WO2002087153A2

WO2002087153A2 - Revenue-optimal admission controller with hard quality of service guarantees for data networks

Info

Publication number: WO2002087153A2
Application number: PCT/CA2002/000523
Authority: WO
Inventors: Eric G. Manning; Robert Kristian Watson; Shahadat Khan; Mostofa Akbar
Original assignee: Newmic Foundation
Priority date: 2001-04-20
Filing date: 2002-04-22
Publication date: 2002-10-31
Also published as: AU2002247589A1; WO2002087153A3; CA2344738A1

Abstract

A system for admission control of service requests from one or more customers into a QoS-guaranteed data network, comprising an admission controller for controlling admission of users to the network, the users being admitted to the network upon satisfying predetermined criteria that ensure optimal revenue and hard QoS for the network. Another aspect of the invention provides for a method of controlling admission of service request into a QoS-guaranteed data network to provide dynamically optimal system revenue, comprising the steps of: mapping parameters for determining admission to a variant of a combinatorial knapsack problem; solving said problem to produce one or more solutions; and using said solutions to determine whether the service request is granted.

Description

REVENUE-OPTIMAL ADMISSION CONTROLLER WITH HARD QUALITY OF

SERVICE GUARANTEES FOR DATA NETWORKS

The present invention relates in general to client/server data communication network and, more particularly, to a method and apparatus for implementing an admission controller for use in a data network.

BACKGROUND OF THE INVENTION

Multimedia traffic, comprising voice, video, images and text, is becoming increasingly common on data networks and in particular on Internets and is expected to account for a large portion of future internet use. Entertainment (movies, television and games) and videoconferences (meetings, classes) are two of the important applications. Multimedia traffic, unlike e-mail or file transfers, requires strict guarantees of the Quality of Service

(QoS) provided by the network. The QoS generally relates to the data rate (transmission speed), error rate and latency (delay) provided to the customer. Video, for example, must be transmitted at the correct rate and with few errors, or users will see jerky or distorted images. Interactive voice conversations must suffer network delays of less than a few hundred milliseconds or participants will become disoriented and unsure of who is speaking. A network, which guarantees the promised level of QoS to each customer, is called an QoS-enabled network.

Given that any network has finite resources - the capacities of its transmission links (measured in bits/sec) and routers or switches (often measured in packets/sec) - the allocation of enough link and switch capacity to each user to guarantee a given QoS means that not all applicants can be admitted to a QoS-enabled network. To do so would invite overbooking of resources and hence the violation of QoS guarantees. If there is not enough free capacity to serve a new applicant, it must be rejected. The entity, which scrutinizes applicants and admits or rejects them, is called a network admission controller. An admission controller that keeps track of committed network resources, and only admits a new applicant if sufficient uncommitted or free resources are available to meet its QoS needs, is called an admission controller with Quality of Service Guarantees. If these guarantees are absolute they are called hard; if they can be broken occasionally without penalty they are soft. Finally, an admission controller that is able to select among all of the customers on offer, so as to admit the subset which yields the highest possible revenue, is termed revenue-optimal.

While there exists a revenue-optimal admission controller for a multimedia server computer, there is currently no revenue-optimal admission controller with QoS guarantees for Internets.

SUMMARY OF THE INVENTION In acccordance with this invention there is provided a system for admission control of service requests from one or more customers into a QoS-guaranteed data network, comprising an admission controller for controlling admission of users to the network, the users being admitted to the network upon satisfying predetermined criteria that ensure optimal revenue and hard QoS for the network.

Another aspect of the invention provides for a method of controlling admission of service requests into a QoS-guaranteed data network to provide dynamically optimal system revenue, comprising the steps of: mapping parameters for determining admission to a variant of a combinatorial knapsack problem; solving said problem to produce one or more solutions; and using said solutions to determine whether the service request is granted.

BRIEF DESCRIPTION OF THE DRAWINGS

A better understanding of the invention will be obtained by reference to the detailed description below in conjunction with the following drawings in which:

Figure 1 is a schematic diagram of a system according to an embodiment of the invention;

Figure 2 is a schematic diagram showing relationships among qualities, utility and resources; and Figure 3 is a schematic diagram graphically illustrating a Multidimensional Multiple

Choice Knapsack Problem (MMKP). DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In the following description, numerous specific details are set forth to provide a thorough understanding of the invention. However, it is understood that the invention may be practiced without these specific details. In other instances, well-known software, circuits, structures and techniques have not been described or shown in detail in order not to obscure the invention. The term data processing system is used herein to refer to any machine for processing data, including the computer systems and network arrangements described herein. In the drawings, like numerals refer to like structures or processes.

Referring to figure 1 there is shown a system according to an embodiment of the present invention. The system 100 includes a plurality of customers or users 110; a QoS enabled network 120 for providing content 140 to the users, an admission controller 150 for controlling admission of users to the network, the users being admitted to the network upon satisfying predetermined criteria that ensure optimal revenue and hard QoS for the network.

The functionality of the admission controller may be better understood by first describing its use in the QoS enabled network.

Typically, customers submit requests for admission to a network. The customer's requests are expressed as Service Level Agreements (SLA's). A SLA exists for each of one or more levels of QoS and includes a data rate requested; maximum acceptable latency; an offered price; start and end times and whether the service is recurring (e.g. every Tuesday).

Depending on the SLA the admission controller must then admit the customer at one of the specified levels of QoS. By convention, a QoS Level 0 means zero data rate, infinite latency, and no charge, i.e. rejection of the SLA. A rejected customer may choose to revise the offered price upwards and bid again to be admitted; hence the admission process incorporates an auction (described later). SLAs arrive randomly in time, and the controller selects a subset of those on offer to be admitted to the network. Accordingly, SLA's are collected or batched for an interval of time which is termed an epoch. At the end of each epoch (a few seconds to a few minutes in practice) the controller selects SLAs for admission from the accumulated batch, and concurrently the next epoch begins.

A user interaction with the network is normally referred to as a session. A session is the flow of datagrams (for example, a telephone call or the viewing of a movie) requested and permitted by an SLA. For each session, a QoS level must be selected by the controller. This level determines the session revenue and the session resource requirements.

In order to guarantee service at the level of QoS selected, the controller binds the necessary resources to the session before it begins, and for the duration of its existence. In this manner we can avoid allocating the same resource to two competing sessions

(analogous to allocating an airline seat to two competing passengers), and thus violating

QoS guarantees.

The system revenue is the arithmetic sum of all session revenues. Referring to Figure 2 there is shows the relationship between system revenue U and session revenues ui(Qi), and between resource mappings r(Qi) and constraints R, established via the choice of session QoS Qi.

One of the objectives of the controller is to maximize the revenue U while respecting the system resource constraints R. This problem is called the Adaptive Multimedia Problem (AMP). The applicants have discovered that the problem of selecting SLAs for admission from an offered batch so as to maximize revenue, while observing all QoS guarantees, can be expressed as a variant of the well known Knapsack Problem in combinatorial mathematics.

By way of background the Knapsack Problem can be explained as follows: h its simplest form, we have a pile of stones, each of which has a weight u and a Volume p, and a knapsack, which has a volume P. The problem is to pick a subset of the stones which maximizes the weight of the knapsack while remaining within its volume constraint, i.e. not overfilling it ∑p≤ P. Here maximum weight is analogous to utility and the volume constraint is analogous to the resource constraint. In a variant of the Knapsack problem, the Multidimensional Multiple Choice Knapsack Problem (MMKP) we have piles of stones, and we must select exactly one stone per pile, so as to maximize weight while respecting the volume constraint. And, volumes are allowed to be vectors rather than single numbers, so the volume constraint is multidimensional, where the volume is 2-dimensional with components p and m. This means that the sum of p-values of the stones chosen must not exceed the P-value of the knapsack, and the sum of m-values of stones chosen must not exceed the M-value of the knapsack. Revenues are u-values, as before. The admission control problem is then converted into a known, well-understood problem - the MMKP. We do this by letting a stone represent a SLA at a particular level of QoS; a pile of stones represent a SLA ( all levels of QoS); the knapsack represent the data network; each volume constraint of a stone represent a requirement for one of the network's resources, i.e. data rate or latency of one link of the network; the weight of a stone be the price offered for this SLA at this level of QoS.

Thus the act of selecting a set of stones to maximize weight becomes the act of selecting a set of SLAs at particular QoS levels which maximizes revenue. To refrain from overfilling the knapsack is to refrain from oversubscribing any of the network's resources - the data rates or latencies that its links can sustain - and thus QoS is assured.

Additionally, however, we must know which links of the network a given SLA will use, in order to know the links whose data rate or bandwidth resource will be (partially or completely) consumed by the SLA. In order to determine this, we must select a path or route through the network from source to destination. A route is a sequence or chain of links connected by switches, along which datagrams flow from source to destination. We must find a route, which has adequate spare (free, uncommitted) capacity or bandwidth on each of its links. We do this by deleting all links with insufficient spare capacity from the network and applying a well-known, standard network routing algorithm to the remainder.

Having now completely modeled the problem mathematically as an MMKP, the controller solves the MMKP using one of two algorithms, described below. The first, BBLP, (Branch & Bound with Linear Programming) yields an exact solution. However, as the problem grows (bigger networks or more SLAs) the time and cost for a solution grows very rapidly (as the problem is NP-hard) so BBLP is impractical for large networks. The second solution algorithm, NHEU, is inexact. However, it is much faster and usually yields solutions, which are within 10 or 20 % of the exact, i.e. truly optimal or best solution. Each of these will be described in detail below.

The admission controller can be programmed on a standard Pentium-based computer running Windows 98. In a preferred embodiment there is implemented on the computer the NHEU solution algorithm for the MMKP, a routing algorithm (OSPF), and procedures to accept a set of SLAs and a description of the subject network's topology and link capacities.

hi one embodiment the SLAs are contained in an input file. The controller reads the network topology and capacity files and builds an internal description of the subject network. It then reads the SLAs. The Procedure NHEU is invoked to solve the resulting MMKP, and the SLAs admitted are displayed on the computer's screen, together with the revenue earned and the states of all network links.

To use the controller to regulate admissions to a real network, SLAs are passed to the controller, which batches them and selects the admitted ones by solving the MMKP. It then instructs the local switch of the network, to which it has a direct connection, to build MPLS (Multi Path Label Switching) paths corresponding to the routes chosen by the controller for the admitted SLAs. It then passes the resulting MPLS path id to the customer, who labels every datagram with this label. The usual MPLS procedures of the switch then ensure that all datagrams of this SLA are routed along this MPLS path. The controller must run fast enough to allow real-time admissions; that is, the decisions to admit SLAs, and if so at which level of QoS, must be taken as the SLAs arrive in real time. Admission of a batch can be done concurrently with the collecting of SLAs for the next batch, so the controller need only complete an admission in less than the epoch time interval: a few seconds to a few minutes are realistic values for the epoch interval.

h one implementation a controller built in Java running on a Pentium 3 microprocessor is able to admit 100 SLAs to a 30-node network in less than 2 seconds. As a re- implementation of the controller in the C programming language would yield about a factor of 10 speedup, a controller to do this task in 200 msec is feasible. Hence the controller using current technology is fast enough for real-time admission to enterprise networks (usually defined as networks of less than 100 switches or nodes).

The algorithms implemented by the controller will now be described in detail. The Admission Controller uses a heuristic I-HEU for solving the MMKP [AkbarOOl

I-HEU Setup: To the bona fide QoS levels, there is added a null QoS level, with no resource requirements and zero revenue. If the final result of the MMKP assigns the null QoS level to an SLA, then that SLA will be rejected: i.e., admission at QoS level 0 is equivalent to rejection. The null QoS level indicates whether a SLA is active or inactive; a SLA becomes active when it gets a non-null QoS level. It is up to the user to decide whether she will withdraw her bid, or wait in the inactive state for the next batch of SLAs to be processed - with or without increasing the offered price to get admission. The I-HEU has three steps as follows:

Initially, the QoS levels (including crossgrades) of an SLA is sorted in ascending order of utility (revenue) before being submitted to the admission controller. The lowest level is by convention the null QoS level. The first step - finding the feasible solution - is irrelevant here, because every null QoS level is feasible by definition. In Step 2 an SLA will be upgraded to a higher QoS level if the necessary resources are available, i.e., a network path can be found with acceptable latency bound and enough unassigned capacity to meet the upgrade's needs. hi step 3, an SLA is downgraded to a lower QoS than the previously selected QoS level and the controller then tries to find upgrades or crossgrades for other SLAs, so that total revenue increases. In each iteration, a particular QoS level and path are selected for upgrading. Additional checking is required to determine whether an upgrade complies with the path restriction, and with the values of the up and down flags. This in turn requires comparison of each QoS level of an SLA with the QoS level and path selected in the previous epoch. If the SLA manager selects a non-null QoS level for an inactive SLA after performing I-HEU, then the SLA is admitted. When an SLA becomes active, the null QoS level is removed from its profile in the next application of I-HEU, as SLAs, once admitted, are not to be rejected.

The concepts of path nearness and mutually near SLAs are used during SLA admission, changes of SLA requirements and changes of link capacities.

Path Nearness: We define a near path of path Pik(Si, Di) as one which has nodes in common with (Si,Di)- more specifically, in order:

1. Those paths which have startpoint Si and endpoint Di, or vice- versa.

2. Those paths which have Si or Di as an endpoint or startpoint

3. Those paths which contain both Si and Di 4. Those paths containing either Si or Di

Two SLAs are considered mutually near or simply near if any of the following conditions apply, in the order given: l.The SLAs share both source and destination nodes, or one SLA's source is the other's destination, or vice- versa. 2. The SLAs share either source or destination nodes, or one SLA's source is the other's destination. 3. Neither source nor destination nodes match, but the routes of one SLA pass through the source or destination node of the other.

The admission controller, during the admission of a batch of inactive SLAs, performs the I-HEU twice. In a first step it simply performs an adaptation considering only the batch of SLAs which are currently candidates for admission. The resulting newly- admitted SLAs are added to the active SLA list; those that are not admitted are sent to the second step.

The second step of admission the controller tries to reroute SLAs near to the unadmitted SLAs to other paths, to free up paths for the unadmitted ones. The controller discovers active SLAs, which are near to the unadmitted ones, and attempts to reroute them. It then performs an adaptation using the unadmitted SLAs from step 1 and the near SLAs discovered in this step

The third step of admission requires the controller to request for additional capacity on the links, which have insufficient capacity to allow admission of the unadmitted SLAs. The underlying assumption is that the facilities-based carrier who provisions our network may be able to expand the capacity of links on request, by leasing or selling additional optical wavelengths ("lambdas").

Any SLAs that remain unadmitted after all three steps are deemed rejected, and are added to the rejected SLA list.

When the capacity of a link is changed (in the worst case it may become zero, and the link may be removed due to failure), or when the cost of a link is changed, the admitted SLAs may be affected and the current set of QoS levels may not yield near-optimal revenue.

The controller first determines if the system is in a critical situation; that is, if the affected link is overbooked. If the link is not overbooked, then the identities of SLAs that are currently using the affected link are determined. An adaptation is performed on these SLAs, while respecting any SLA restrictions.

However, if the link is overbooked, then the network is in a resource contention situation. In this case, we cannot guarantee that all previous admissions will be respected. All affected SLAs are first downgraded to the lowest QoS level, and the controller then performs an adaptation on these SLAs. During this adaptation, the SLA manager ignores all SLA restrictions. If the link remains overbooked, the SLA manager tries the following strategies, in order:

It first finds nearby SLAs near to the affected ones and performs the same adaptation, ignoring SLA restrictions. If this fails, it downgrades the affected SLAs (and SLAs near to them) to null QoS level and attempts to re-admit them. Those SLAs, which remain at the null QoS level after the second step, are considered rejected.

A change in SLA requirements will occur when an SLA's QoS parameters (bandwidth requirement, delay requirement, or utility) are changed by the customer. Requirements can also change when a customer adds a new QoS level to an existing SLA. The procedure followed in either case is similar.

When a QoS level is added to an existing SLA, the controller discovers any near SLAs, and performs an adaptation on these. That is, it considers only the subgraph of the network graph defined by the paths of the SLA and those near to it , and attempts to get the necessary additional resources for this SLA by trying the following tactics: look at the existing path for free bandwidth, reroute this SLA only, reroute this SLA and SLAs near to it.

If the SLA manager fails to accommodate the new requirement using these tactics of adaptation, it finally asks for more network resources (link bandwidth, lambdas) from the underlying facilities-based carrier. IF the net utility (utility gain earned by accommodating the new requirement less the cost of the additional bandwidth) is positive, it acquires the additional resources and accommodates the new requirement. The admission control algorithm is approximate in the following respects:

First, we use the approximate algorithm I-HEU rather than the exact algorithm BBLP to solve the knapsack problem. Secondly, near SLAs and near paths as defined here do not enumerate all paths, which could, if rerouted, allow us to admit or upgrade an SLA under consideration. Third, the rerouting algorithm (kth shortest path algorithm) is used.

Our experiments suggest that the results obtained are usually within 20% of the truly optimal revenue (AkbarOl).

The Branch and Bound Algorithm BBLP

Solution State

The solution state of a node is represented by a solution vector x= {xy} where i=\,...,n and j=\,...,l_t. At any node, if an item is picked from group i, group i is said to be fixed, otherwise it is said to be free. The fixed/free statuses of the groups are indicated using the group status vector g . For i=l,...,n, if group i is free, g[i] is 0; otherwise, it is 1 and the group is fixed. For a fixed group i, if

it implies that item / is picked, and otherwise =0 implying that item is not picked. If group i is free, the value of xy may be fractional after the simplex computation. The simplex computation takes the solution state P as input and provides the value of vector elements xy, and the optimal objective value ULP^P₎ as outputs. The LP(P) is supplied using the values of n, m, /,-, uy, r_fa and R_k.

Algorithm The branch and bound algorithm for the MMKP involves the iterative generation of a search tree:

Step 1:

Start with a solution state where all groups are free. Compute the upper bound, select the branching group and initialize the tree with this node as the only live node. Step 2:

Find node e, which has the largest upper bound among the live nodes. This node is called the branching node, the expanding node or simply the e-node. Step 3:

If node e does not have any free group (i.e., all groups are fixed), then this node represents the optimal solution, and the algorithm terminates.

If node e has at least one free group, then fixing the branching expands this node group b. Fixing group b involves the following steps for each item j of this group:

From a new node t where the picked items are the picked items of node e and the item j of group b. Compute the upper bound at node t. Select the branching group if there exists any free group in t and

If node t is feasible put node t as a live node into the search tree.

Go back to step 2.

The following is the pseudo code of BBLP for solving the MMKP /* A branch and bound algorithm for MMKP */ /*Legend n: groups, lj:items in group j v :value vector, r required resource vector, R : total resource vector x : solution vector, g : group status vector, R₂ available resource vector,

V\ objective value, n₂: free groups, b: branching group and U: upper bound

procedure BblpQ

{ 1 Vι=0, n₂=n, R₂=R , g=0; /^initialize all groups free*/

2 t/=SolveLP( x, g, v ,r ,R₂ ,V ); /*Find upper bound*/

find x[b] j 1 = max x[i] [j] /*Find branching group b*/

1=1, ,n

J=L I, 4 Treelnsert( x , g , V_\ ,«₂ , R₂ ,b, U); /*Insert first node to data tree*/

5 while(l){

6 t = TreeExtractMax(); /* extract live node with highest U */

7 x = t — >x,g = t→g,b = t→b

8 if(t→ra₂ ^≤0)

9 return t -> x /* return if no free variable */

10 g[b] = 1, x[b] = 0, n₂ =( - B₂)-1 g[b]=l /* fix branching group b */

11 for(/=l,..., ) {

12 y = x,R₂ =t→R₂;

13 y[b] []= 1 ; /*pick itemy^' of group b */

14 R₂ = R₂ - ?[b] [j], V_λ=t→V_λ+ v[b] [j] ; /* Update R₂ ^a»d V\ */

15 U=SolveLP( y,g,v,r,R₂, V_\); /*find upper bound of U*/

16 lf(UX){

7=1. g[i]= branching group b*/

18 Treelnsert( x , g , V_\ ,m , R₂ ,b ', U);

/*Insert new live node to data tree*/

19 " }

20 } 1 }

22}

Computation of Upper Bound Suppose at a certain node t, there are n_\ fixed groups, and n =n-n_\ free groups. The value of the objective function achieved for the decisions already taken to reach node t can be found from the fixed groups:

;=1, ,n

J= gM=l The amount of resource available for free groups can be expressed using the m-integer vector R₂ where

R₂ = R - ∑η. j=l, ,71

7=1- /,

_?M=1

Now consider the MMKP instance P₂(t) where the «₂ free groups of P are considered as the groups of items, and vector R₂ denotes the amount of available resources. Suppose

U(t) = V_x(t) + V_LP(m)

Where V_LP^_{P m} is the LP relaxation of P (t). Since V_LP(Pι{t)) ≥ V ₎ , U(f) indicates an upper bound of the objective value of P achievable from node t. For an efficient solution of LP(P) problem, we use the Simplex Method of linear programming.

Selection of Branching Group

The node, which has the maximum value of upper bound j is chosen as the next branching node. The solution of the simplex method is also used for selecting the branching group. After the simplex method computation, the free group, which has the maximum value of xy is used as the branching group. This is based on the hypothesis that a high fractional value of xy in the solution of LP( ) would also lead us to xy =1 in the solution of P. However, the optimality of the algorithm does not depend on this hypothesis.

During the search, we use a tree data-structure, called data-tree, for maintaining the live nodes, i.e., the nodes that contain the solution states, which are still feasible. Each node of the data-tree has the data structure, which contains the following fields:

Groups status vector (g): a vector of n binary digits to indicate the fixed or free status of the groups in the current solution.

Solution vector (x): a vector of floating point numbers to store xy for

We require floating point numbers because this vector is also used to store the result of the simplex computation.

Objective value (V): a floating-point number indicating the value of objective function achieved from the fixed groups.

Number of free groups (n₂): an integer indicating the number of free groups.

Available resource vector (R₂): is a vector of m integers indicating the amount of available resources.

Branching group (b): an integer indicating the group to be fixed next if current node becomes branching node.

Upper bound (U): a floating point number indicating the upper bound of the value of the objective function achievable from the current node.

Left and Right Child Pointers (lchild and rchild): Two pointers to the data-tree nodes.

The tree is maintained in such a way that a parent node would always have a larger upper bound than any of its children. The branch and bound method requires two functions on this tree: insert a node into the data-free (function Treelnsert()) and extract the node with the largest upper bound from the data-tree (function TreeExtractMax()).

In algorithm BBLP, the data-tree is first initialized by inserting a node where all the groups are free (lines 1-5). The upper bound and branching variable are based on the solution by simplex method. Here function SolveLP() solves the linear program which is passed through its arguments. It produces the solution in vector x and returns the upper bound achievable from the current node. However if the LP is unbounded or infeasible, SolveLP returns a negative value.

Lines 5-21 form the main loop of the algorithm. Each of the iterations starts by extracting the data-tree node t, which has the highest upper bound. If node t does not have any free group (n₂=0), the algorithm terminates, and the current value of x gives the solution of P. Otherwise, we fix the branching group b. For each itemy within group b, we do the following: (1) extend x by picking item j to generate a new partial solution y ; (2) compute upper bound U(y ) by solving LP(P₂(j )) (3) select a branching group using the results of the simplex method (4) if simplex method is feasible, which implies picking item j is feasible, we insert a new node with partial solution y into the data-tree.

It may be seen then that the present invention provides an efficient and effective revenue-optimal admission controller with QoS guarantees for networks and while an embodiment of the controller is described with respect to a general-purpose computer programmed to perform the funcions of the controller, the controller could be equally well be implemented in an embedded system.

The above-described invention may be implemented in all software, all hardware, or a combination of hardware and software, including program code stored in firmware format to support dedicated hardware. A software implementation of the above described embodiment(s) may comprise a series of computer instructions either fixed on a tangible medium, such as a computer readable media. The controller may be used for access to any type of network or medium which can be either a tangible medium, including but not limited to optical or analog communications lines, or may be implemented with wireless techniques, including but not limited to microwave, infrared or other transmission techniques.

The present invention may also be implemented as a computer program product for use with a computer system capable of executing an application under the control of an operating system on the computer system, the computer program product comprising a computer usable medium having program code stored thereon.

Although the invention has been described with reference to specific embodiments, various modifications will become apparent to a person skilled in the art with departing from the spirit of the invention.

Claims

THE EMBODIMENTS OF THE INVENTION IN WHICH AN EXCLUSIVE PROPERTY OR PRIVILEGE IS CLAIMED ARE DEFINED AS FOLLOWS:

1. A system for admission control of service requests from one or more customers into a QoS-guaranteed data network, comprising:

(a) an admission controller for controlling admission of users to the network, the users being admitted to the network upon satisfying predetermined criteria that ensure optimal revenue and hard QoS for the network.

2. A method for control of admission of service requests into a QoS-guaranteed data network to provide dynamically optimal system revenue, comprising the steps of: mapping parameters for determining admission to a variant of a combinatorial knapsack problem; solving said problem to produce one or more solutions; and using said solutions to determine whether said service request is granted.

2. A method as defined in claim 2, said service request being one of a plurality of service level agreements, ones of said agreement corresponding to a predetermined level of QoS, wherein each level specifies a. the requested data rate, b. the maximum acceptable, c. the offered price, and d. start and end times.

3. A method as defined in claim 2, including a routing algorithm for setting up a route in order to meet the QoS guarantee of an admitted service request.

4. A method as defined in claim 2, including using a heuristic to find a real-time and near-optimal solution of the knapsack problem.

5. A computer-readable medium containing programming instructions for control of admission of service requests into a QoS-guaranteed data network to provide dynamically optimal system revenue, the instructions for: mapping parameters for determining admission to a variant of a combinatorial knapsack problem; solving said problem to produce one or more solutions; and using said solutions to determine whether said service request is granted.