MXPA99010769A - Methods and apparatus for locating channels to design internet protocol networks congarantias de funcionamie - Google Patents

Abstract

Methods and apparatus are provided for designing IP networks with substantially improved performance compared to existing IP networks, such as, for example, those networks designed under the best effort criterion. Particularly, the invention includes methods and apparatus for: calculating the worst-case link capacity requirements, optimizing the topology of the network, and determining the location of the router within a network.

Description

METHODS AND APPLIANCES FOR LOCATING ENCAMINATORS TO DESIGN INTERNET PROTOCOL NETWORKS WITH OPERATING GUARANTEES Field of the Invention The invention relates to methods and apparatus for designing packet-based networks and, more particularly, to design IP (Internet Protocol) networks with performance guarantees.

BACKGROUND OF THE INVENTION Traditional IP networks are constructed with a planning or design optimization of very limited capacity. These networks can only provide a service with a better effort without guarantees of operation. However, customer expectations can only be met if IP networks are designed to provide predictable performance. In particular, network service providers have to guarantee the support bandwidth for their virtual private network (VPN) clients. In addition, included in any network design considerations, is the fact that there are several types of network chassis that can be used in a given network. For example, a packet switch such as REF .: 31753 Lucent PacketStarMR IP Switch (from Lucent Technologies, I?> Murray Hill, NJ) supports traffic scheduling and buffer management capabilities new features, including the flow waiting line with a perfectly weighted waiting line (WFQ) and longer waiting line drop (LQD), which allow a minimum bandwidth for VPNs while achieving a high level of resource utilization. It is also known that existing linked encapshers, on the other hand, do not support adequate flow isolation and their first-in-first-out (FIFO) programming, even when combined with the administration policy of the initial detection buffer random (RED), results in little control over the bandwidth that is shared between the VPNs and their performance is dictated to a large extent by the dynamic properties of the TCP (Transmission Control Protocol) which is the dominant transport protocol used in IP networks. Consequently, there is a need for a network design tool that allows users, ie, network designers, to design IP networks that have the same (homogeneous) or different (heterogeneous) types of routers that provide substantial performance guarantees. for a variety of applications such as, for example, VPN. Specifically, there is a need for a design tool which: automatically calculates the binding capacity requirements in the worst case and optimal on the basis of the specifications, of a designer; that optimizes the topology of the network; and that determines the optimal location of the router in the network.

Brief Description of the Invention The present invention provides methods and apparatus for designing IP networks with improved performance substantially compared to existing IP networks such as, for example, those networks designed under the best effort criterion. Particularly, the invention includes methods and apparatus for: calculating the requirements of bond capacity in the worst case and optimal; optimize the topology of the network; and determine the location of the router within the network. In a first aspect of the invention, the methods and apparatus for calculating the requirements to calculate the requirements of the link capacity of the links of the network are provided. Particularly, the upper and lower link capacity limits are calculable to provide the user with the design methodology with worst case and optimal results as a function of the different design parameters.

That is, given a network topology, specific IP demands and network delays, the design methodology of the invention allows a user to calculate the link capacity requirements for various network congestion scenarios., for example, events of formation of multiple bottlenecks across the network, for each link of the given network. In this design methodology, the user can design the IP network, given a specific topology, without the need to know where specific bottlenecks are located within the specific network. Also, the methods and apparatus for computing the link capacity of the invention handle the case where one or more conditions exist within a given demand. In a second aspect of the invention, methods and apparatus are provided to optimize the topology of the network associated with a network design. In particular, an optimal network topology was formulated according to the invention which aims to reduce the total costs of the network. In some modality, an iterative increase methodology was provided which aims to reduce network costs by packing small demands on the available capacity of some existing links rather than introducing additional links previously used in the topology of the network. In another modality, an iterative download methodology was provided which aims to reduce network costs by removing identified links which are slightly loaded to form an optimal network topology. In a third aspect of the invention, methods and apparatus are provided for determining the location of the WFQ / LQD routers to replace the FIFO / RED routers in an existing network so that the cost savings of the network are maligned. The methodology of the invention fulfills such a determination using a pure mixed programming model. These and other objects, features and advantages of the present invention will become apparent from the following detailed description of the illustrative embodiments thereof, which should be seen in connection with the accompanying drawings.

Brief Description of the Drawings FIGURE 1 is a block diagram of an IP network design system according to one embodiment of the present invention; FIGURE 2 is a flow chart of a design methodology according to an embodiment of the present invention; FIGURE 3 is a flowchart of a method for calculating link capacity requirements according to an embodiment of the present invention; FIGURE 4 is a flow chart of a method for calculating link capacity requirements according to another embodiment of the present invention; FIGURE 5 is a flow diagram of a method for calculating link capacity requirements according to yet another embodiment of the present invention; FIGURES 6A through 6D are flow diagrams of a method for optimizing a network topology according to an embodiment of the present invention; FIGURE 7 is a flow chart of a method for optimizing a network topology according to another embodiment of the present invention; FIGURE 8 is a flow chart of a method for determining the location of routers in a network according to an embodiment of the present invention; FIGURE 9 is a diagram of an exemplary network topology for implementing the present invention; FIGURES 10A through 10H are examples and tabular results associated with case studies performed with respect to the present invention; and FIGURE 11 is a graph illustrating the dynamics of the TCP congestion window.

Detailed Description of Preferred Modes The invention will be described below in the context of a VPN structure; however, it should be understood that the invention is not limited to such applications or system architectures. Rather, the teachings described here are applicable to any type of packet-based network, including any IP applications and system architectures. In addition, the term "processor" as used herein, is intended to include any processor device, including a CPU (Central Processing Unit) and associated memory. The term "memory" as used herein, is intended to include memory associated with a processor or CPU, such as a RAM, ROM, a fixed memory device (e.g., hard disk drive), a removable memory device (for example, a hard drive). example, computer floppy disk). In addition, the processing device may include one or more input devices, for example, a keyboard, for feeding data to the processing unit, as well as one or more output devices, for example, a CRT visualization device and / or printer to provide results associated with the processing unit. It should also be understood that the different elements associated with a processor can be shared with other processors. Accordingly, the instructions or code of the programs and programming systems for effecting the methodologies of the invention described herein, may be stored in one or more of the associated memory devices (ROM, fixed or removable memory) and, when ready to be used, be loaded into RAM and be executed by a CPU. Furthermore, it should be appreciated that, unless otherwise noted, the terms "node", "switch" and "router" as used herein are interchangeable. As mentioned, the optimal design of an IP network with guarantees of quality of service (QoS) has been a critical open research problem. In fact, with the commercialization of the Internet and the growing dependence on Internet businesses, IP networks are becoming a critical mission and the service with the best effort of current IP networks is no longer adequate. The present invention provides the methodologies for designing IP networks that provide bandwidth and other QoS guarantees for such networks, for example, VPN. For example, given the connectivity of the network and the traffic demand, the design procedures of the invention generate the network topology and the corresponding link capabilities that satisfy the demand when all the routers in the object network have WFQ / LQD capabilities. , such as the PacketStar IP Switch. The same can be done in a second design for the case of legal routers that use the FIFO / RED scheme. A comparison can then be made with respect to the savings calculated in terms of network costs. To achieve this comparison, models were used to determine TCP performance under FIFO programming with RED. In both of the previous cases, the routing restrictions imposed by the Shorter First Path (OSPF) routing protocol were taken into account. The present invention also addresses the problem of the location of the router to move a conventional router network to a guaranteed operational network via the replacement of legacy routers with WFQ / LQD routers. In particular, the methodologies of the invention identify the strategic places in the network where WFQ / LQD routers need to be introduced to produce the maximum savings in network costs. FIGURE 1 illustrates a block diagram of one embodiment of an IP network design system according to the present invention. The network design system IP 10, by itself, includes several interconnected functional processors, namely: a routing processor 12, a worst-case link capacity requirement processor 14, operatively coupled to the processor routing 12; an optimal link capacity design processor 16, operatively coupled to the link capacity requirements processor in the worst case 14; a processor for optimizing the topology of the network 18, operatively coupled to the processor of the link capacity requirements in the worst case 14; and a router replacement processor 20, operatively coupled to the optimal link capacity design processor 16. It should be appreciated that the functional processors 12 through 20, illustrated in FIGURE 1, can be implemented via the indicated processing devices ( for example, CPU and associated memories) collectively via one or more processing devices. That is, that the IP network design system 10 of the invention can be implemented via a single processing device or multiple processing devices. As input to the IP 10 design system and its associated methodologies, i.e., in the form of data signals stored or user-supplied to the design system to the processing devices, an initial main network topology is provided in the form of a graph G = (V, E) where V is the set of nodes that correspond to the points of presence (POP where the routers are located) and £ is the set of links that can be used to provide direct connectivity between the POPs. It should be appreciated that, as will be explained, an initial network topology must be provided by means of the optimization processor of the network topology 18. Also, given as an input to the system is the travel vector L = [L? , L? . . , L / E /] where L is the actual length of the link leE. A set of point-to-point IP traffic demands is also given as an input to the design system 10 where each IP flow demand i is specified by fl r given as a 6-tuple: fi = (sl f ti f to f ni f di, ri) where s and t are the origin and destination of the nodes in V, respectively ai. is the type of transport protocol (either TCP or RDP (Use Datagram Protocol)), nJ is the number of TCP or UDP connections within the flow, di is the aggregate minimum performance requirement for the flow assumed to be bidirectional, and r is the minimum between the speed of the access link of the client's site at the origin as? and the speed of the client's link access in the destination to tj. Let F be the set that contains all the f and; the subset of F whose elements are those demands which are routed through link 1 according to some routing algorithm R.

It should be appreciated that the selected routing algorithm J? is executed by the routing processor 12 (FIGURE 1) on the basis of such entries prior to this. The output of the routing processor 12, denoted by the reference designation A in FIGURE 1, is the routing information as a function of the demand flow and the topology of the network, that is, the flow (traffic) found on each link or pass through each link. The design system of the invention focuses on routing by the shortest path similar to that used in the standard OSPF protocol. The shortest paths are calculated on the basis of a given link metric 1 \? for each link 1 in £. Let I = [l? , l ?, ..., 1 / e /] the vector of those link metrics. It is assumed that the break of the link is used so that there is a unique route between any node of origin and destination. Let the capacity of link 1 Cl t be expressed in the unit of the capacity of the network of lines (for example DS3, OC3, etc.) with the assumption that a single value of the size of the network of lines was used (or capacity) through the network. Be C = [C, C2,. . . , C / e /] denotes the vector of the link capabilities.

In general, the network design system 10 and the associated methodologies deal, inter alia, with the following problem of capacity allocation: finding the vector of required capacity C, so that the performance demand requirements given by F are satisfied while minimizing the total cost of the network at the same time. Note that assigning a capacity of zero to a subset of links in E, the G topology must, in effect, be changed by reducing its connectivity and subsequently influencing the routing of the demands. Therefore, as will be discussed, the IP network design methodologies of the invention also include a topology optimization component. Referring to FIGURE 2, one embodiment of a general design algorithm 200 of the system proceeds as follows. First, the mixture of traffic Fi in such link is calculated (by means of the routing processor 12) on the basis of an initial network topology Gs (of optimization processor 18) which is a subgraph of G, the routing algorithm R, the link metric vector 7, and the set of IP requests F (step 202). Secondly, the capacity of each link required to satisfy the bandwidth demands in i (by means of the processors of the link capacity requirements 14 and 16) is calculated based on the types of routers in the network, the different assumptions about the congestion scenario, and in some cases the end-to-end delays of the TCP requests (step 204). Third, the design system determines whether the final network design was obtained (by means of the optimization processor 18) (step 206). If not, in step 208, the topology of the network is disrupted (by the optimization processor 18) and the new cost of the network is evaluated according to steps 202 and 204. This iteration of the design is then repeated until the final network design is obtained. The results of the final design are sent (step 210), for example, in the form of information presented to the user of the design system, including: (1) vector C; (2) the route of each traffic flow fl f- and (3) the corresponding costs of the network. One of the important characteristics of the methodologies of the invention is that both in the homogeneous case, ie for networks with only conventional FIFO / RED routers or those that use only WFQ / LQD headers, as in the heterogeneous case, that is, using A mixture of both types of routers can be accumulated in the design. Therefore, those methodologies of the present invention also serve as the central engine 16 certain of those sections are divided by themselves into subsections. 1. 0 Link Capacity Calculations Since the approach of the methodologies of the present invention is to guarantee the support bandwidth for guaranteed VPNs in IP, where the TCP is the main transport protocol in use, the present invention determines how much capacity it should have. have any given link in the network to guarantee a given set of demands, each associated with a group of TCP connections sent through this link. Typically, each group of connections belongs to a different VPN, and represents one of the point-to-point demands of the VPN. The answer to this question depends mainly on the programming of packages and the strategies of administration of the intermediate memory used in the routers of the network. The most popular on the Internet currently uses FIFO RED programming as the packet descent policy. Advanced next-generation routers such as the PacktStar IP Switch, however, use the WFP scheduler with a Longer Wait-Line Drop-Down (LQD) policy to provide bandwidth guarantees at the VPN level with cleaning and insulation properties at the flow (or connection) level. As will be evident, the conventional FIFO combined with the RED (FIFO / RED) typically can not provide bandwidth guarantees unless it is designed with a link capacity greater than the WFQ combined with LQD (WFQ / LQD). It should be appreciated that the link capacity processors 14 and 16 calculate the link capacity requirements given the demand for flow requirements and selected scheduling and buffering schemes. Later, design considerations under discussion will be discussed due to the use of the FIFO / RED scheme, as well as the methodologies implemented by link capacity design processors 14 and 16 to address those aspects of the design. Next, the design considerations for the WFQ / LQD scheme are discussed, the link capacity requirements of which can be calculated by any of the processors 14 and 16. Finally, the modalities of the different design methods are explained in detail. of capacity according to the invention. 1. 1 First On Enter-First On Exit with Random Initial Detection (FIFO / RED) Consider a set of TCP connections, which are routed through a link 1 with a bottleneck of capacity C? "F0. / RED and with the assumption that each TCP origin is a glutton and operates in a regime that avoids congestion, can be shown, as explained in Section 5.0 below, - based on the results of: S. Floyd , "Connections with Multiple Congested Gateways in Packet-Switched Networks Part 1: One-way Traffic," ACM Computer Comm. Review, Vol. 21, No. 5, pp. 30-47 (Oct. 1991); and M. Mathis, J. Semke, J. Mahdavi, and T. Ott, "The Macroscopic Behavior of the TCP Congestion Avoidance Algorithm," ACM Computer Comm, Review, Vol. 27, No. 3, pp. 67-82 (July 1977 ), that sharing the link capacity that any given connection € F \ obtained is given by: r, J = - = ^ - ITALc / 1 ™, Wi = (1) where tx and ha are the round trip delay (RTD) and the number of congested links in the path or congestion path i, respectively. A congested link refers to a link that has a sufficiently high load that its waiting line experiences packet loss, in contrast to a non-congested link, which has zero packet losses. In other words, the binding capacity is shared by the 19 TCP connections competing in proportion to their weight, which is given in equation (1) as the inverse of the product of the round trip delay t and the square root of h . Note the following fundamental difference between FIFO / RED and WFQ / LQD. In both cases, the distribution of the link of any connection is proportional to its weight. In the WFQ / LQD, as will be explained later, the weight of the connection is under the control of the network operator, and can be set arbitrarily. However, in the FIFO / RED, the weight is dictated by the characteristics of the trajectory or path of the connection, namely t and h, which to a large extent are not controllable. Each point-to-point demand i of a given VPN actually corresponds to a number of TCP connections, each of which has the same weight wl t since they all follow the same shorter path characterized by ti and hi. Let neither the number of TCP connections that constitute the demand i, i and Fi. It follows from equation (1) that the distribution of the link obtained by the demand i is now given by: '20 Assuming that the number of TCP connections is not within, a given demand i is proportional to the value of the real demand d', equation (2) becomes: w. FIFO r, = Wj (3) jeF W: = t, h (4) To have a link capacity with FIFO that satisfies the demands, we require that rt = di, Vi e Fí r which implies that the minimum link capacity capable of satisfying all the demands is given by: with w and t i e Fi, given by equation (4). The capacity of .FIFO link c, is obviously greater than or equal to the sum of all demands: 21 Also, if i * is the demand that reaches the maximum in equation (5), then combining equation (3) and equation (5) we obtain: which implies that r. = d. , r = di Vi e F. In other palacras, i * is the demand that has assigned its exact requirement and all other demands were assigned a bandwidth greater than or equal to its required value. The parameters involved in the calculation of the required link capacity according to equation (5) are dl t tA, and hl t i e Fi. Given the demands dlf the fixed part of the delays tj is determined. (propagation delays) of the shortest path and are expected to be dominant over a wide area (a delay component of the average wait line could also be added), the value of the third hi parameter is not trivial. 22 To treat the aspect of determining hi, we introduce some notations. Let i be the number of jumps that correspond to the shortest connection path i. Obviously, the number of congested jumps satisfy hi < x, Vi e F? . Let] be the vector representing the state of congestion of all the links in the network with I-¡being the indicator for the link j and equal 1 if the link j is congested and 0 in other circumstances. Let H =. { hj, h2,. . . , h \ F \] the vector of the number of congested links in the path or trajectory of each end-to-end demand, that is, i. -? í where P < i > J * P (?) Represents the sequence of links (paths) traveled by demand i. ^ [h, hl,. . . , h \ r? \] the vector of hi for those demands i e Fj. Vectors I and H have values in the sets / =. { 0.1 } lel and fí =. { 0, 1, ...,? X ... X { 0, 1, ...,? | R |} , respectively. Let g be the map line between I and H as defined above: V el, VHeH, H ** g (I) if and only if hi-? Ij, i = l, 2, ..., | F | JVC) 23 The set / is plotted under g to a subset H denoted by Hf and represents the set of elements feasible of H: Hf - (H and H: 31 and I s.t.H - g (I).}.

In other words, not every H and H is feasible.

. { H.}. denotes the set of Hi, 1 and E. An entry for hl that corresponds to the demand i appears in each H? that satisfies 1 e p (i), that is, all the links in the connection path i. If all entries for the same demand are equal, it is said that (H) is consistent. When (H) is consistent, the vector H with hj being a common input for the demand i in all Hl f 1 e p (i), is referred to as the common vector of. { H.}. . Finally, it is said that. { H.}. it is feasible if: (i) it is consistent; and (ii) its common vector is feasible. The calculation of the link capacity in equation (5) does not take into account the effect of multiple bottleneck in a wide network scenario where a demand i may not reach its distribution r in the J link if its distribution in another link along its path it is smaller. Therefore, the difference between r11 and the minimum distribution of the demand i in all the other links is its path or trajectory (which is greater than given that 1 * = di in all the links p (i)) can deduced from C? FIF0 without violating the demands, in other circumstances this extra capacity can be calculated by other valid demands that go through the link 1 which does not have its own requirements satisfied. In this sense, we consider ClFIF0 in equation (5) as a superior entity on the capacity of required link and we denote this one by c.FIF0 which is rewritten as: where we emphasize the dependence of weights Wj and in consequence of c / FIF0, on the value of H? . Also, taking into account the multiple bottleneck effect mentioned above we obtain a lower limit c / FIF0 on the required link capacity as fws. Based on / FIF0, we obtain the distribution of the demand i as: 25 r'i (Hx) * c / FIF0 (Hx) (%) and then we calculate the minimum distribution along the path or trajectory: n ( { B)) = min. { ri1 '(H1.), l'ep (i); r, (7) where the value of r- > di could be fixed in some value representing any potential limitation due for example to the speed of the VPN access link to the network. When the minimum in equation (7) corresponds to ^ * (Hi '), we say that demand i is obstructed by link 1 *, or that link 1 * is the bottleneck link for demand i. finally, you get c / FIF0 as: c.FIF0 ( { *.}.) "J ( { /.}.) (8)? * h < which is now a function not only of H but of H for all i 'ep (i), which is a subset of (H). The reason why C / GIGO is the lower limit is because for any feasible (H), there may be some 26 non-demand scenarios that result in the deviation of the bottleneck of some demands, which will then require a capacity greater than c / FIF0 to satisfy the requirements of the active demands. So far we have discussed the capabilities of upper and lower limit c "/ FIF0 and c / FIF0 as a function of Hx for each link 1 in the network. Since we do not know if in fact that Hi, the fwing limits were determined: c.FIF ° ur?) - ma? cT? F0 m) o- C / nro (ífln.} ^ ^ ^ pp. ( { R.}.) (1Q) tHf ~ where for each H and H f feasible, we form the Hx corresponding and we calculate c / FIF0 (Hx) and c.FIF0 ( { H.}.). He exact value of the required capacity C "F0 satisfies c / FIF0 (lf? n) = c1nro = / FIF0 (lf **). Advantageously, these limits do not provide upper and lower limits on the link capacities independent of the real value of H 6 H f. Although those limits are calculable, one more way practical, that is, easier to calculate, the set of limits is given as: ? FO (for) m ma? AF1F0 (Hx) (11) rtert C / nro (lfejor) = minClnro (¡J?) (] > 2) m Htinr where the maximum and minimum are taken on all values of H regardless of feasibility. It is evident that #peaker is obtained by choosing Hx for each i in the equation (6) with hi - h and h -, = 1 for j? i (each demand i e Fx has at least one link 1 as a congested link in its path or trajectory). Similarly, H? M: ior is obtained by taking Hx for each i in equation (6) with i - 1 and hj - hj and for j? i.

However, by definition, it can be noted that the fwing inequalities exist: c.FIF0 (H * e > OI) = / FIF0 (Ifin) < c / FIF0 (Hx) = c.FXF0 afax) < c.FIF0 (lfot); and c? FIF0 (lf? n) < ¿, F1F0 (H * 1 *) 28 where .FIF0 (H * in) is defined as equation (9) taking the minimum place of the maximum. Therefore, c.FIF0 (If * 0 ') It is used as an upper limit. c / FIF0 (H e: ior) is a good candidate for the lower limit since c / FIF0 (ífe: ioc) and C / FIFO (fam) SQn both less than c-FIF0 (H in) and in the case of the studies presented in Section 5.0, this proves that c / FIF0 (H * jor) = c / FIF0 ( { H.}.) for two typical values of. { H.}. which correspond to H equal to H3alto and H "n ° where hi =? 2 and hi = 1, i = 1, 2,..., | F |; respectively. H" po corresponds to the case where each demand has a single Congested link in your path or path, which may not be feasible. / fsaJt0 is feasible and corresponds to the case where all demands are active and large and each link contains at least one demand with a jump. The modalities of these methodologies will be explained later in Section 1.4. 1. 2 Waiting Time Perfectly Weighted with the Longest Waiting Line Decline (F Q / LQD) Next, the following discussion assumes that the designer who is using the system 10 of the invention has selected WFQ / LQD as the programming / buffering scheme. It is known that the PacketStar IP Switch supports waiting lines with a flow of 640000 per output link with three-level WFQ programmer, for example VP Kumar, TV Lakshman, and D. Stiliadis, "Beyond Best Effort: Architecture of Router for the Differentiated Services of the Internet of Tomorrow ", IEEE Comm. Magazine, Vol. 36, No. 5, pp. 152-164 (May 1998). At a higher level of the programmer's hierarchy, the link capacity is shared between the different VPNs, within each VPN it can be distributed based on the application classes (for example, TCP flows such as the Ftp, TCP flows such as the Telnet, UVP flows, etc.), and finally the bandwidth of each application class can be further divided among the flows belonging to that class (typically equal to the level of flow within each class, but it could be different). To make efficient use of buffer resources, the IP PacketStar switch uses the soft repartition of a set of shared buffers between all flows. A flow can reach its best link capacity weight (that is, get a link distribution proportional to its weight) if it can maintain an adequate pool of packets in the shared buffer. In other words, the WFQ programmer provides better opportunities for each flow to have access to the link for packet transmission but this may not translate into a better sharing of the link capacity if the flow is unable to sustain an appropriate packet reservation. due to inadequate control 5 of access to the shared buffer. This problem could occur for example when loss-sensitive traffic such as TCP is competing with loss insensitive traffic such as unhandled UDP, or when TCP connections with different round-trip delays (RTD) are competing for the ability to link. In the first scenario, since the TCP movement is insensitive to the loss of packets (TCP resources reduce their speed by contracting their window when a packet loss is detected), applications that do not adapt their speed according to the conditions Losses (either non-adaptive UDP resources or aggressive TCP resources that do not comply with the standard TCP behavior) are capable of capturing a poor allocation of the common buffer pool. The second scenario of TCP C connections with different RTDs is a well-known program, for example, S. Floyd and V. Jacobson, "On the Effects of the Traffic Phase on Compounds Commutated by Package", Interworking: Research and Experience, Vol 3, No. 3, pp. 115-156 (September 1992); T. V. Lakshman and U. Madhow, "Analysis of the Operation of Window-Based Flow Control 31 Using TCP / IP: The Effect of Delayed Products of Superior Bandwidth and Random Loss", IFIP Trans. High Perf. Networking, North Holland, pp. 135-150 (1994). The reason behind the imperfection of TCP connections with large RTDs is that the TCP constant window is increased by RTD during the congestion impeding phase that allows connections with smaller RTDs to increase its window faster, and when they create a reservation In some router in its path, this reserve grows rapidly for smaller RTD connections since it grows in an RTD packet. To solve these problems resulting from the complete distribution of the buffer memory, the PacketStar IP Switch uses the following buffer management strategy: each flow is assigned a nominal buffer space, which will always be guaranteed, and the occupation of the flow buffer is allowed to exceed its nominal allocation when there is space available in the buffer. This nominal assignment is ideally set proportional to the product of the bandwidth delay of the connection, but in the absence of delay information it could be set proportional to the weight of the connection in the WFQ programmer. When a packet above can not be accommodated in the buffer, some packet already in the intermediate memory 32 is pushed out. The flow waiting line from which the package is discarded is one with the largest excess beyond its nominal allocation (if the nominal assignments are equal, this is equivalent to a decrease in the longer waiting line). Since, as explained above, TCP flows with shorter RTD are likely to have longer waiting lines than their nominal allocation, the LQD policy alleviates imperfection at larger RTD connections. In addition, non-adaptable resources are likely to have longer waiting lines and will be penalized by the LQD policy. In this way, the flow isolation provided by the WFQ, combined with the protection and perfection provided by the LQD, allows each flow, class of flows, or end-to-end demand of the VPN, to obtain an allocation of capacity of link proportional to their assigned weight, for example see B. Sutter, TV Lakshman, D. Stiliadis, and AK Choudhury, "Design Considerations for Supporting TCP with Waiting Line by Flow", Proc. IEEE Infocom, pp. 299-306, San Francisco (March 1998). As a result, the weight of the programmer and the nominal buffer allocations are set on a given link 1, such that the link capacity 33 necessary to satisfy a set of VPN demands point-to-point di is simply equal to the sum of the demands, that is, where Fi is the set of all VPN demands that are routed through link 1 (that is, they have link 1 in their shortest path). However, compared to a WFQ / LQD capability router, a greater capacity is needed to satisfy the same demands when the router can only support FIFO / RED. 1. 3 Capacity Requirements due to TCP and UDP Up to this point, TCP traffic has only been taken into account to calculate the link capacity requirements given by the limits in equations (11) and (12) for the FIFO / RED case and equation (13) for the case WFQ / LQD. Due to the isolation provided by the flow waiting line, we only need to add the UDP demand to obtain the total capacity requirements for the WFQ / LQD case. We apply the same for the FIFO / RED case assuming that the aggregated UDP traffic probably does not exceed its demand. Therefore, the link capacity requirement for the WFD / LQD case is given by: and for the FIFO / RED case, the upper limit is: cFIFO. { Hpeor) = cFfFO (HmeJor) = ü¡riFO (Hmejor) + Y di UDP (16) ieF, the lower limit is: In addition, under special cases tf "1 * 0 and nunof the upper and lower limits are given by: C'F'FO. { Hsa "o) +? dl UDP cF¡F0. { Hsa "°) = (17) ieF, 35 u: FIFO. { one) where d denotes the UDP performance requirement for the demand i, and the function of the upper limit. was introduced to take into account that the link capacity is discrete (in units of line network capacity). 1. Modalities of Calculation of the Linking Capacity Given the equations derived above, the following are several modalities of the methodologies of the invention for calculating the binding capacity requirements relevant to the particular design criteria selected from the users of the network design system. of the invention. As will be indicated, the methodologies are effected by the processor of 36 link capacity requirements in the worst case 14 and / or the optimal link capacity design processor 16 (FIGURE 1). Referring to FIGURE 3, a method 300 is shown to calculate the requirements for link capacity in the worst case based on a FIFO / RED according to the invention. It should be understood that the notation c denotes the link capacity taking into account only the TCP traffic, while C denotes the link capacity taking into account both TCP and UDP traffic. Consequently, as is evident from the terms in the equations, the first addition term is the link capacity requirement for the TDP traffic and the second addition term is the link capacity requirement for the UDP traffic. Furthermore, it should be appreciated that such calculation is performed by the link capacity requirements processor in the worst case 14 (FIGURE 1) on the basis of input from the routing processor 12 and the user. Accordingly, such a design methodology provides the user with the system 10 with a calculation, based on particular input specifications, of the link capacity requirements on the basis of link by link. First, in step 302, the processor 14 receives the input parameters of the routing processor 1 ^ and the user. The inputs of the processor 12 include the set of VPN demands point-to-point, di, the round trip delay TÍ, associated with the connection i. Of course, those entries are initially specified by the user. In addition, the user specifies the programming / buffering or placement scheme in the buffer memory, which, in this case is FIFO / RED, and the congestion option H ° (for example, Ifor. H ** 1 and H "no It should be understood, as explained above, that the congestion option is an assignment of some value? For the given design criteria.,? Refers to the number of congested links in the path or path of connection i. Referring again to section 1.2, tfeor is obtained by choosing Hx for each i in equation (6) with hi = h3 - 1 for j? i (each demand i e F has at least one J link as a congested link in its path or trajectory). Recall that tf 01 is the upper limit defined in equation (15) and, together with the lower limit ¡f * z (calculated by the optimal link capacity processor 16) apply to all values of H. In addition, } pfl and H "n0 as defined in step 303, which are special cases of the upper limit lf * or, correspond to? j = í, and hi = 1, i = 1, 2, ..., | F |, respectively, H "does not correspond to the case where each demand has a single congested link in its path, which may not be feasible. rf >; *? to is feasible, and corresponds to the case where all the demands are active and avid, and each link contains at least one demand for a jump. It should be understood that according to the invention, the user need only specify the option H °, since the values of h which correspond to this are preferably stored in the memory associated with the processor 14. Next, in step 304, depending on the option chosen by the user, the link capacity requirements are calculated in the worst case. That is, the link capacity for each link in the current network topology is calculated based on the demands, the delay, the scheduling / buffering or buffering scheme and the selected congestion option. It should be noted that the equations for} f * oz, "1 0 and H" n ° shown in step 304 of FIGURE 3, are respectively the same as equations (15), (17) and (19) above, with the right side of the equation (6) inserted as the first term of the upper limit function. Finally, the link capacity requirements (denoted by the reference designation B in FIGURE 1) for each link of the current topology are sent to the user via, for example, a visual representation associated with the processor 14.

Referring now to FIGURE 4, a method 400 for activating the optimal link capability based on a FIFO / RED according to the invention is shown. Again, as is evident from the terms in the equations, the first addition term is the link capacity requirement for the TDP traffic and the second addition term is the link capacity requirement for the UDP traffic.

Furthermore, it should be appreciated that such calculation is effected by the optimal link capacity design processor 14 (FIGURE 1) based on the input of the routing processor 12, the user and the link capacity requirements processor in the worst case. the cases 14 (ie, since, the rx distribution is a function c,). Accordingly, such a design methodology allows the user of the system 10 to calculate, based on the particular input specifications, the link capacity requirements on a link-by-link basis. First, in step 402, the processor 16 receives similar inputs as the processor 14, i.e., the topography of the network, the origin-destination demands, the round trip delay and the selection of the congestion scenario made by the user. Also, the link capacity requirements calculated by the processor 14 are provided to the processor 16. Again, it should be understood that according to the invention, the user need only specify the option H ° (eg, H ** ior, H). "lto, H" no), since the values of hi corresponding to it are preferably stored in the memory associated with the processor 16. As mentioned above, tf * ei ° r is obtained by taking Hx for each i in the equation (6) with h¿ = 1 and h-, = h j and for j? i. Also, tf * lto and # ° not as defined in step 404 correspond to h¿ =? j and h¿ = 1, i = 1, 2,. . . , I FI, respectively. Then, depending on the option chosen by the user, the optimal link capacity requirements are calculated. That is, the link capacity with each link in the current network topology is calculated based on the demands, the delay, the scheduling / buffering scheme of buffering and the selected congestion option. It should be noted that the equations for H * e:, oc, tf "ito and.}" Not shown in step 404 of FIGURE 4, are respectively the same as equations (16), (18), and (20) ) above, with the right side of equation (6) inserted as the first term in the function of the upper limit in equation (16) and the right side of equation (8) inserted in the first term in the limit function superior in equations (18) and (20) Finally, the 41 link capacity requirements (denoted as the reference designation D in FIGURE 1) for each link of the current topology are sent to the user via, for example, The visual representation It should be understood that an input device and an output device are used for all the inputs and outputs selected by a user associated with the system 10 of the invention.Thus, the following point should be noted with respect to the output of the processor 16 compared to n the output of the processor 14: the network topology remains unchanged, however, the link capacity of some of the links can be reduced due to the consideration of the multiple bottleneck effect across the network. Referring now to FIGURE 5, a method 500 for calculating the input link capacity in a WFQ / LQD C? HF0 according to the invention is shown. This mode assumes that the user has selected the programming / buffering scheme or placement in the WFQ / LQD buffer memory for its design. Therefore, it should be appreciated that calculations based on congestion limits and options are not necessary and, as a result, only VPN demands are required for each J link with respect to TCP / UDP traffic as input, in step 502, to calculate the link capacity requirements. It should also be appreciated that since it is not necessary to calculate the upper and lower limits in the WFQ / LQD scheme, either the processor 14 or the processor 16 can be used to calculate such link capacity requirements. Thus, in step 504, the programmer weights and the final allocations of the buffer memory are set to a given link 1 such that the link capacity that a set of VPN demands needs to be met point-to-point is simply equal to the sum of the demands, as for the equation shown in step 506 (which is the same as equation (14)), where Fx is the set of all the VPN demands that are routed through the link 1 ( that is, they have link 1 in their shortest way). In step 508, the link capacity is sent to the user (for example, via a visual representation). 2. 0 Optimization of the Network Topology Remember that for the problem of allocation of capacity under consideration, we have the flexibility to eliminate some of the links in the topology of the original network G by making capacity assignments of zero. The motivation to remove links is to rid some of the network facilities previously used to reduce the total cost of the network. Despite the total process design 43 of the invention, the cost of the network is calculated based on the following function: l-rl J (M (C1, Ll) = T (C?)) (21) where M (.,.) and T (.) are the path and termination cost functions, respectively. It should be noted that this cost function was selected for its simplicity and ease of illustration. Other forms of more general cost functions with the present invention can be easily incorporated. It should be appreciated that the optimization processor of the network topology 18 preferably calculates the cost of the network. However, processor 14 or 16 could do the same. In the following subsections, we will consider two modalities and their variants for the optimization of the topology of the network. This is a method of increasing or improving the link and the link download method. The network topology optimization process is also performed by the optimization processor of the network topology 18. Also, the topology of the network is initially provided by the routing processor 12 to be used by the system 10 which can be provided by the topology optimization processor 44. of the network 18, or alternatively, by the user of the system 10. .1 Enhancement of the Increase Referring to FIGS. 6A through 6D, a method 600 is shown to optimize the topology of the network using the increase method according to the invention. In the augmentation method, starting with an appropriate subgroup Gs of G and we increase this with additional links and / or capacities until all the demands can be routed. Initially, a subset of arcs in G is selected to form G3. The manner of forming G5 is as follows: First, in step 602, all end-to-end demand flows i are divided into two sets, namely, guardians and defectors based on their minimum throughput demand d2. In one modality, the criterion is to make the demands that require at least one unit of the bandwidth of the network of lines with guardians, while the rest, that is, those demands with fractional line size, become defectors. In step 604, guardian demands are then routed over the complete G graph according to the routing algorithm of choice such as, for example, the routing of the shortest path, which is the algorithm used by the routing protocol. OSPF. Once the route for the gatekeeper is calculated, the necessary capabilities are provided along the route in line network size units to accommodate the guardian's demand. Due to the discrete nature of the network size of lines, it is likely to have available capabilities along the guardian routes. After routing all the demands of the guardian, the links in G have been used to convey the demands of the guardian form the initial subgraph Gs (step 606). Note that Gs can not provide connectivity between the source and destination nodes of some of the defectors. In this way, the present invention provides the designer with an option to increase the connectivity of the defector, in step 608 to provide connectivity in all pairs of source-destination nodes. If the option is selected, all defectors are placed on a work list Ll, in step 610. If the list is not empty (step 612), the defector is removed from Ll, in step 614. The chosen defector is know as f,,. Then, in step 616, it is determined if there is a path or path in Ga between the source node and the destination node of τ t without considering the capacity restrictions on the links. If there is such a path, then it is removed from the work list Ll in step 618. If not, then in step 620, the defector fi is converted into a guardian, and then it is routed 46 along the way more short in G from its origin to its destination node, adding the required capacity along its path or trajectory. For those links along their path that are not included in the current Ga, those links are added over Gs to form the new Gs. Then, fi of the work list Ll is removed in step 618. After f is removed from Ll, the list is checked to see if some other dropouts were left (step 612). If they exist, then steps 614 through 620 are repeated. If not, the processor proceeds to step 622. Recall that in step 608, the option of selecting the connectivity of the defector to the user of the design system 10 is given. select the deserter connectivity option or the deserter connectivity option was selected and completed, the following procedure is performed. All dropouts are placed on a work list L2 in step 622. As the process progresses, the work list L2 is checked iteratively to see if there are any more dropouts left (step 624). In step 626, a defector, fj, is removed from L2 (step 626). In step 628, f3 is routed along the shortest path between its origin and destination nodes in Gß. Next, in step 630, the method includes determining whether there is adequate connectivity and capacity along this shorter path to accommodate fj. If so, remove fj from list L2 (step 632) and check the list again to see if there are any remaining dropouts in the list (step 624), so that the process can be repeated. However, if there is no adequate connectivity and capacity to accommodate fj along this shorter path, then the designer can choose between two alternative augmentation methods, in step 636. A method is referred to as an augmentation method only. capacity and the other is known as a method of increasing connectivity plus capacity. In the increase of only the capacity, the method is to conserve the initial Gs without change from now on. If the defector can not be accommodated in Gs, additional capabilities are added along their route (step 638). One advantage of this method is computational efficiency because G9 remains unchanged after the initial phase. Therefore, the routes of defectors are not influenced by the increase in capacity subsequent to G3. Note that this is not the case for the other methods, where increased connectivity may take place after some of the defectors have been routed. After step 638, f3 is removed from the list (step 632) and the process is repeated for the remaining defectors in L2.

An alternative enhancement strategy of the invention provides that when a defector can not be routed over Gs due to the lack of available connectivity or capacity in Gs (the latter case should not occur if the complete optional connectivity procedure has been performed (steps 610 to 620) for Gs, the additional defectors are converted to guardians to enrich the specificities and connectivity available from G3.The conversion from deserter to guardian can be accomplished via one of the following two methods.The designer can choose the method, in step 640. A first method is known as threshold controlled defector conversion, the method is to convert some demands of the defector to guardian demands by lowering the threshold between the demand values of the guardians and the defectors (step 642). This threshold is initially set to a unit network line size.The newly converted guardians are then routed, at step 644 over its shorter paths in the full topology G. Links are added with the necessary capabilities assigned. Any newly activated link is then used to increase the current Gs and form a new Gs. Although capacity and connectivity can be added when the threshold is lowered, newly added resources may not directly address the need of the target deserter to be sent. Also, since there may be change of connectivity in Gs, the shorter paths or trajectories of the defectors routed above, may be altered in the new Gs (but not the guardians, since they are not routed over G). As a result, it is desirable to undo the routing of all already routed defectors (step 648) and then return to step 622 to reroute the defectors. This process of lowering the threshold, converting the defector to guardian and increasing G5 is repeated (step 624) until all defectors can be routed over Gs. The resulting Gs is then converted to the final network topology. It is evident because the option to complete the connectivity is made (step 608) to form the initial Gs, that is, without this option, there is a small demand for a defector which has no connectivity in Gs, the threshold can continue to decrease until the small demand of the defector is converted to a guardian. This can be very wasteful from the perspective of accumulation of capacity, as well as computational inefficiency. The first one refers to the introduction of unnecessary available capacities in the wrong places of the network, and the last one refers to the fact that the deserters are routed many times, that is, 50. as long as the threshold is decreased and the connectivity of Ga changes. An alternative lag conversion method is known as direct lag conversion, where a lag is directly converted to a common demand when it can not be routed over the current Gs (step 646). Again, the converted lag is then routed over G while adding the links (if necessary) and extra capabilities to increase the Gs. Due to possible changes in the shorter paths or paths after the increase in Gs, all formerly lagging routers have to be scrapped, in step 648, and then rerouted (step 622) as in the case of the controlled conversion by the threshold . Then, regardless of the selected conversion option, once all the lags are re-routed and no further lags are present in the work list L2, the network optimization process is completed (block 634) thus producing the topology of the final network. 2. 2 Link Download Optimization In the link download mode, the method is to start with the full topology G and then try to improve the new network cost by removing some slightly loaded links to produce the final topology. Due to the use of the shortest single path routing, all the networks of lines in a link are removed to change the routing pattern in the network. Referring now to FIGURE 7, there is shown a method for downloading links according to the invention. First, the candidate links to be downloaded are identified, in step 702, on the basis of the traffic flow they can support and a Thddesarga tuneable usage threshold. Specifically, in the case of FIFO / RED router networks, a link 1 is a candidate to be downloaded if (c.FIF0 ( { H.}.) +? d,? DP) < (Thddeaarga * capacity of? * Fl network of unit lines For links in a WFQ / LQD router network, the corresponding criterion is? D? DP) = imFI (Thdaeaargß * network capacity of unit lines). Once the candidate links are selected, they are sorted according to their potential impact on the existing network design when the traffic transported by them is rerouted (step 702). For this purpose, provided that the candidate list is not empty (step 704), the sum of the product of the demand and the exact control of flows through each candidate link is calculated in step 708.

Next, in step 710, the candidate link with the smallest sum of the product is removed tentatively from the topology of the network. The motivation is to minimize disruption of the topology / capacity during download to avoid rapid changes in the cost of the network. After a candidate link is tentatively removed, the new routes, capacity requirements and the cost of the resulting network are recalculated in step 712. If the removal of the link results in a reduction in the cost of the network, the link is removed permanently in step 716. However, if the removal of the network does not result in a reduction in the cost of the network, this candidate link which is retained in the topology but removed from the candidate list (step 718). The download process is then repeated (with the topology updated if the previous candidate or the same topology was removed if the candidate was retained) for the next candidate link with the smallest area of the product in the list (step 704). If the candidate list is empty, the download process ends (block 706). It should also be noted that several heuristic optimization variants of the topology discussed in Section 2.0 have been implemented and tested. For example, we have tried different orderings in which defectors are routed, as well as the combined use of the increase method 53 with the link download method. That is, it should be appreciated that the link download method can be used as a stand-alone optimization method or can be used in conjunction with the augmentation method, for example, the link download method can follow the augmentation method. The resulting functioning is presented in Section 4.0 where studies of different cases are discussed. In the augmentation method for topology optimization, the cost of network configurations generated during intermediate iterations is not explicitly considered. However, the minimization of the cost of the network is implicitly carried out via the traffic packaging and the increase of the topology. This is based on the observation that the cost of the network can be reduced by packing small demands on the available capacity of some existing links instead of introducing additional poorly used links. It should be appreciated that when an additional capacity is needed over a link to accommodate a new defector in the augmentation method, the actual required capacity can be calculated in a simple, direct additive form for the case of the WFQ / LQD routers. However, in the case of FIFO / RED router networks, the unloading method is preferred because the capacity requirement of 54. The link as shown in Section 1.0 can vary considerably when the mix of traffic in a link changes. Also, it should be appreciated that, with the routing of the demand based on the shortest path, there is a subtle difference between a link that has no available capacity and one with a total capacity of zero. The routes of the demands in these two cases can be significantly different and need to be distinguished. Consequently, given the implementation of any of the illustrative modalities of optimization of the topology of the network explained above, the output of the optimization processor 18 (denoted with the reference designation C in FIGURE 1) is a network topology that is preferably provided to the processor of link capacity requirements in the worst case 14, through the processor of routing 12, which then calculates the link capacity requirements given the received topology, as explained above. Next, as explained in the context of FIGURE 2, it is determined whether this final topology meets the criteria of the designers, for example, due to the cost of the network or validation considerations. If not, the optimization process (whichever has been implemented) is repeated until the final network topology is formulated. 3. 0 Replacement of the Router Assume an existing IP network which uses only the linked FIFO / RED routers. Let the network be represented by an indirect simple graph Gs = (V, E ') where V is the set of nodes that correspond to the set of routers and E' is the set of links that connect the routes. Here, we consider the following problem P. Given a maximum of WFQ / LQD Nmax routers that can be used for the one-to-one replacement of any FIFO / RED routers in V, find the set of FIFO / RED routers to be replaced, so that the savings in the cost of the network are maximized. Let TFIF0 (C) and T * 1 (C) be the termination cost for a FIFO / RED router and the WFQ / LQD router, respectively, to terminate a capacity link C. Let M (C, L) be the cost of the trip of a link of capacity C and length L, regardless of what types of routers are used. By replacing some of the FIFO / RED routers in the existing network with WFQ / LQD routers, the resulting changes in the total cost of the network can be divided into two separate components. First there are the expenses related to the upgrade of the selected FIFO / RED routers to the WFQ / LQD routers. Second, there are the cost savings derived from the reduction in transmission capacity requirements when the administration of the FIFO / RED programming and buffer in the legacy routers are replaced with WFQ / LQD in the next generation advanced routers. To understand the details of the savings / expenses involved in the replacement process, the present invention provides the following two steps. First, we perform a one-by-one replacement of a selected set of FIFO / RED routers using WFQ / LQD routers which have the same number of interfaces and terminating capacity as the routers they replace. Denote such replacement cost by a FIFO / RED router selected by Ci. Second, for a transmission link 1 = (i, j) connecting the FIFO / RED routers i and j, if both i and j are replaced by WFQ / LQD routers, the link capacity requirement i is reduced due to the administration of Packet programming and improved router memory. As a result, cost savings can be derived from (i) obtaining a "reimbursement" of the interfaces / extra termination capacity over the newly replaced WFQ / LQD routers and (ii) the reduction of the travel cost associated with link 1. 57 Specifically, if and only if we replace both of the terminal routers i and the 1 = (i, j) link with WFQ / LQD routers, we can realize the savings of YES, J given by: Yes tj =. { M (C? FIF0, Lx + f (Cf8J -MfCj ™ 0, L?) -T »(Cx).}. (22) where CjFrF0 and C? WFQ are the corresponding capacity requirements for the FIFO / RED cases and WFQ / LQD Note that Sj.- is a conservative estimate of the actual savings derived from such replacement because it only considers the impact of the capacity requirement on a link-by-link basis, it is possible that capacity reduction may be achieved extra somewhere in the network when the WFQ / LQD router is added due to its stricter bandwidth control over the flows that pass through the J link. Advantageously, based on the structure described above, the problem P can be formulated, according to the invention, as the following intermingled programming problem (MIP): maximize (OJ subj ect T ^ x = N "ax, Iti" 58 Vi, je V, O = yirj = i, (a) Vi, je VJ O = yirj = xjr (b) Vi e VJ i = 0 or 1 (c ) where £? Is the cost updated by router i as described above. YES, J is the saving of costs as defined in equation (22), understanding that Si, j - 0 if (i,) - £ E 'or i-j. i is a binary decision variable such that x¿ = 1 if and only if the router i is selected to be replaced by the router WFQ / LQD. yíf3 is a dependent variable that reflects the realization of cost savings associated with the link 1 = (i, j): according to the restrictions specified by restrictions (a) and (b), ylrJ may be different from zero only - c if both a «1 and x¡- = 1. Otherwise, yitj = 0. This corresponds to the fact that cost savings can only be realized when both ends of a link are connected to a WFQ / LQD router . Note that there is no need to specify yi fj as a binary variable because with Si 7 = 0, the maximization of the objective function will automatically force yl? 3 to become 1 if it is allowed by the values of x ^ yx¿ in based on restrictions (a) and (b). Otherwise, yitj will be forced to 0 by the restrictions in any way. Nmax is an input parameter that specifies the maximum number of FIFO / RED routers that can be replaced. If Naíx is set to I V I, the solution to this MIP problem will determine both the optimal number of routers to be replaced as well as the corresponding replacement locations. On the basis of the above MIP formulation, the router replacement processor 20 implements a program of the optimal router replacement programming and programs using standard MIP optimization packages. For example, such standard MIP optimization packages that can be used include: AMPL, as is known in the art and described in R. Fourer, Gay DM and BW Kernighan, "AMPL - A Modeling Language for Mathematical Programming" Boyd &; Fraser Publishing Company (1993); and CPLEX Intermediate Resolution Device from CPLEX division of ILOG, Inc. When running on a Pentium II 333 MHZ PC, the optimal location of WFQ / LQD routers in a large linked FIFO / RED network with approximately 100 nodes and 300 Links can be determined within seconds. .0 Case Studies In this section, the results of some case studies (examples) are discussed with respect to the allocation of IP capacity and the replacement of the optimal router according to the invention. The first case study was based on the topology of the NSFNET at the end of 1994. FIGURE 9 shows this topology, which can be used as the complete topology G in the design structure of the invention. The size of the line network was set at 240 units through the network. FIGURE 10A gives the matrix whose inputs are the corresponding point-to-point traffic demand used for the example. The relative magnitudes of the traffic demands were fixed according to the 1994 statistics reported in Merit Network Information Center Services, "Statistical Reports Belonging to the NSFNET Main Networks", (1994), using the scaling method proposed in RH Hwang, "Routing in High Speed Networks", PhD. Dissertation, University of Massachusetts at Amherst (May 1993). We also scale the absolute volume of each demand to reflect the growth of demand since 1994. The total cost of network J was calculated based on some arbitrary termination cost per line network and some arbitrary cost per unit length. It was assumed that the termination and travel costs are the same for the FIFO / RED and WFQ / LQD routers. While designing the network, we have experimented with several topology optimization strategies and different network congestion assumptions described in the previous sections. 4. 1 Design with Homogeneous Routers First, we consider the case of homogeneous networks, which use FIFO / RED routers or WFQ / LQD routers exclusively. We refer to those as the cases all FIFO / RED and all WFQ / LQD, respectively. For our design problem, the total cost of the network is governed by two key factors, namely: (1) the final topology of the network as a result of the heuristic of network utilization; and (2) the capacity requirements of the links, which are a function of the programming and buffer management capabilities available in the routers. We will discuss the impact of those two factors separately in the following subsections. 4. 1.1 Impact of the Router Capabilities To focus on the impact of the router capabilities alone, we made use of the same final topology for all design cases. This was achieved by activating the topology optimization module and we used the initial structure G as the final one, that is, final Gs = G. As a result, the designs all WFQ / LQD and all 62 FIFO / RED had both a topology identical to the shown in FIGURE 9, where all the links are active. FIGURE 10B summarizes all corresponding results. Note that with the same final network topology (and thus the traffic route), the cost difference between all cases WFQ / LQD and all FIFO / RED is solely due to capacity savings derived from programming and administration capabilities of the advanced buffer of the WFQ / LQD routers. Although the cost of the case all WFQ / LQD does not vary with respect to the different scenarios of network congestion, the cost of the configuration all FIFO / RED varies depending on the assumptions of the network congestion scenario. The configuration of all WFQ / LQD costs less than 1/3 of all FIFO / RED under the scenario of congestion in the worst case, that is, based on ClF 'or (Hnmr). The cost is still considerably lower even when the scenario of optimal congestion based on Ha jor is assumed. Note that there are significant cost differences for the FIFO case when different Uf congestion scenarios are assumed, "H" 110, H "" 0,? E? ° c). however, the cost difference due to the multiple bottleneck effect (between C, "H) (H m) and C" FO (Hfnr) as well as between -.

ClF '(Hμn ") and C (fíwm)) is relatively small when compare with the cost difference due to the choice of H. 63 As mentioned above, since the congestion scenario is dynamic and unpredictable in practice, if you want to provide minimum performance guarantees under all traffic scenarios, you have few choices, but you assume the worst-case scenario of the cases, that is, rf »* 01" for the calculation of the FIFO / RED link capacity. "FIGURE 10B also includes a column called the" Network Width Reaccumulation Factor "denoted by K. Given the topology of final network Ga (V, E ') and the associated link capabilities C, to satisfy the minimum performance requirements, dl f is defined as: To motivate the definition of K, let us consider an individual link 1 and the corresponding relationship C / YV. Ideally, if the capacity of link is available in continuous units as opposed to the discrete steps of the network size of lines, and if bandwidth programming and administration of the intermediate unit of the ideal link is available, it will be sufficient to have CJ / YÍ /, equal to one to satisfy the minimum performance requirements of all demands. When the same argument is applied to all links in the network, it is clear that the ideal (minimum) value of K also is equal to one. Thus, K is a measure of "overaccumulation" due to non-ideal situations, such as the discrete nature of the link capacity (ie, the need to round to the nearest whole number of line networks) and the absence of a 10-band wide sophistication bandwidth and sophisticated buffering in the routers. From FIGURE 10B, it can be seen that K is slightly greater than one for the case WFQ / LQD, which is simply due to the discrete nature of the capacity of i c, link. On the other hand, a much greater capacity for overaccumulation is required for FIFO / RED cases due to the absence of adequate traffic management capabilities in FIFO / RED routers, that is, an excess of link capacity is needed to overcome the inherent inaccuracy of the TCP to meet the minimum performance requirements. In addition to the design of the NSFNET structure, we also conducted a similar study on the design of a large-scale carrier-class network. The results are summarized in FIGURE 10C. The findings are qualitatively similar to those of the NSFNET study, except that the relative cost difference between all FIFO / RED and all WFQ / LQD configurations becomes even greater. This is due to the increase in the size and diversity of traffic on the network when compared to NSFNET. Recall from equation (5) that with FIFO / RED routers, the capacity requirement of a link is dominated by the demand that has the maximum relation (/ (w./Yw). come back J * F, the end-to-end delays of traffic demands, and thus the maximum ratio dl / (wl / 'YwJ). J * F? 4. 1.2 Comparison of Topology Optimization Heuristics We will now proceed to compare the effectiveness of several topology optimization heuristics discussed in Section 2.0. Here, we use the example of the NSFNET structure for illustration. FIGURE 100 reports the results of several optimization options using WFQ / LQD routers exclusively. The cases of the FIFO / RED router were not included because it is preferred to download the link to an employee in such cases. As shown in FIGURE 10D, the resulting network costs according to the 66 different heuristics are very close to each other, except for the "capacity only" increase method, which works considerably worse, 30% increase in cost. Such performance is typical among other design cases we have tested, and can be attributed to the effects of a lack of adaptation while classifying demands as common or lagging once a threshold is selected, each demand is classified, and typically is converted . In the example of the NSFNET, the selected threshold is such that the connectivity of the subgraph Ga formed by the common demands is not sufficiently rich so that some demands have to be traced by longer paths or trajectories than in the case of other topologies optimal. One possible way to improve the "capacity only" increase method is to deal with multiple thresholds, and select one that produces the lowest network cost. For the study of the NSFNET, due to the available nature of the initial structure G and the existence of demands for each pair of modes, there are few links in G that can be removed via optimization of the topology. As a result, the topology optimization heuristic produces a maximum cost reduction of approximately 3% over the non-optimized topology G. However, we have observed that in other test cases, the use of that topology optimization heuristic produces 67 much higher cost savings, from approximately 15% to more than 90%. In general, the actual savings are a strong function of the initial G topology and the distribution of traffic demands. In terms of computational efficiency, it may be relatively more expensive to rely on link discharge alone, especially when the initial G skeleton has very rich connectivity and a large portion of the demands are much smaller than the size of the line network. In those cases, the only link download method tries to download almost all the links in G, while the augmentation method can accelerate a process by quickly selecting the subset of links to form the "central" topology path in the routing of the Guardians. Finally, since the download of the link typically results in an equal or better network cost, it is contemplated to apply this after completing the augmentation method. Doing this, we observe additional savings in the costs that fluctuate from 0 to 15% in several design cases that we carry out. To conclude this subsection, FIGURES 10E and 10F give the combined impact of topology optimization and router capabilities on the network costs for the NSFNET and the carrier class structure examples, respectively. The cost of the network 68 reported in FIGURES 10E and 10F was based on the most "optimal" topology we have found for a given configuration. Even with the most conservative assumptions, there are still substantial cost benefits using the WFQ / LQD routers instead of the FIFO / RED. 4. Location of the Router Now we will consider the heterogeneous case where only a fixed number N of WFQ / LQD routers can be used together with other FIFO / RED routers in the construction of the NSFNET structure described above. Based on the location or placement method of the WFQ / LQD router described in Section 3.0, FIGURE 10G shows the optimal location for the WFQ / LQD routers when N varies. The corresponding network cost based on C, F'F? (H '? or) was calculated. Due to the assumption of bidirectional traffic demand, at least two WFQ / LQD routers are required to obtain as a result any capacity savings (and thus costs). Let us use the cost for all FIFO / RED configuration as a base. The first pair of WFQ / LQD routers, when optimally deployed, can result in cost savings of approximately 12.1%. This is achieved by producing the requirement of 69 capacity of a single long path link between node 8 (San Francisco) and node 10 (Chicago). This link results in the most significant savings due to its greater absolute capacity (in terms of the absolute number of line networks) and the greater travel or mileage. As the other WFQ / LQD routers are optimally placed, you can form a group of the links that have higher absolute capacity with a higher overaccumulation factor due to the maximum ratio f / t and /? Wj) greater than the traffic J * F, transported by the links. We also conducted a similar router location study for the carrier class network example. FIGURE 10H shows the corresponding cost reductions when different portions of the routers are replaced in the all FIFO / RED configuration with optimally located WFQ / LQD routers. Initially, when 10% of the FIFO / RED routers are optimally replaced by WFQ / LQD routers, there is a reduction of approximately 15% in the cost of the network. Again, the greatest cost reduction is due to the greater capacity for reducing expensive links (long distance). When the fraction of routers of WFQ / LQD increases to 20% and subsequently to 30%, the cost reduction is still considerable. This is due to the formation of groups of 70 WFQ / LQD, which rapidly increase the number of "beneficiary" links, that is, intragroup and intergroup. Subsequently, the speed of cost reduction is gradually leveled out, so that most of the greater savings have been extracted via the first optimization. . 0 Performance Allocation Under FIFO / Network With reference to Subsection 1.1 above, the assumptions used in M. Mathis, J. Smeke, J. Mahdavi, and T. Ott, "The Macroscopic Behavior of the Algorithm that Prevents TCP Congestion" , ACM Computer Comm. Review, Vol. 27, No. 3, pp. 67-82 (July 1997) to calculate the performance of a TCP connection are: (i) The links operate under a light or moderate packet loss, so that the dynamic window mechanism of the TCP is governed mainly by the scheme that avoids congestion, where the congestion window is divided in two when the loss of a packet is detected. Note that under conditions of severe loss, the TCP window flow control may experience downtimes, which cause the window to decrease to a packet value followed by a slow start mode. 71 (ii) Packet losses along the path or trajectory of the connection are represented by a constant loss probability p with the assumption that a packet drop or fall takes place every 1 / p transmitted packets. Under those assumptions, the congestion window of the connection behaves like a periodic wave as shown in FIGURE 11. In FIGURE 11, it is assumed that the maximum size of the Waax window is large enough, so that the window of congestion does not reach saturation (W < Wmax) • If the receiver recognizes each packet, the window is opened once each round trip (which implies that the slope in FIGURE 11 is equal to one) and that each hard cycle W / 2 round trips (r W / 2). The number of packets transmitted per cycle is given by the area under the wave, which is: Under assumption (ii), this is also equal to 1 / p, which implies that: W = 3./72.

The performance of the TCP connection is then given by: _ packages per cycle _ pags. time per cycle sec.

To calculate the performance of each connection i in the set Fi of all the TCP connections routed through link 1, we make the following assumption: (iii) Let S be the best set of congested links Xj the packet descent process in the link j. It is assumed that Xj, j and S, are independent and that each one is represented by the same probability of loss p °. A similar assumption was also used in S,. Floyd, "Connections with Multiple Congested Gates in Networks Switched by Packet Part I: One Way Traffic", ACM Computer Comm. Review, Vol. 21, No. 5, pp. 30-47 (October 1991) to calculate the TCP throughput in a linear topology of n links and n + l connections consisting of a n-hop connection that traverses all n links and a connection of one hop per link, so that each link is crossed by two connections. 73 Under this assumption, the probability of path loss for connection i is given by: where h? is the number of congested jumps in the path or path of the TCP connection. Actually, be j? f j2f. . . , jm the ordered set of congested links in the path or connection path i with jx and jm being the first and last congested links traversed by connection i, respectively. If N packets of connection i are transmitted at the origin, then p ° N are lowered on a link jj. and (l -p °) N are successfully transmitted. "Out of (l -p °) N packets arriving at the link j2, p ° (lp °) N are lowered and (lp °) 2-V are successfully transmitted By a simple induction argument, it can easily be shown that (l -p °) hl "1N constitute a bond jhl out of which p ° (lp °) hi '1N are lowered and (lp °) hlN are released. Therefore, the total number of lost packets is N- (l-p °)? LN, which corresponds to a given loss ratio in equation (25). For small values of probability of loss p °, p is equal to i - p ° when the terms of highest order of p ° in the equation (25) are neglected. Substituting in equation (24) we obtain: 74 where : / 2 • 2 j. w, = - / A.

For a given link, I and S with capacity c, FIFO, be r¡ the detachment of the TCP connection i ß Fi.

If we ignore the multiple bottleneck effect discussed and taken into account in Subsection 1.1 above and focus on link 1 only, we can treat r1 in equation (26) as if it were r and we have: Obviously, the buffer of the link must be sufficiently large (of the order of the product of the bandwidth delay) to be able to achieve a 75 aggregate yield of c, F0 as in equation (27). Yes This is not the case, the link can be underutilized. From equation (27) we obtain the value of 6 as: and the distribution of the connection performance i as: ./ _ -c w rv¡. r '= S • w. = "FIFO (28)? EF Simulations of the Mathis et al. Article, referred to above, were used to validate the result in equation (24). In the Floyd article referenced above, a different method was used to derive and validate identical simulations that result in two connections per link for the special case mentioned above. However, the result we obtained in equation (28) is a generalization of an arbitrary topology with an arbitrary TCP traffic pattern and recognizes the weighted nature of the performance allocation. 76 6.0 NP Hardness of the Router Replacement Mode Considering the following graphic problem P (G, N): Given a weighted, simple, non-directed graph G = (V, E, S) where S is a weight matrix] V "| x | V | with Sifj entries which correspond to a pair of nodes i, j € V? e so that: M (CIFIF0, L?) + T (C? "FO) -M (C? 'ÍF0, L?) -T (C? N) otherwise A saving of YES, J becomes a reality if both nodes i and j are selected. The object is to maximize the savings while maintaining the total number of selected nodes less than or equal to N. It is clear that the above graphical problem P (G, N) is a specialized version of the P problem described in Section 3.0 setting all the Q in P of Section 3.0 in zeros. The selection of the nodes in P (G, N) corresponds to the choice of the FIFO / RED locations for the WFQ / LQD update. Since P (G, N) is a special case of P, it is sufficient to show that P (G, N) is hard in NP to prove that P is hard in NP. The proof that P (G, N) is hard in NP is through the reduction of the decision version of the maximum common problem complete NP to a case of P (G, N). The common problem 77 is known in the art, for example, as described in M.R. Garey and DS Johnson, "Computers and Intractability": A Guide to the Theory of Fullness NP ", Freeman (1979) and CH Papadimitriou et al.," Combined Optimization: Algorithms and Complexity ", Pentrice Hall (1982). decision of the common problem maximum Q (G, N) can be established as follows: Given a simple, non-directed graph G = (V, E) and a positive integer N, there exists a subgroup Gs = (VJ E ') of G of so that \ V '\ = N and for all i, je V different, (i, j) e E? To prove that P (G, N) is hard in NP, we can reduce Q (G, N) in one case of P (G, N) setting: 1 otherwise.

It is evident that Q (G, N) being the answer "Yes" if and only if the maximum saving derived from P (G, N) is equal to N (N-l) / 2 and this completes the reduction. Problem P can also be transformed into a generalized backpack problem using the following , Steps. First, set the size of the backpack to Nmax and treat the FIFO / RED routers in G as points to be 78 packed where each point is of size 1. Second, assign any pair of points (i, j) a utility value S ?, which can be done if and only if both i and j points are packed in the backpack. Third, for each point i, there is an associated penalty Qi if the backpack is packed. Define the total utilitarian value for a set of points chosen as the sum of the utilitarian values by pairs Sl rJ minus the sum of the Qi penalties of the set. The selection of the optimal set for the FIFO / RED routers to be replaced then becomes the selection of a set of points to be packed derived from the restriction of the size of the backpack while maximizing the total utilitarian value. In this way, as explained, the need to provide operational guarantees in IP-based networks is more and more critical as the number of carriers is moving to IP packeting directly over SONET to provide main Internet connectivity. Advantageously, the present invention provides design and optimization algorithms and network capacity to solve those and other aspects. Also, the present invention provides algorithms that produce designs for both the homogeneous case (all networks WFQ / LQD or all FIFO / RED) as well as for heterogeneous networks where we solve the problem of the optimal location of 79 WFQ / LQD routers in an embedded network of routers FIFO / RED. It should be appreciated that although detailed descriptions of the preferred embodiments of the invention were given above, the invention is not limited thereto. For example, the system and network design methodologies of the invention can be applied to other methods to provide VPN services such as, for example: (i) the use of the service type field (TOS) in IP packets with surveillance to the routers of the edges of the network and marking of the traffic access of the VPN contract and (ii) hybrid methods that combine (i) and WFQ / LQD. Also, the invention can be implemented with a more sophisticated routing such as, for example, sending the next step through the no shortest circuit-free path via deflection and the use of TOS-based routing in OSPF. In addition, the invention can also implement other heuristics to solve the problem of location of the hard router in NP. The invention can be used to design, for example, an infrastructure to support differentiated services via a two link network combining a WFQ / LQD router network with a homogeneous FIFO / RED router network. Also, the performance model of TCP can be extended to cover regimes where dead times and / or maximum receiver / window size of the emitter, as well as to cover 80 other types of packet programming such as the WFQ between FIFO classes between flows of each class. Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it should be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be made thereto by one skilled in the art without depart from the scope and spirit of the invention.

It is noted that in relation to this date, the best method known to the applicant to carry out the aforementioned invention, is that which is clear from the present description of the invention.

Claims (6)

1. 81 CLAIMS Having described the invention as above, the content of the following claims is claimed as property: 1. A method for redesigning a packet-based communications network from an existing packet-based communications network employing a first type of router, such that a portion of the redesigned network employs a second type of router instead of the first type of router, characterized in that it comprises the steps of: specifying the replacement parameters as a whole programming representation defined as: maximize submit to? '. S?' V ?, e VJ C = y. , - = Xi, V? , j e V, C < y ,,, =? J f Vi, e V,? Y = 0 or l, 82 where V represents a set of nodes that correspond to a set of the first type of routers, E 'represents a set of links that connect the routers in V, Q is an updated cost value for a router i of the first type, Sl tj is a cost saving value, is a binary decision variable such that x2 is equal to one if the router i is selected to be replaced by a linker of the second type, and l tJ is a dependent variable that reflects the realization of the cost savings associated with a J bond equal to (i, j) according to two of the constraints, and Nmax is a parameter that specifies the maximum number of routers of the first type that can be replaced; and determining one of an optimal number of routers of the first type to be replaced by routers of the type type and the corresponding replacement places in the inmate from the entire programming representation mixed to form the network of packet-based communications, redesigned, so that Si # J is maximized substantially.
2. 2 . The method according to claim 1, characterized in that the first type of router is an encapr.adcr of the FIFO / RED type and the second type of 83 router is a WFQ / LQD router.
3. 3. The method according to claim 2, characterized in that the value of the cost saving is 5 represented as: Sl t. = [M (C2 FIFC, LxJ + f (C1 FIF0) -M (C1 WFO, L?) -fF0 (ClWF °) j where c IF0 and C? WF0 are the requirements of 10 corresponding capacity for the FIFO / RED routers and WFQ / LQD, respectively, M is a function of path cost or mileage, T is a function of the termination cost, and L is a length associated with a link 1. Four . An apparatus for redesigning a network of packet-based communications from an existing packet-based communications network employing a first type of router, so that a portion of the network indicated network employs a second type of router instead of first type of host, characterized in that it comprises C the passes of: at least one processor to specify the parameters of the replacement as a representation of crccra? a :: cr. Whole defined as: 84 maximi zar. { ¿Y, j 'StJ ~ 2_, x > -2,} (/ ,;) € £ '? E' submit to? -t, =? ^, l € l " See, j € V, 0 = yl? = xl f Vi, j € VJ 0 = y, j = xJ f Vi e VJ x, = 0 or 1, where V represents a set of nodes that correspond to the set of the first type of routers, £ "represents a set of links that connect the routers in VJ Q: is an updated cost value for a router i of the first type, S; .. is a cost-saving value, x is a binary decision variable such that j is equal to one if the keyboard is selected to be replaced by a router of the second type, and., is a dependent variable which reflects the realization of cost savings associated with a link 1 equal to (x, j) according to two of the constraints, and Ke.a> is a parameter that specifies the maximum number of cameras of the first type that it is allowed to replace, the processor also to determine one of an optimal number of routers of the first type to be replaced by enmersors of the second type and the corresponding replacement places 85 in the network from the programming representation n integer mixed to form the packet-based communications network redesigned so that YES, J is substantially maximized; and 5 memory to store the results associated with the calculation. The apparatus according to claim 4, characterized in that the first type of router is a FIFO / RED router and the second type of router is a FIFO / RED router. 10 router is a WFQ / LQD router. 6. The apparatus according to claim 5, characterized in that the value of the cost savings is represented as: 1: S;.: =. { K (C: r: r :, L:) + ^ r: (C. r: rc) -M (CjFQ, L2) - T »F0 (C? WF)) dopae ^ .F-r- and - ^ are corresponding capacity requirements for the FIFO / RED routers and WF./--- QC, respectively, K is a function of the cost of the journey or mileage, 7 is a function of the cost of completion, and i is a length associated with a link 2. A manufacturing article to redesign a communication based on packets from an existing packet-based communications network that employs a Zr first type of routing :, so that a portion of the redesigned network 86 employs a second type of router instead of the first type of router, characterized in that it comprises a readable medium in a machine containing one or more programs which, when are implemented implement the steps of: specifying the replacement parameters as a whole programming representation defined as: maxi hoist submit to L = Nm (U,? ef VJ, J € VJ 0 = yi? 3 = i, V?, J and VJ 0 = yltJ = x3. I saw VJ Xi = 0 or 1, where V represents a set of nodes that correspond to a set of the first type of routers, E 'represents a set of links that connect the routers in VJ Qi is an updated cost value for a router i of the first type, S, 3 is a cost saving value, x is a binary decision variable such that Xj is equal to one if the router i is selected to be replaced by a router of the second type, and l t3 is a dependent variable that reflects the realization of the cost saving associated with a link 1 equal to (i, j) according to two of the constraints, and Nma is a parameter that specifies the maximum number of routers of the first type that is allowed to be replaced; and determining one of an optimal number of routers of the first type to be replaced by routers of the second type and the corresponding replacement places in the network from the entire mixed programming representation to form the redesigned packet-based communications network of so that S.,; is maximized substantially. 8. The article of manufacture according to claim 7, characterized in that the first type of host is a routed: FIFO / RED and the second type of ermacker is a WFQ / LQD camper. 9. The article and manufacture according to claim 6, characterized in that the value of the cost savings is 5. .. =. { C /: FJ L - + Z * "(Xr: r-) -M (C:" rJ L :) -l "FC (C ^).}. where ~ .rF- and, ~? r- are the corresponding 2-tier requirements for the FIFO / RED and 88 WFQ / LQD routers, respectively, M is a function of path cost or mileage, T is a cost function of termination, and L is a length associated with a link J.