Side_1_360

As a first approximation, suppose that the esti-

mated bit-rate B^tis between Ciand Ci+1at time t,

then the CDR is set to Ci+1. The first time B^t

crosses one of the levels Ciand Ci+1we change

the CDR to Ciif the crossing is downwards and

to Ci+2if the crossing is upwards.

It might also be necessary to introduce some

safety factor safeso that CDR is set to safe⋅Ci.

This could for instance be done for high priority

traffic such as the EF class (overbooking case)

or for real-time traffic in general.

Rate reductions can be less frequent. This can be

achieved by using a larger integration time for

such decisions. Another solution could be to take

into account past observations and for example

use exponential weighting of the observations.

If this is not done, only one measurement with

CDR less than the reduction threshold should in

any case normally not trigger a rate reduction.

At times with low traffic, no rate reductions are

necessary. The control system should then keep

track of bandwidth demands for the different

LSPs, so that when the network starts to get

loaded bandwidth can be freed from LSPs not

needing it.

A problem arises when a request for bandwidth

increase is rejected. Different options then apply,

like (in the relevant order)

• Release bandwidth from other LSPs not need-
  ing it;
  
  
  • Pre-empt lower priority traffic if possible;
  
  
  • Set-up a new path using other physical
    resources with spare bandwidth;
  
  
  • Try to re-optimise the network.
  
  
  
  

A Control Framework for a

Core Network by Use of MPLS

and DiffServ

Discussion of Different Alternatives

To be able to offer QoS to users it is necessary,

as we have seen, to have some control over the

use of network resources. A basic part of such

control builds on agreements with the users of

the network (SLA). Another part could be moni-

toring of LSPs.

Figure 5 shows the overall scenario. User traffic

is monitored and enforced at the ingress to the

network (SLA monitoring). This can be in an

edge router, or preferably as near to the user as

possible.

In principle we can distinguish between three

types of services:

i Service with connectivity to a predefined set

of destination points. An example can be Vir-

tual Leased Line service. In this case SLA

specifies the allowed traffic towards these

destination points (pipe model, Committed

Information Rate – CIR-SLA), we have no

spatial gambling and the necessary resources

can be reserved in the network.

ii Service with call admission control function-

ality. VoIP may be implemented in this way.

The call admission control will decide

whether to permit or deny a given call request

based on knowledge about the available re-

sources in the network. If the conclusion is

negative the service is blocked, otherwise the

necessary resources can be reserved and the

call set up.

iii A one-to-any service without call admission

control functionality. In this case the SLA

only controls the volume of the traffic flow-

ing over the user-network interface (of each

class and both ways). This is called a hose

SLA (Committed Access Rate – CAR-SLA).

The SLA is therefore not enough to control

the volume of traffic in a given direction, i.e.

we have a kind of spatial gambling on traffic

volume. (In the downstream direction the

SLA will be any-to-one, also called a funnel

SLA.)

The latter case is of most concern from a QoS/

control viewpoint. This is the service type that

Figure 5 Overall scenario

SLA

monitoring

Core with

DiffServ /

MPSL network

Edge