GB2416088A - Allocating traffic carrying capacity of network links - Google Patents

Abstract

In a communications system having at least one source linked via a network of links to a plurality of destinations and having a controller at the input of the source to the network, a method of controlling the network comprising: obtaining a measure of the data traffic in terms the statistical variation of traffic data in at least one link; and allocating the capacity of link(s). This may be determining a capacity margin which is added to the "non-robust" link capacity, and involves e.g. determining the mean and variance of the data traffic on the links and/or selecting a tolerance parameter beforehand.

Description

241 6088 - 1

TELECOMMUNICATION SYSTEM

This invention relates to networks, in particular telecommunication networks, where a plurality of inputs (sources) are connected, in a network of links, to a number of destinations such that data i.e. traffic can flow from source to destination.

One fundamental prerequisite for the network planning is demand forecasting.

Such planning includes network design, routing and provisioning. Demand forecasting presents the expected traffic demand which can conceptually be represented by a traffic matrix that defines the amount of data transmitted between all possible source-destination pairs of the considered network. Deriving accurate and reliable traffic matrices by means of measurements solely is typically very costly and usually not feasible. For this reason, statistical inference techniques are known have been proposed as a more realistic yet accurate alternative to the direct measurement approaches.

One approach to dimension networks (i.e. set up link capacities) is well known which is applied in circuit-switched networks. This depends on a fixed traffic matrix and assumes only a fixed number of calls from a source to a targeted destination. This classical way is not robust to any traffic deviation that might occur and will be referred and defined hereinafter as such.

It is an object of the invention to provide means to evaluate a required link capacity that is capable of transmitting the given traffic demand with high enough quality, from traffic demand data. This data can be obtained using known traffic estimation techniques; by direct measurements or mathematical estimation techniques.

The invention comprises in a communications system having at least one source linked via a network of links to a plurality of destinations and having a controller at the input of the source to the network, a method of controlling the network comprising: i) obtaining a measure of the data traffic in terms the statistical variation of traffic data in at least one link, and ii) allocating the capacity of link(s) dependent on the result of step (i).

The determining in step (ii) may be determining a capacity margin which is added to the "non-robust" link capacity. In step i) the mean and variance of the data traffic on the links may be determined. - 2

Preferably, there is an additional step of selecting a tolerance parameter and using it with the measure from step (i) in step (ii).

In most cases the mean value of the input traffic demand into a network can be determined. Based on the given mean traffic demand and a target call blocking probability, traditional telecommunication networks are dimensioned according to Erlang formulae that take into account traffic variations around the mean and output the number of lines (trunks) required to serve the offered traffic volume. The term dimensioned is the evaluation and setting up control of the capacity of the links. This is achieved by means of call admission control (CAC) mechanism which is applied on a link basis. If the available resources on any link of the routing path are insufficient to accommodate a new connection, the call is simply blocked. Otherwise, once established, it is guaranteed a fixed transmission rate from source to destination all through its activity period. As a result, the actual traffic volume is controlled to conform with the demand. Inaccurate demand forecasting might lead to a higher call blocking probability, but a premium service is guaranteed to all admitted active connections.

In order to control the data (e.g. packet) flow in a network with integrity, the invention proposes network planning by means of a capacity margin that accounts for variation of traffic demand inside the network. The invention includes determining an adequate capacity margin, this margin is that capacity which should be added to a link in order to account for a particular value of traffic variation.

The main objective of network provisioning is the assignment of appropriate network resources which result in desired service quality. Real-time applications generate a stream type of traffic which do not tolerate quality degradation once the session is started. To experience a consistent quality level, each call requires a guaranteed effective rate r to all through its activity period. In practical IP networks, however, no complex resource control mechanisms are deployed; and thus the actual traffic has the freedom to largely deviate from the given demand causing dramatic degradation in quality. Therefore, classical ways of planning are not robust to traffic variation and thus unsuitable for IF networks.

In circuit-switched networks, resource control is performed on a link basis where no call is admitted unless all links along its routing path have sufficient resources to - 3 establish the call. Less tight resource control mechanisms are possible, varying in complexity of implementation and in the level of yielded quality. The invention uses resource control at the input of a network.

The invention will now be described in more detail and with reference to the following figures of which: Figure I shows an outline of a communication system.

Figure 2 shows the statistical representation of the established connections between sources and N destinations.

Figure 3 shows the Poisson type relationship with various values of connections K and capacity margin.

Figure 4 plots in terms of Mc for selected values of K and Pn.

Figure 5 shows the link topology of an example.

Figure 6 presents comparatively the dimensioning results of a standard approach and that of the invention.

Figure 1 shows an outline of a communication system 1 having edge routers 2 deploying a call admission control (CAC) units 3 at the input of a network to limit the maximum number of simultaneously active calls that are admitted into the network. Each call requires an effective bit rate r to perform well in the network and that the r's are additive values.

The invention sets up a sufficiently robust networks so as to perform efficient dimensioning by determining a link capacity from statistically determined inference; in other words the invention determines a suitable link capacity that accounts for traffic deviation inside the network.

To further explain the background, an example of the problem is further explained. Say a CAC unit is installed at a node A at the network input, the maximum number of connections entering the network at A is limited. The maximum number of connections is 10. So, at most 10 connections enter the network from point A. Assuming that the given or projected (estimated) traffic demand says that 6 connections go to a certain destination B and the other 4 go to another destination, C. This given traffic demand is only an estimation but in reality it is not guaranteed that only 6 go to B and 4 go to C; this is because there is no control inside the network so it might happen that 7 go - 4 to B and 3 go to C. So, if the network was dimensioned based only on the given traffic demand and accounted for only 6 connections going to B. then when 7 go to B. all the 7 connections suffer from quality degradation since the link capacity does not afford 7 connections but only 6. The invention accounts for such deviation of traffic by adding extra capacity to the links so that they can afford 7 connections for example and not only 6.

A strictly robust network is guaranteed when the total traffic allowed in to the network is accounted for in any degree of variability. In the example above, then if we want to make our network strictly robust then we should dimension the link capacities based on the worst case scenarios and that is when all 10 connections go to B or all of them go to C. So, if the given (projected) traffic demand was 6 connections from A to B and 4 from A to C, we dimension the network using a new traffic demand which is 10 connections from A to B and 10 from A to C. By doing so, it is sure that no quality degradation can occur to any of the connections since the links can now support all the 10 connections whether all of them are directed to B or to C. This way, the planned link capacities can amply serve the active connections at all times irrespective of the way they are distributed in the network. This approach is referred to hereinafter as "Strictly Robust". There is a need for a solution than what that offers a trade-offbetween network resources and performance, using adequate use of the available traffic demands to account for variability. To perform the invention a statistical distribution of the traffic volume directed to the different destinations needs to be obtained in the network and is a prerequisite to network planning problem which is to evaluate the needed link capacities that can support the given traffic demand in high quality. So, the statistical distribution of the traffic volume directed to the different destinations needs to be known: if the given traffic demand says that 6 connections are directed to destination B. cannot be trusted as in the example above we see that 7 connections might go to B since we have no control inside the network. So, what we should do is to calculate the probability that 7 connections go to B and the probability that 8 connections go to B and so on. So, we need to calculate the probability of all cases: whether 1, 2, 3, 4..., 10 connections are going to B. Then, if we found out that the probability that all 10 connections are going to B is very - 5 small then we ignore this case and do not account for 10 connections but 9 at most for example] . The resulting approach is then referred to as "Statistically Robust." To obtain the statistical distribution of traffic inside the network, we start our analysis with one-source-multiple-destination scenario. The expected traffic demand offered from source S to destinations Dn, n = 1, 2, ..., N. is given as: Dl D2... DN (1) S [ Al A2... AN] where AN represents the mean traffic load in Erlang offered from S to Dn in the busy hour. The total traffic volume A at source S is computed as:

N

A = An (2) n=l Depending on the network model in Figure 1, a call admission control unit is placed at the ingress of S and thus the maximum number of active connections allowed into the network can be bounded to K. K is the any number of connections in a system. K can be computed by numerically inverting Erlang B formula which takes A and the desired call blocking probability as input variables.

If calls are assumed to arrive according to a Poisson process and the service times have a negative exponential distribution, then the resulting system is a Poisson system as assumed in traditional telecommunication networks. AS the number of active calls in the network is limited to K; i.e. we are sure that no more than K connections can enter the network at a certain ingress point. This is because we assumed that we have a CAC unit at this ingress point. The task of this CAC unit is to limit the number of coexisting connections to a preconfigured value. We refer to this preconfigured value as K;. the resulting distribution is a truncated Poisson process. Similarly, the distribution of active calls from S to Dn is a truncated Poisson process. It is truncated since the number of connections can not be larger than K since the CAC unit at the network ingress restrict the maximum number of coexisting connections to K) with a different mean value of the l - 6 actual number of active connections. So, the maximum number is K but the mean can be something else. Since a high degree of robustness is desired, the case is considered where effectively the maximum number of co-existing connections are active and aim at evaluating the distribution of the active connections out of the K connections that are S directed to any destination Dn. For this case, the truncated Poisson process might not apply anymore.

Assuming that connections are started independently of each other and that each connection k has a probability Pk. n to be destined to Dn: Pk,n =Pn = A l <k <K (3) By this, supplementary benefit from the provided estimates of traffic demand to statistically model the system. For example, if one-source-two-destination scenario is given where Al = 6 Erlang and A2 = 4 Erlang, then we assume that a newly incoming connection is destined to D' with a probability of 0.6 and to D2 with a probability of 0.4.

Note that empirical probability values of the possible destinations of a given call can be obtained in e.g. corporate networks where traffic statistics can be more easily collected.

For example, an employee at a given branch connects to other branches of the enterprise or to external networks with certain probabilities. If, for example 100 connections are started from one branch to all other branches (let's call them branch B and branch C), and if 60 connections are directed to branch B and 40 to branch C, then a connection is directed to branch B with a probability of 60/100 = 0.6 and a connection is directed to branch C with a probability of 40/100 =0.4] Figure 2 shows the statistical representation of the established connections between S and the N destinations. Note that the traffic demand from S to each destination Dn is treated independently.

Let Kn denote the number of connections destined to Dn out of the K connections. Then,

K

Kn = Ik,n, k=l where Ik n is a Bernoulli random variable having the following probabilities: P { Ikn= 1} =Pn' (5) S P { Ik,n = 0} = 1 - Pn (6) Ik n serves as an indicator that tells whether the connection k is destined to Dn. Based on the central limit theorem, Kn then follows a normal distribution with mean pKn and variance INK computed es: Ale K Pn (7) (SKI =KPn (1-P. ) (8) The normal distribution estimation of Kn gets more accurate as K increases.

1 S Figure 3 shows that K = 50 leads already to accurate estimation. K represents the maximum number of connections that are allowed to enter the network at the ingress point, which have a CAC. So, if K=50, we already have accurate results; however, normally we have K>50 and so we have always accurate results. Note also that as K increases, the mean and variance of Kn increase as well.

Now that the distribution of the number of active connections directed to each Dn is known, we need to dimension the network properly so as to assure high performance under traffic demand variations. Our objective is to define a capacity margin Mc such that P { Cn required Cn planned + Mc} 5 it, (9) P {Kn r > Cn planned + Mc} S it, (10) where Cn required is the link capacity required to serve all active flows Kn on the link connecting S and Dn (Kn is a random variable), Cn planned is the capacity calculated based on the given traffic demand, and His a parameter that determines the tolerance of quality degradation or lack of robustness. Since Kr' is shown to be normally distributed, then Mc can be numerically calculated using the following equation: - 8 Q( Preplanned < 6' (11) where Q(x) is the Gaussian error integral known as the Q function and it is given by: Q(x)=1(1-erf( x): (12) where erf (x) is the error function given by: erf(x)=;fe(-U)du. (13) Figure 4 plots in terms of Mel for selected values of K and Pn' where Cn planned is calculated according to the mean value of Kn and r is set to 0.5 Mbps. In the left plot of Fig. 4, we observe that more capacity margin is required for the same tolerance degree. A very high capacity margin is not required since it does not increase linearly with Cn,planned. If K = 1000 and e = 5%, the required capacity margin is Mc = 10 Mbps.

Note, however, that Cn planned = 100 Mbps for K = 1000, and 1000 Mbps for K = 10000, i.e., Cn planned increased 10 times while the required Mc increased in just 3.3 times. In regards to the right plot of Fig. 4, a similar trend applies for increasing Boas long as Pn stays less than 0.5. If Pn exceeds 0.5, then higher values of Pn corresponds to less capacity margin for the same tolerance degree. The curve corresponding to Pn = 0.8 falls below that of Pn = 0.5 and exactly coincides with the curve of Pn = 0.2. This observation is expected since as given in (7) and (8), distributions with Pn= p, Vp, and Pn = 1 - p have the same mean and variance and are thus identical. But, what causes this behaviour? It can be explained when knowing that Pn = 0.5 corresponds to the most uncertain scenario. If Pn = 0, we are certain that no calls are directed to destination Dn and thus we require no extra capacity margin to account for variability. If Pn = 1, again we have a certain scenario that all calls are directed to Dn and thus no extra capacity margin is required since anyway we are considering the maximum number of connections. Therefore, for Pn = 0.5, the maximum capacity margin is required in order to account for the highest uncertainty in traffic demand. Pn is defined previously as An/A (refer to equation 3) and these values are given from the beginning in the given traffic - 9 - demand, so we can calculate Pn easily and when Pn=0.5 then we need to add the maximum capacity margin which is more than when Pn=0.3 or 0.9 or any other value.

There is no definite fixed value of the capacity margin, the value differs and can be calculated according to the given traffic demand.

The previously introduced approaches to network provisioning are now applied on a sample scenario network N 1 1 whose topology is presented in Fig. 5. Network N 1 1 represents a DiffServ network with six edge routers and five core routers. The edge routers are marked dark in the figure; a CAC unit is assumed to be placed at their input to limit the number of simultaneously active stream calls. At each edge router, 500 users are connected where each user is expected to generate a traffic load of 0.1 Erlang in the busy hour. The outgoing traffic of each edge router is evenly distributed to the other edge routers. The effective bit rate of each connection is set to r and the target call blocking probability at each CAC unit is set to 1%.

According to Erlang B formula, we calculate the maximum number of active calls at each edge router. The resulting number of connections are then distributed according to which category is selected e.g. "Not Robust" or "Strictly Robust" or "Statistically Robust". If the first approach is considered, active connections at one edge router directed to each destination (other edge routers) are distributed in proportion to the traffic ratio given in (3). Afterwards, connections are routed through the network to reach their destinations according to OSPF (Open Shortest Path First) OSPF is well known routing scheme and it computes the shortest route from origin to destination- I do not think there is a need to describe it as it could be any other routing scheme, so it is irrelevant to our invention which routing scheme are we using. At this point, the total number of active connections K' routed on each link 1, (here we are giving each link of the network an index I to differentiate among them) is known and the link capacity Cal is calculated as: C, = K, r, (14) where r is the effective bit rate required for each stream connection. When Strictly Robust is applied to achieve a completely robust network, dimensioning is performed based on the assumption that all active connections at each edge router can possibly be directed in total to each destination. OSPF Routing is then performed to obtain the - lo - number of connections traversing each link which can be used to calculate the corresponding link capacity.

Finally, the Statistically Robust approach is applied to achieve a highly robust network by accounting for traffic variations. So, described above, is the calculation of the distribution of the traffic deviation inside the network, or computing the probabilities of traffic deviation inside the network. Given = 1% and Cn piannet'= INK r, the capacity margin Mc is calculated according to (11) which is then added to all links constituting the routing path from source to destination. Since OSPF routing is used in this example and it can result in multiple paths from source to destination, then an alternative approach is followed where the maximum number of connections Kn corresponding to the given His calculated according to ( 10). Routing is afterwards performed, using known routing schemes, e.g. OSPF) and the link capacities are computed as in the previous approaches: "Strictly Robust" and "Not Robust" Figure 6 presents comparatively the dimensioning results of each approach. It is shown that Not Robust requires relatively the lowest amount of capacity but at the same time does not offer any kind of quality guarantees as soon as the actual traffic deviates from the given values. If a completely=robust network is desired, Strictly Robust is used and almost five times capacity is required in comparison to the former case. However, for a slight smoothening of quality criteria, one is able to assure a highly robust network for relatively low costs (reduced capacity resources). Statistically Robust is shown as a reasonable trade off solution which calls for around 1.5 times more capacity as compared to Not Robust. As a result, a capacity margin of 350 r is needed in our example to account, up to a certain extent, for uncertain demands in order to assure a high performance for active stream connections, which require strict quality of service constraints. - 11

Claims (4)

1. In a communications system having at least one source linked via a network of links to a plurality of destinations and having a controller at the input of the source to the network, a method of controlling the network comprising: i) obtaining a measure of the data traffic in terms the statistical variation of traffic data in at least one link, and ii) allocating the capacity of link(s) dependent on the result of step (i).

2. A method as claimed in claim 1 comprising determining in step (ii) a capacity margin which is added to the "non-robust" link capacity.

3. A method as claimed in claim 1 or 2 wherein step i) comprises determining the mean and variance of the data traffic on the links.

4. A method as claimed in claim 1, 2 or 3 comprising the additional step of selecting a tolerance parameter and using it with the measure from step (i) in step (ii) . A method as claimed in claim 2 or 4 wherein the mean is 1lKn = K-Pn and the variance is OK =K Pn (} -Pn) where Kn denote the number of connections destined to Dn out of the K connections Let Kn denote the number of connections destined to Dn out of the K connections Kn = Ik n where Ik,n is a Bernoulli random variable having the k=l following probabilities: P { Ik,n = 1} = Pn and P t Ik,n = 0} = 1 - Pn 6. A method as claimed in claim I, 2 or 3 wherein the link capacity is determined by determining the traffic demand, determining from this a Cn, panned, the capacity calculated - 12 based on the given traffic demand, and deciding a it, a parameter that determines the tolerance of quality degradation and adding a capacity margin Mc to Cn,panned, such that P { Cn,requ,red Cn,planned + Mc} 5 F, P {Kn r > Cn,planned + Mc} 5 8,.

where Cnrequred is the link capacity required to serve all active flows Kn on the link connecting S and Dn (Kn is a random variable), and. Qua c np/anned < e, r OK Pn (1-Pn) ) where Qx) is the Gaussian error integral known as the Q function and it is given by: Q(x) = 2 (1 -erf ( ; )), where or] (x) is the error function given by: erf (x) = : f e( A you 6. A method as claimed in any previous claim wherein said system is an IP network.