Chapter 7 - Applying Policies in the Sample Network
On This Page
Policy is a much-overused term. There are policies in selecting network equipment, policies in selecting vendors, policies in selling services, policies in provisioning networks, policies in granting resources, and so on. In this section we discuss policies specifically related to the granting of network resources after the network has been built and all long term provisioning has been applied. The policies in which we are interested determine specifically how much resource of each type is granted to which users and applications.
Granting Resources Based on Policy vs. Availability
To the first order, resources are granted based on availability. For example, a provider's SLA, from the customer's perspective, specifies resources available at each service level, without regard for the particular customer’s user or application that may claim these resources1. A customer policy might specify which users and/or applications in the customer's network are allowed to make use of these resources. Thus, rather than use available resources on a first-come-first-serve basis, the customer applies policy that restricts resource usage to certain consumers. Similarly, an RSVP-enabled router might be configured to admit requests for up to 100 Kbps traffic for the guaranteed service level. Policy would tell it which users or applications are entitled to use the 100 Kbps capacity.
It is possible to provision certain policies in a top-down manner. For example, a network provider might provision devices in the provider's network to provide a specific customer a specific capacity at a certain service level. This is a fairly coarse grain policy. It can be simply provisioned, so long as there is an easy way to identify traffic originating from the customer. Assuming that all traffic from the customer originates from source addresses on, for example, subnet 188.8.131.52, then the network provider can provision devices within the network to recognize this source address and to police traffic sent from this address to the appropriate limits. This constitutes a provisioned policy of the provider regarding the specific customer.
We will now look at the resources available in the provider's network, from the customer's perspective. The customer would like to apply finer-grain policies. For example, the customer may want to restrict the usage of capacity in the expensive service level to a group of privileged users running important applications. Fine-grain policies such as these can be applied using a relatively static provisioning approach, however, the finer-grain the policies, the more cumbersome this approach becomes.
In order to apply fine-grain policies, it is first necessary to define these policies in terms of classification criteria and the resources to which classified packets are entitled. Let's say, for example, that all users from the marketing department are entitled to certain privileges distinct from those to which users from the engineering department are entitled. In this case, classification criteria would have to include the set of IP source addresses for all marketers and the set of IP source addresses for all engineers. If these IP groups of users were separated by subnet, this classification criteria could be expressed in a relatively compact form, however, in the general case, management of the required classification criteria would be extremely cumbersome.
An additional complication of such statically provisioned policy information is that it is hard to reconcile it with resource availability. For example, assume that it is necessary to install a policy to the effect that only executives using the IP telephony application are entitled to make use of the low latency services in the network. Assume that it is possible to define classification criteria that recognize traffic from executives using IP telephony, and also assume that the classification criteria can be used to direct this traffic to the low latency queue in each device. Recall that this queue has limited capacity and it may be possible to accommodate only 10 simultaneous users at any given node, out of the set of all executives.
The desired effect is a combination of resource availability and policy criteria, in the form: allow up to 10 simultaneous executives using IP telephony to access these resources. This is very difficult to implement using a static provisioning approach to policy. It would be necessary to provision classification criteria for only ten executives at a time. The problem is in determining which executives to allow at any point in time. The subset of executives that should be allowed at any time changes dynamically. Of course, this applies primarily to policies regarding applications that require high quality guarantees. For applications that do not require high quality guarantees, considerations regarding resource availability are not as strict and it is therefore possible to simply allow all executives, based on statistical assumptions regarding executive resource usage.
Dynamic Enforcement of Policies
From the example in the previous section, we see that it is difficult to enforce fine-grain policies in a useful manner by using a provisioning approach. Policies, by their nature, are relatively static. However, efficient enforcement of these policies requires a more dynamic approach than provisioning. In this section we discuss the application of policy based on dynamic signaling. This approach is particularly applicable to applications that signal and that require relatively high quality guarantees.
We have previously discussed the use of signaling to effect dynamic admission control based on the availability of resources. This approach relies on the appointment of admission control agents in the signaling path. These agents consider the availability of requested resources along a path before admitting a resource request. If the resources are available, devices along the path install classification criteria corresponding to the traffic for which resources were requested. We can enforce policies by requiring the admission control agents to consider not only resource availability but also policies regarding who is entitled to these resources.
In addition, this approach installs classification criteria in devices dynamically, in response to the results of a signaled request. The results of the signaled request are based on merging of both resource availability and policy information. The effect of this approach (as applied to the example of policies regarding executives using IP telephony) is that at any time, the classification criteria installed in a device will allow resources only to a subset of executives that does not exceed the capacity available.
Dynamic enforcement of policies in response to signaling is implemented by requiring certain admission control agents to apply a policy check in the process of admitting or rejecting a signaled resource request. This is typically implemented by enabling a device in the network through which data (and resource requests) flow, to outsource policy requests. Outsourcing policy requests consists of stripping policy objects describing the requesting user and application from a signaled resource request, and forwarding the policy objects, with a specification of the requested resources, to a policy server. The policy server then consults a database of users, applications, and privileges to which they are entitled and returns an admit/reject decision to the network device. The network device acts as a PEP and the policy server is the PDP.
Scope of Policies
The scope of a specific set of policies generally does not extend beyond a single administrative domain. For example, the policies of the large network provider determine the allocation of resources among customers of the provider. The provider does not care which of each customer's users are making use of the resources allotted to the customer. That is a matter of each customer's internal policies.
The policy objects inserted by the hosts that originate resource requests are very fine-grain. They describe individual users and the application used by the user. The scope of these objects is the customer network. These objects can be acted upon by PEPs and PDPs in the customer network, but are not useful to the provider's network. In the short run, the provider's networks will tend to be relatively statically provisioned, supporting static SLAs only. As a result, providers have little need to dynamically enforce policies. Their policies can be enforced via static provisioning, as described in section 7.2.
However, in the long run, providers will offer dynamic SLAs to their customers. These will allow customers to grow or shrink their resource usage, subject to policies. This flexibility will be reflected in the form of dynamic SLAs and will be supported by signaling-based admission control within the provider's network. In this environment, the provider will need to dynamically enforce policies regarding varying resource usage by different customers. The provider can be expected to apply the same approach described for dynamic enforcement of policies in the customer's network. However, the provider's admission control agents will apply policies based on policy objects describing customers, not individuals within a customer's network. If the provider processes per-conversation signaling requests from customers, it will be necessary to insert policy objects describing the customer in the signaling request. These may either replace (or be appended to) the original user/application policy objects. In general, as resource requests traverse administrative domain boundaries, it will be necessary to insert policy objects that are meaningful to each domain interested in dynamic enforcement of policies.
Note that, if aggregate signaling is used within the provider's network, as described in section 184.108.40.206, then policy objects pertinent to the provider are easily separated from those pertinent to the customer. In fact, the customer's policy objects remain invisible to the provider.
Multicast and Policy Objects
RSVP signaling is designed to merge reservations for multicast resources as appropriate. For example, requests from two receivers for resources for the same multicast session will automatically be merged along paths that are common to both receivers. This is appropriate from a resource perspective as both receivers can be satisfied using only one set of resources. However, such merging of requests complicates policy decisions. If policies dictate that one receiver is entitled to resources and another is not, what is the appropriate policy decision? Is it to admit the request, thereby enabling a free-rider, or is it to reject the request, thereby penalizing an entitled receiver? Furthermore, how are the policy objects conveyed upstream? Should merged requests include all policy objects from each pre-merge request? This approach could lead to unmanageably large sets of policy objects. Many of the multicast issues affecting policy have not yet been resolved.
|1||From the provider's perspective, these resources represent the fraction of available resources at the provider's ingress (and further into the network) that are granted to the specific customer and, as such, represent a policy regarding the specific customer.|