understood that the adaptive routing methodology illustrated herein is not limited to specific forms disclosed by way of example and illustration, but may assume other embodiments limited only by the scope of the appended claims.

Claims

What is claimed is:

1. A method for routing offered traffic through a network comprising nodes wherein each node is adapted to communicate with each of the other nodes, said method including the step of generating a set of routes between a pre-selected node pair wherein the routes comprise links interconnecting specific nodes, said method characterized by the step of

controlling the selection of the traffic route between the preselected node pair from the set of routes by regularly measuring link data locally and by generating admissible routes in the set as determined by the measured link data and the current usage of the routes in the set.

2. A method for routing offered traffic through a network composed of a plurality of nodes interconnected by links according to a preselected pattern, wherein each node is arranged to communicate with the other nodes, said method comprising the steps of

generating, for use at preselected intervals, sets of routes, each route composed of at least one link and each set including at least one route between each node pair,

over predetermined periods, measuring link data to estimate an equivalent offered load for each link and computing route occupancy factors based on said equivalent offered load,

upon a request for service between a particular node pair, computing occupancy values from said occupancy factors in relation to the current usage of the links comprising the routes in the set associated with the node pair, and

if the minimum of the occupancy value is less than a preselected threshold, routing the request over the route having the minimum value; otherwise, blocking the service request.

3. The method as recited in claim 2 wherein the equivalent offered load for each link is given by the ratio ɸ(1-α), where ɸ is the carried load and o is the blocking, as obtained from measured link data.

4. The method as recited in claim 3 wherein ɸ is equal to the average number of busy servers in each said link in a given period and α is equal to the proportion of time all servers in each said link are busy in the same given period.

5. The method as recited in claim 2 wherein the step of computing occupancy factors includes the step of equating the occupancy factors to the Δ(k,j) factors obtained from the relation

where B is the Erlang-B formula, s. is the number of potential servers in link k, 0<j<s. and y. is the equivalent offered load for each link k.

6. The method as recited in claim 5 wherein the step of computing occupancy values includes the steps of

evaluating for each of the routes in the associated set the quantities

where R is one route and Δ(k,X_k) is the occupancy factor evaluated such that j= X_k, X_k being the number of busy servers in link k at the initiation of the service request, and the summation is over all links k comprising each route R, and

equating said occupancy values for the routes in the set with the V(R) quantities.

7. The method as recited in claim 6 wherein y _k = ɸ/(1-α), where ɸ is the carried load and a is the blocking, as obtained from the measured link data for each link k.

8. The method as recited in claim 7 wherein ɸ is equal to the average number of busy servers in link k in a given period and α is equal to the proportion of time all servers in link k are busy in the same given period.

9. The method as recited in claim 2 further comprising the step of initially computing, before said step of measuring, the route occupancy factors using a presumed set of nominal link loads in order to service requests received prior to the first estimate of equivalent offered load.

10. The method as recited in claim 9 wherein the nominal link loads equal the number of servers in each of the corresponding links.