Internet Engineering Task Force               Raj Yavatkar, Intel
  INTERNET-DRAFT                                Dimitrios Pendarakis, IBM
                                                Roch Guerin, U. Of Pennsylvania
                                                November 1998
                                                Expires: June 1999
                A Framework for Policy-based Admission Control
                             Status of this Memo
  This document is an Internet-Draft.  Internet-Drafts are working
  documents of the Internet Engineering Task Force (IETF), its areas,
  and its working groups.  Note that other groups may also distribute
  working documents as Internet-Drafts.
  Internet-Drafts are draft documents valid for a maximum of six months
  and may be updated, replaced, or obsoleted by other documents at any
  time.  It is inappropriate to use Internet-Drafts as reference
  material or to cite them other than as ``work in progress.''
  To learn the current status of any Internet-Draft, please check the
  ``1id-abstracts.txt'' listing contained in the Internet-Drafts Shadow
  Directories on (Africa), (Europe),
  (Pacific Rim), (US East Coast), or (US West
  This document is a product of the RSVP Admission Policy (RAP) working
  group in the Transport Area of the Internet Engineering Task Force.
  Comments are
  solicited and should be addressed to the working group's mailing list at, and/or the author(s).
                                                                    [Page 1]

                A Framework for Policy-based Admission Control     Nov. 1998
  1. Abstract
  The IETF working groups such as Integrated Services (called "int-serv")
  and RSVP [1] have developed extensions to the IP architecture and the
  best-effort service model so that applications or end users can request
  specific quality (or levels) of service from an internetwork in addition
  to the current IP best-effort service. Recent efforts in the Differen-
  tiated Services Working Group are also directed at definition of mechan-
  isms that support aggregate QoS services. The int-serv model for these new
  services requires explicit signaling of the QoS (Quality of Service)
  requirements from the end points and provision of admission and traffic
  control at Integrated Services routers. The proposed standards for RSVP
  [RFC 2205] and Integrated Services [RFC 2211, RFC 2212] are examples of a
  new reservation setup protocol and new service definitions respectively.
  Under the int-serv model, certain data flows receive preferential treat-
  ment over other flows; the admission control component only takes into
  account the requester's  resource reservation request and available capa-
  city to determine whether or not to accept a QoS request. However, the
  int-serv mechanisms do not include  an important aspect of admission con-
  trol: network managers and service providers must be able to monitor, con-
  trol, and enforce use of network resources and services based on policies
  derived from criteria such as the identity of users and applications,
  traffic/bandwidth requirements, security considerations, and time-of-
  day/week. Similarly, diff-serv mechanisms also need to take into account
  policies that take into account various criteria such as customer iden-
  tity, ingress points, and so on.
  This document is concerned with specifying a framework for providing
  policy-based control over admission control decisions. In particular, it
  focuses on policy-based control over admission control using RSVP as an
  example of the QoS signaling mechanism. Even though the focus of the work
  is on RSVP-based admission control, the document outlines a framework that
  can provide policy-based admission control in other QoS contexts. We argue
  that policy-based control must be applicable to different kinds and quali-
  ties of services offered in the same network and our goal is to consider
  such extensions whenever possible.
  We begin with a list of definitions in Section 2. Section 3 lists the
  requirements and goals of the mechanisms capable of controlling and
  enforcing access to better QoS.  We then outline the architectural ele-
  ments of the framework in Section 4 and describe the functionality assumed
  for each component.  Section 5 discusses example policies, possible
  scenarios, and policy support needed for those scenarios. Section 6 speci-
  fies the requirements for a client-serv protocol for communication between
  a policy server (PDP) and its client (PEP) and evaluates suitability of
  some of the existing protocols for this purpose.
                                                                    [Page 2]

                A Framework for Policy-based Admission Control     Nov. 1998
  2. Terminology
  The following is a list of terms used in this document.
    -    Administrative Domain: A collection of networks under the same
         administrative control and grouped together for administrative pur-
    -    Network Element or Node: Routers, switches, hubs are examples of
         network nodes. They are the entities where resource allocation
         decisions have to be made and the decisions have to be enforced. A
         RSVP router which allocates part of a link capacity (or buffers) to
         a particular flow and ensures that only the admitted flows have
         access to their reserved resources is an example of a network ele-
         ment of interest in our context.
         In this document, sometimes we use the terms router,  network ele-
         ment, and network node interchangeably, but should be interpreted
         as reference to a network element.
    -    QoS Signaling Protocol: A signaling protocol that carries an admis-
         sion control request for a bandwidth resource, e.g., RSVP.
    -    Policy: The combination of rules and services where rules define
         the criteria for resource access and usage.
    -    Policy control: The application of rules to determine whether or
         not access to a particular resource should be granted.
    -    Policy Object:  Contains policy-related info such as policy ele-
         ments and is carried in a request or response related to resource
         allocation decision.
    -    Policy Element: Subdivision of policy objects; contains single
         units of information necessary for the evaluation of policy rules.
         A single policy element carries an user or application identifica-
         tion whereas another policy element may carry user credentials or
         credit card information.  Examples of policy elements include iden-
         tity of the requesting user or application, user/app credentials,
         etc. The policy elements themselves are expected to be independent
                                                                    [Page 3]

                A Framework for Policy-based Admission Control     Nov. 1998
         of which QoS signaling protocol is used.
    -    Policy Decision Point (PDP): The point where policy decisions are
    -    Policy Enforcement Point (PEP): The point where the policy deci-
         sions are actually enforced.
    -    Policy Ignorant Node (PIN): A network element that does not expli-
         citly support policy control using the mechanisms defined in this
    -    Resource: Something of value in a network infrastructure to which
         rules or policy criteria are first applied before access is
         granted. Examples of resources include the buffers in a router and
         bandwidth on an interface.
    -    Service Provider: Controls the network infrastructure  and may be
         responsible for the charging and accounting of services.
    -    Soft State Model - Soft state is a form of the stateful model that
         times out installed state at a PEP or PDP. It is an automatic way
         to erase state in the presence of communication or network element
         failures. For example, RSVP uses the soft state model for instal-
         ling reservation state at network elements along the path of a data
    -    Installed State: A new and unique request made from a PEP to a PDP
         that must be explicitly deleted.
    -    Trusted Node: A node that is within the boundaries of an adminis-
         trative domain (AD) and is trusted in the sense that the admission
         control requests from such a node do not necessarily need a PDP
  3. Policy-based Admission Control: Goals and Requirements
                                                                    [Page 4]

                A Framework for Policy-based Admission Control     Nov. 1998
  In this section, we describe the goals and requirements of mechanisms and
  protocols designed to provide policy-based control over admission control
    -    Policies vs Mechanisms: An important point to note is that the
         framework does not include any discussion of any  specific policy
         behavior or does not require use of specific policies. Instead, the
         framework only outlines the architectural elements and mechanisms
         needed to allow a wide variety of possible policies to be carried
    -    RSVP-specific: The mechanisms must be designed to meet the policy-
         based control requirements specific to the problem of bandwidth
         reservation using RSVP as the signaling protocol. However, our goal
         is to allow for the application of this framework for admission
         control involving other types of resources and QoS services (e.g.,
         Diff-Serv) as long as we do not diverge from our central goal.
    -    Support for preemption: The mechanisms designed must include sup-
         port for preemption. By preemption, we mean an ability to remove a
         previously installed state in favor of accepting a new admission
         control request.  For example, in the case of RSVP, preemption
         involves the ability to remove one or more currently installed
         reservations to make room for a new resource reservation request.
    -    Support for many styles of policies: The mechanisms designed must
         include support for many policies and policy configurations includ-
         ing bi-lateral and multi-lateral service agreements and policies
         based on the notion of relative priority.  In general, the determi-
         nation and configuration of viable policies are the responsibility
         of the service provider.
    -    Provision for Monitoring and Accounting Information:  The mechan-
         isms must include support for monitoring policy state, resource
         usage, and provide access information. In particular, mechanisms
         must be included to provide usage and access information that may
         be used for accounting and billing purposes.
    -     Fault tolerance and recovery: The mechanisms designed on the basis
         of this framework must include provisions for fault tolerance and
         recovery from failure cases such as failure of PDPs, disruption in
                                                                    [Page 5]

                A Framework for Policy-based Admission Control     Nov. 1998
         communication including network partitions (and subsequent merging)
         that separate a PDP from its peer PEPs.
    -    Support for Policy-Ignorant Nodes (PINs):  Support for the mechan-
         isms described in this document should not be mandatory for every
         node in a network. Policy based admission control could be enforced
         at a subset of nodes, for example the boundary nodes within an
         administrative domain. These policy capable nodes would function as
         trusted nodes from the point of view of the policy-ignorant nodes
         in that administrative domain.
    -    Scalability:  One of the important requirements for the mechanisms
         designed for policy control is scalability. The mechanisms must
         scale at least to the same extent that RSVP scales in terms of
         accommodating multiple flows and network nodes in the path of a
         flow. In particular, scalability must be considered when specifying
         default behavior for merging policy data objects and merging should
         not result in duplicate policy elements or objects. There are
         several sensitive areas in terms of scalability for policy control
         over RSVP. First, not every policy aware node in an infrastructure
         should be expected to contact a remote PDP. This would cause poten-
         tially long delays in verifying requests that must travel up hop by
         hop. Secondly, RSVP is capable of setting up resource reservations
         for multicast flows. This implies that the policy control model
         must be capable of servicing the special requirements of large mul-
         ticast flows. Thus, the policy control architecture must scale at
         least as well as RSVP based on factors such as the size of RSVP
         messages, the time required for the network to service an RSVP
         request, local processing time required per node, and local memory
         consumed per node.
    -    Security and denial of service considerations: The policy control
         architecture must be secure as far as the following aspects are
         concerned. First, the mechanisms proposed under the framework must
         minimize theft and denial of service threats. Second, it must be
         ensured that the entities (such as PEPs and PDPs) involved in pol-
         icy control can verify each other's identity and establish neces-
         sary trust before communicating.
  4. Architectural Elements
  The two main architectural elements for policy control are the PEP (Policy
                                                                    [Page 6]

                A Framework for Policy-based Admission Control     Nov. 1998
  Enforcement Point) and the PDP (Policy Decision Point). Figure 1 shows a
  simple configuration involving these two elements; PEP is a component at a
  network node and PDP is a remote entity that may reside at a policy
  server.  The PEP represents the component that always runs on the policy
  aware node. It is the point at which policy decisions are actually
  enforced. Policy decisions are made primarily at the PDP. The PDP itself
  may make use of additional mechanisms and protocols to achieve additional
  functionality such as user authentication, accounting, policy information
  storage, etc. For example, the PDP is likely to use an LDAP-based direc-
  tory service for storage and retrieval of policy information[6]. This
  document does not include discussion of these additional mechanisms and
  protocols and how they are used.
  The basic interaction between the components begins with the PEP. The PEP
  will receive a notification or a message that requires a policy decision.
  Given such an event, the PEP then formulates a request for a policy deci-
  sion and sends it to the PDP.  The request for policy control from a PEP
  to the PDP may contain one or more policy elements (encapsulated into one
  or more policy objects) in addition to the admission control information
  (such as a flowspec or amount of bandwidth requested) in the original mes-
  sage or event that triggered the policy decision request.  The PDP returns
  the policy decision and the PEP then enforces the policy decision by
  appropriately accepting or denying the request.  The PDP may also return
  additional information to the PEP which includes one or more policy ele-
  ments. This information need not be associated with an admission control
  decision. Rather, it can be used to formulate an error message or
  outgoing/forwarded message.
        ________________         Policy server
       |                |        ______
       |  Network Node  |        |     |------------->
       |    _____       |        |     |   May use LDAP,SNMP,.. for accessing
       |   |     |      |        |     |  policy database, authentication,etc.
       |   | PEP |<-----|------->| PDP |------------->
       |   |_____|      |        |_____|
       |                |
  Figure 1: A simple configuration with the primary policy control
  architecture components. PDP may use additional mechanisms and protocols
  for the purpose of accounting, authentication, policy storage, etc.
  The PDP might optionally contact other external servers, e.g., for access-
  ing configuration, user authentication, accounting and billing databases.
  Protocols defined for network management (SNMP) or directory access (LDAP)
  might be used for this communication. While the specific type of access
  and the protocols used may vary among different implementations, some of
                                                                    [Page 7]

                A Framework for Policy-based Admission Control     Nov. 1998
  these interactions will have network-wide implications and could impact
  the interoperability of different devices.
  Of particular importance is the "language" used to specify the policies
  implemented by the PDP. The number of policies applicable at a network
  node might potentially be quite large. At the same time, these policies
  will exhibit high complexity, in terms of number of fields used to arrive
  at a decision, and the wide range of decisions. Furthermore, it is likely
  that several policies could be applicable to the same request profile. For
  example, a policy may prescribe the treatment of requests from a general
  user group (e.g., employees of a company) as well as the treatment of
  requests from specific members of that group (e.g., managers of the com-
  pany). In this example, the user profile "managers" falls within the
  specification of two policies, one general and one more specific.
  In order to handle the complexity of policy decisions and to ensure a
  coherent and consistent application of policies network-wide, the policy
  specification language should ensure unambiguous mapping of a request pro-
  file to a policy action. It should also permit the specification of the
  sequence in which different policy rules should be applied and/or the
  priority associated with each one. Some of these issues are addressed in
  In some cases, the simple configuration shown in Figure 1 may not be suf-
  ficient as it might be necessary to apply local policies (e.g., policies
  specified in access control lists) in addition to the policies applied at
  the remote PDP. In addition, it is possible for the PDP to be co-located
  with the PEP at the same network node. Figure 2 shows the possible confi-
  The configurations shown in Figures 1 and 2 illustrate the flexibility in
  division of labor. On one hand, a centralized policy server, which could
  be responsible for policy decisions on behalf of multiple network nodes in
  an administrative domain, might be implementing policies of a wide scope,
  common across the AD. On the other hand, policies which depend on informa-
  tion and conditions local to a particular router and which are more
  dynamic, might be better implemented locally, at the router.
                                                                    [Page 8]

                A Framework for Policy-based Admission Control     Nov. 1998
     ________________                        ____________________
    |                |                      |                    |
    |  Network Node  |  Policy Server       |    Network Node    |
    |    _____       |      _____           |  _____      _____  |
    |   |     |      |     |     |          | |     |    |     | |
    |   | PEP |<-----|---->| PDP |          | | PEP |<-->| PDP | |
    |   |_____|      |     |_____|          | |_____|    |_____| |
    |    ^           |                      |                    |
    |    |    _____  |                      |____________________|
    |   \-->|     | |
    |        | LPDP| |
    |        |_____| |
    |                |
  Figure 2: Two other possible configurations of policy control
  architecture components. The configuration on left shows a local decision
  point at a network node and the configuration on the left shows PEP and
  PDP co-located at the same node.
  If it is available, the PEP will first use the LPDP to reach a local deci-
  sion. This partial decision and the original policy request are next sent
  to the PDP which  renders a final decision (possibly, overriding the
  LPDP). It must be noted that the PDP acts as the final authority for the
  decision returned to the PEP and the PEP must enforce the decision ren-
  dered by the PDP. Finally, if a shared state has been established for the
  request and response between the PEP and PDP, it is the responsibility of
  the PEP to notify the PDP that the original request is no longer in use.
  Unless otherwise specified, we will assume the configuration shown on the
  left in Figure 2 in the rest of this document.
  Under this policy control model, the PEP module at a network node must use
  the following steps to reach a policy decision:
    1.   When a local event or message invokes PEP for a policy decision,
         the PEP creates a request that includes information from the mes-
         sage (or local state) that describes the admission control request.
         In addition, the request includes appropriate policy elements as
         described below.
    2.   The PEP may consult a local configuration database to identify a
         set of policy elements (called set A) that are to be evaluated
         locally. The local configuration specifies the types of policy ele-
         ments that are evaluated locally. The PEP passes the request with
                                                                    [Page 9]

                A Framework for Policy-based Admission Control     Nov. 1998
         the set A to the Local Decision point (LPDP) and collects the
         result of the LPDP (called "partial result" and referred to as D(A)
    3.   The PEP then passes the request with ALL the policy elements and
         D(A) to the PDP. The PDP applies policies based on all the policy
         elements and the request and reaches a decision (let us call it
         D(Q)). It then combines its result with the partial result D(A)
         using a combination operation to reach a final decision.
    4.   The PDP returns the final policy decision (one after the combina-
         tion operation) to the PEP.
  Note that in the above model, the PEP  *must* contact the PDP even if no
  (or NULL) policy objects are received in the admission control request.
  This requirement would help ensure that a request cannot bypass policy
  control by omitting policy elements in a reservation request. However,
  ``short circuit'' processing is permitted, i.e., if the result of D(A),
  above, is ``no'', then there is no need to proceed with further policy
  processing at the policy server. Still, the PDP must be informed of the
  failure of local policy processing. The same applies to the case when pol-
  icy processing is successful but admission control (at the resource
  management level due to unavailable capacity) fails; again the policy
  server has to be informed of the failure.
  It must also be noted that the PDP may, at any time, send an asynchronous
  notification to the PEP to change its earlier decision or to generate a
  policy error/warning message.
  4.1. Example of a RSVP Router
  In the case of a RSVP router, Figure 3 shows the interaction between a PEP
  and other int-serv components within the router.  For the purpose of this
  discussion, we represent all the components of RSVP-related processing by
  a single RSVP module, but more detailed discussion of the exact interac-
  tion and interfaces between RSVP and PEP will be described in a separate
  document [3].
                                                                   [Page 10]

                A Framework for Policy-based Admission Control     Nov. 1998
    |                              |
    |           Router             |
    |  ________           _____    |            _____
    | |        |         |     |   |           |     |
    | |  RSVP  |<------->| PEP |<--|---------->| PDP |
    | |________|         |_____|   |           |_____|
    |      ^                       |
    |      |      Traffic control  |
    |      |      _____________    |
    |      \---->|  _________  |   |
    |            | |capacity | |   |
    |            | | ADM CTL | |   |
    |            | |_________| |   |
  --|----------->|  ____ ____  |   |
    |   Data     | | PC | PS | |   |
    |            | |____|____| |   |
    |            |_____________|   |
    |                              |
  Figure 3: Relationship between PEP and other int-serv components
  within an RSVP router. PC -- Packet Classifier, PS -- Packet Scheduler
  When a RSVP message arrives at the router (or an RSVP related event
  requires a policy decision), the RSVP module is expected to hand off the
  request (corresponding to the event or message) to its PEP module. The PEP
  will use the PDP (and LPDP) to obtain the policy decision and communicate
  it back to the RSVP module.
  4.2. Additional functionality at the PDP
  Typically, PDP returns the final policy decision based on an admission
  control request and the associated policy elements. However, it should be
  possible for the PDP to sometimes ask the PEP (or the admission control
  module at the network element where PEP resides) to generate policy-
  related error messages. For example, in the case of RSVP, the PDP may
  accept a request and allow installation and forwarding of a reservation to
  a previous hop, but, at the same time, may wish to generate a
  warning/error message to a downstream node (NHOP) to warn about conditions
  such as "your request may have to be torn down in 10 mins, etc." Basi-
  cally, an ability to create policy-related errors and/or warnings and to
  propagate them using the native QoS signaling protocol (such as RSVP) is
  needed. Such a policy error returned by the PDP must be able to also
  specify whether the reservation request should still be accepted,
  installed,  and forwarded to allow continued normal RSVP processing. In
  particular, when a PDP  sends back an error, it specifies that:
                                                                   [Page 11]

                A Framework for Policy-based Admission Control     Nov. 1998
    1. the message that generated the admission control request should be
    processed further as usual, but an error message (or warning) be sent in
    the other direction and include the policy objects supplied in that
    error message
    2. or, specifies that an error be returned, but the RSVP message should
    not be forwarded  as usual.
  4.3. Interactions between PEP, LPDP, and PDP at a RSVP router
  All the details of RSVP message processing and associated interactions
  between different elements at an RSVP router (PEP, LPDP) and PDP are
  included in separate documents [3,8]. In the following, a few, salient
  points related to the framework are listed:
    *    LPDP is optional and may be used for making decisions based on pol-
         icy elements handled locally. The LPDP, in turn, may have to go to
         external entities (such as a directory server or an authentication
         server, etc.) for making its decisions.
    *    PDP is stateful and  may make decisions even if no policy objects
         are received (e.g., make decisions based on information such as
         flowspecs and session object in the RSVP messages). The PDP may
         consult other PDPs, but discussion of inter-PDP communication and
         coordination is outside the scope of this document.
    *    PDP sends asynchronous notifications to PEP whenever necessary to
         change earlier decisions, generate errors etc.
    *    PDP exports the information useful for usage monitoring  and
         accounting purposes. An example of a useful mechanism for this pur-
         pose is a MIB or a relational database. However, this document does
         not specify any particular mechanism for this purpose and discus-
         sion of such mechanisms is out of the scope of this document.
       4.4. Placement of Policy Elements in a Network
       By allowing division of labor between an LPDP and a PDP, the policy
                                                                   [Page 12]

                A Framework for Policy-based Admission Control     Nov. 1998
       control architecture allows staged deployment by enabling routers of
       varying degrees of sophistication, as far as policy control is con-
       cerned, to communicate with policy servers. Figure 4 depicts an exam-
       ple set of nodes belonging to three different administrative domains
       (AD) (Each AD could correspond to a different service provider in
       this case).  Nodes A, B and C belong to administrative domain AD-1,
       advised by PDP PS-1, while D and E belong to AD-2 and AD-3, respec-
       tively. E communicates with PDP PS-3. In general, it is expected that
       there will be at least one PDP per administrative domain.
       Policy capable network nodes could range from very unsophisticated,
       such as E, which have no LPDP, and thus have to rely on an external
       PDP for every policy processing operation, to self-sufficient, such
       as D, which essentially encompasses both an LPDP and a PDP locally,
       at the router.
                                AD-1                    AD-2        AD-3
              ________________/\_______________   __/\___      __/\___
             {                                 }  {       }     {       }
                    A            B            C            D            E
               +-------+   +-----+    +-------+    +-------+    +-------+
               | RSVP  |   | RSVP|    | RSVP  |    | RSVP  |    | RSVP  |
       +----+  |-------|   |-----|    |-------|    |-------|    |-------|
       | S1 |--| P | L |---|     |----| P | L |----| P | P |----|   P   |    +----+
       +----+  | E | D |   +-----+    | E | D |    | E | D |    |   E   |----| R1 |
               | P | P |              | P | P |    | P | P |    |   P   |    +----+
               +-------+              +-------+    +------+     +-------+
                  ^                         ^                           ^
                  |                         |                           |
                  |                         |                           |
                  |                         |                       +-------+
                  |                         |                       | PDP   |
                  |         +------+        |                       |-------|
                  +-------->| PDP  |<-------+                       |       |
                            |------|                                +-------+
                            |      |                                   PS-2
                Figure 4: Placement of Policy Elements in an internet
       5. Example Policies, Scenarios, and  Policy Support
       In the following, we present examples of desired policies and
       scenarios requiring policy control that should possibly be addressed
                                                                   [Page 13]

                A Framework for Policy-based Admission Control     Nov. 1998
       by the policy control framework. In some cases,  possible
       approach(es) for achieving the desired goals are also outlined with a
       list of open issues to be resolved.
       5.1. Admission control policies based on factors such as Time-of-Day,
       User Identity, or credentials
       Policy control must be able to express and enforce rules with tem-
       poral dependencies. For example, a group of users might be allowed to
       make reservations at certain levels only during off-peak hours.  In
       addition, the policy control must also support policies that take
       into account identity or credentials of users requesting a particular
       service or resource. For example, an RSVP reservation request may be
       denied or accepted based on the credentials or identity suppled in
       the request.
       5.2. Bilateral agreements between service providers
       Until recently, usage agreements between service providers for
       traffic crossing their boundaries have been quite simple. For exam-
       ple, two ISPs might agree to accept all traffic from each other,
       often without performing any accounting or billing for the
       ``foreign'' traffic carried.  However, with the availability of QoS
       mechanisms based on Integrated and Differentiated Services, traffic
       differentiation and quality of service guarantees are being phased
       into the Internet. As ISPs start to sell their customers different
       grades of service and can differentiate among different sources of
       traffic, they will also seek mechanisms for charging each other for
       traffic (and reservations) transiting their networks. One additional
       incentive in establishing such mechanisms is the potential asymmetry
       in terms of the customer base that different providers will exhibit:
       ISPs focused on servicing corporate traffic are likely to experience
       much higher demand for reserved services than those that service the
       consumer market. Lack of sophisticated accounting schemes for inter-
       ISP traffic could lead to inefficient allocation of costs among dif-
       ferent service providers.
       Bilateral agreements could fall into two broad categories; local or
       global. Due to the complexity of the problem, it is expected that
       initially only the former will be deployed. In these, providers which
       manage a network cloud or administrative domain contract with their
       closest point of contact (neighbor) to establish ground rules and
       arrangements for access control and accounting. These contracts are
       mostly local and do not rely on global agreements; consequently, a
       policy node maintains information about its neighboring nodes only.
       Referring to Figure 4, this model implies that provider AD-1 has
       established arrangements with AD-2, but not with AD-3, for usage of
                                                                   [Page 14]

                A Framework for Policy-based Admission Control     Nov. 1998
       each other's network. Provider AD-2, in turn, has in place agreements
       with AD-3 and so on. Thus, when forwarding a reservation request to
       AD-2, provider AD-2 will charge AD-1 for use of all resources beyond
       AD-1's network.  This information is obtained by recursively applying
       the bilateral agreements at every boundary between (neighboring) pro-
       viders, until the recipient of the reservation request is reached. To
       implement this scheme under the policy control architecture, boundary
       nodes have to add an appropriate policy object to the RSVP message
       before forwarding it to a neighboring provider's network. This policy
       object will contain information such as the identity of the provider
       that generated them and the equivalent of an account number where
       charges can be accumulated. Since agreements only hold among neigh-
       boring nodes, policy objects have to be rewritten as RSVP messages
       cross the boundaries of administrative domains or provider's net-
       5.3. Priority based admission control policies
       In many settings, it is useful to distinguish between reservations on
       the basis of some level of "importance".  For example, this can be
       useful to avoid that the first reservation being granted the use of
       some resources, be able to hog those resources for some indefinite
       period of time.  Similarly, this may be useful to allow emergency
       calls to go through even during periods of congestion.  Such func-
       tionality can be supported by associating priorities with reservation
       requests, and conveying this priority information together with other
       policy information.
       In its basic form, the priority associated with a reservation
       directly determines a reservation's rights to the resources it
       requests.  For example, assuming that priorities are expressed
       through integers in the range 0 to 32 with 32 being the highest
       priority, a reservation of priority, say, 10, will always be
       accepted, if the amount of resources held by lower priority reserva-
       tions is sufficient to satisfy its requirements.  In other words, in
       case there are not enough free resources (bandwidth, buffers, etc.)
       at a node to accommodate the priority 10 request, the node will
       attempt to free up the necessary resources by preempting existing
       lower priority reservations.
       There are a number of requirements associated with the support of
       priority and their proper operation.  First, traffic control in the
       router needs to be aware of priorities, i.e., classify existing
       reservations according to their priority, so that it is capable of
       determining how many and which ones to preempt, when required to
       accommodate a higher priority reservation request.  Second, it is
       important that preemption be made consistently at different nodes, in
       order to avoid transient instabilities.  Third and possibly most
                                                                   [Page 15]

                A Framework for Policy-based Admission Control     Nov. 1998
       important, merging of priorities needs to be carefully architected
       and its impact clearly understood as part of the associated policy
       Of the three above requirements, merging of priority information is
       the more complex and deserves additional discussions.  The complexity
       of merging priority information arises from the fact that this merg-
       ing is to be performed in addition to the merging of reservation
       information.  When reservation (FLOWSPEC) information is identical,
       i.e., homogeneous reservations, merging only needs to consider prior-
       ity information, and the simple rule of keeping the highest priority
       provides an adequate answer.  However, in the case of heterogeneous
       reservations, the * two-dimensional nature} of the (FLOWSPEC, prior-
       ity) pair makes their ordering, and therefore merging, difficult. A
       description of the handling of different cases of RSVP priority
       objects is presented in [7].
       5.4. Pre-paid calling card or Tokens
       A model of increasing popularity in the telephone network is that of
       the pre-paid calling card. This concept could also be applied to the
       Internet; users purchase ``tokens'' which can be redeemed at a later
       time for access to network services. When a user makes a reservation
       request through, say, an RSVP RESV message, the user supplies a
       unique identification number of the ``token'', embedded in a policy
       object. Processing of this object at policy capable routers results
       in decrementing the value, or number of remaining units of service,
       of this token.
       Referring to Figure 4, suppose receiver R1 in the administrative
       domain AD3 wants to request a reservation for a service originating
       in AD1. R1 generates a policy data object of type PD(prc, CID), where
       ``prc'' denotes pre-paid card and CID is the card identification
       number. Along with other policy objects carried in the RESV message,
       this object is received by node E, which forwards it to its PEP,
       PEP_E, which, in turn, contacts PDP PS-3. PS-3 either maintains
       locally, or has remote access to, a database of pre-paid card
       numbers. If the amount of remaining credit in CID is sufficient, the
       PDP accepts the reservation and the policy object is returned to
       PEP_E. Two issues have to be resolved here:
  *    What is the scope of these charges?
  *    When are charges (in the form of decrementing the remaining credit)
       first applied?
                                                                   [Page 16]

                A Framework for Policy-based Admission Control     Nov. 1998
  The answer to the first question is related to the bilateral agreement
  model in place. If, on the one hand, provider AD-3 has established agree-
  ments with both AD-2 and AD-1, it could charge for the cost of the com-
  plete reservation up to sender S1. In this case PS-2 removes the
  PD(prc,CID) object from the outgoing RESV message.
  On the other hand, if AD-3 has no bilateral agreements in place, it will
  simply charge CID for the cost of the reservation within AD-3 and then
  forward PD(prc,CID) in the outgoing RESV message. Subsequent PDPs in other
  administrative domains will charge CID for their respective reservations.
  Since multiple entities are both reading (remaining credit) and writing
  (decrementing credit) to the same database, some coordination and con-
  currency control might be needed.  The issues related to location, manage-
  ment, coordination of credit card (or similar) databases is outside the
  scope of this document.
  Another problem in this scenario is determining when the credit is
  exhausted. The PDPs should contact the database periodically to submit a
  charge against the CID; if the remaining credit reaches zero, there must
  be a mechanism to detect that and to cause revocation or termination of
  privileges granted based on the credit.
  Regarding the issue of when to initiate charging, ideally that should hap-
  pen only after the reservation request has succeeded. In the case of local
  charges, that could be communicated by the router to the PDP.
  5.5. Sender Specified Restrictions on Receiver Reservations
  The ability of senders to specify restrictions on reservations, based on
  receiver identity, number of receivers or reservation cost might be useful
  in future network applications. An example could be any application in
  which the sender pays for service delivered to receivers. In such a case,
  the sender might be willing to assume the cost of a reservation, as long
  as it satisfies certain criteria, for example, it originates from a
  receiver who belongs to an access control list (ACL) and satisfies a limit
  on cost. (Notice that this could allow formation of "closed" multicast
  In the policy based admission control framework such a scheme could be
  achieved by having the sender generate appropriate policy objects, carried
  in a PATH message, which install state in routers on the path to
  receivers. In accepting reservations, the routers would have to compare
  the RESV requests to the installed state.
  A number of different solutions can be built to address this scenario;
  precise description of a solution is beyond the scope of this document.
                                                                   [Page 17]

                A Framework for Policy-based Admission Control     Nov. 1998
  6. Interaction Between the Policy Enforcement Point (PEP) and the Policy
  Decision Point (PDP)
  In the case of an external PDP, the need for a communication protocol
  between the PEP and PDP arises. In order to allow for interoperability
  between different vendors networking elements and (external) policy
  servers, this protocol should be standardized.
  6.1. PEP to PDP Protocol Requirements
  This section describes a set of general requirements for the communication
  protocol between the PEP and an external PDP.
  *    Reliability:  The sensitivity of policy control information necessi-
       tates reliable operation. Undetected loss of policy queries or
       responses may lead to inconsistent network control operation and are
       clearly unacceptable for actions such as billing and accounting. One
       option for providing reliability is the re-use of the TCP as the
       transport protocol.
  *    Small delays: The timing requirements of policy decisions related to
       QoS signaling protocols are expected to be quite strict. The PEP to
       PDP protocol should add small amount of delay to the response delay
       experienced by queries placed by the PEP to the PDP.
  *    Ability to carry opaque objects: The protocol should allow for
       delivery of self-identifying, opaque objects, of variable length,
       such as RSVP messages, RSVP policy objects and other objects that
       might be defined as new policies are introduced. The protocol should
       not have to be changed every time a new object has to be exchanged.
  *    Support for PEP-initiated, two-way Transactions:  The protocol must
       allow for two-way transactions (request-response exchanges) between a
       PEP and a PDP. In particular, PEPs must be able to initiate requests
       for policy decision, re-negotiation of previously made policy deci-
       sion, and exchange of policy information. To some extent, this
       requirement is closely tied to the goal of meeting the requirements
       of RSVP-specific, policy-based admission control. RSVP signaling
       events such as arrival of RESV refresh messages, state timeout, and
       merging of reservations require that a PEP (such as an RSVP router)
       request a policy decision from PDP at any time. Similarly, PEP must
       be able to report monitoring information and policy state changes to
       PDP at any time.
                                                                   [Page 18]

                A Framework for Policy-based Admission Control     Nov. 1998
  *    Support for asynchronous notification: This is required in order to
       allow both the policy server and client to notify each other in the
       case of an asynchronous change in state, i.e., a change that is not
       triggered by a signaling message. For example, the server would need
       to notify the client if a particular reservation has to be terminated
       due to expiration of a user's credentials or account balance. Like-
       wise, the client has to inform the server of a reservation rejection
       which is due to admission control failure.
  *    Handling of multicast groups: The protocol should provision for han-
       dling of policy decisions related to multicast groups.
  *    QoS Specification: The protocol should allow for precise specifica-
       tion of level of service requirements in the PEP requests forwarded
       to the PDP.
       6.2. Evaluation of Existing Protocols
       In view of the above requirements, we believe that protocols
       developed earlier within a different context are inadequate for pro-
       viding policy support for QoS. In the following sections, we summar-
       ize the reasons for some candidate protocols.
       6.2.1. RADIUS
       The Remote Authentication Dial In User Service (RADIUS) [5] protocol,
       specifies how authentication, authorization and configuration infor-
       mation is exchanged between a network access server wishing to
       authenticate its users and a (shared) authentication server. RADIUS
       has been designed to operate at the level of a "session", which is
       defined as the interval between the time that service is first pro-
       vided and the time service is ended. The latter definition is geared
       towards login-type service, for example network access for dial up
       users through PPP.
       RADIUS is an important protocol for authenticating login users. How-
       ever, some of its characteristics make it unsuitable for use, at
       least in its current form, in the context of policy control for QoS
    *    Use of the UDP protocol: As mentioned in section 6.1, the sensi-
         tivity of policy control information requires reliable operation.
         Use of UDP would mandate specifying retry and fallback algorithms,
         which can be quite complicated. In addition, while timing
                                                                   [Page 19]

                A Framework for Policy-based Admission Control     Nov. 1998
         requirements for a user logging onto a network might permit a delay
         of several seconds ([5]), this will not be the case for reserva-
         tions proceeding in the network. Moreover, acknowledgments are
         essential for actions like billing and accounting.
    *    Difficulty in carrying opaque objects: Use of RADIUS for QoS policy
         control would require defining additional RADIUS * attributes. To
         carry opaque objects of variable length, the format of RADIUS
         attributes would have to be extended; it currently supports four
         data types of pre-specified length; string, address, integer and
    *    No support for asynchronous notification: Asynchronous notification
         is required in order to allow both the PDP and a PEP to notify each
         other in the case of an asynchronous change in state, i.e., a
         change that is not triggered by a signaling message. For example,
         the server would need to notify the client if a particular reserva-
         tion has to be terminated due to expiration of a user's credentials
         or account balance. Likewise, the client has to inform the server
         of a reservation rejection which is due to admission control
    *    Restricted Message Types: The restriction to messages of type
         ``access-request'' and ``access-accept'' (or reject) does not
         facilitate information gathering for monitoring and accounting pur-
         poses. New message types would have to be defined.
    *    Multicast groups: It is not clear how requests from a group of mul-
         ticast receivers could be handled in the context of RADIUS.
    *    Identifying Requests: Several changes will be needed in the way
         requests are identified. Currently this is done using a user
         name/IP address attribute. This might be too narrow for reservation
         protocols, i.e. one might need to specify a source port, a destina-
         tion address/port and provide some user credentials as well.
    *    QoS Specification: Changes will be needed to incorporate the new
         service (resource reservation) and the ability to specify the
         desired level of service (QoS, Tspec).
                                                                   [Page 20]

                A Framework for Policy-based Admission Control     Nov. 1998
  In short, we believe that if the RADIUS protocol were to be used for the
  purpose of QoS policy control, the number of changes required would make
  the task as difficult or more than defining a new protocol designed with
  QoS reservations in mind.
  6.2.2. LDAP
  LDAP is very effective as a directory service protocol. Nevertheless, it
  restricts the flexibility of an implementation by requiring a very
  specific and well understood format of policy information and policy
  rules, common to both a policy server and a client (e.g., a standardized
  schema). The client must be able to interpret the format of policy infor-
  mation and rules directly.  In addition, LDAP is more suitable for access-
  ing directory information, information that is predefined and changes
  Given the rather experimental state of policy control for QoS reservation
  protocols, the dynamic nature of policy decisions and rules and the limi-
  tations of clients (routers) in interpreting policy information, an LDAP
  based solution for client-server communication would be inadequate at this
  point. Also, a wider variety of policy element types and objects need to
  be represented using a policy control protocol than supported by LDAP.
  LDAP currently does not provide support for asynchronous notification from
  servers to clients and is not suitable for exchange of dynamic information
  such as policy state and resource usage information.
  However, LDAP could conceivably be used in downloading policy rules from
  an LDAP server to policy servers. Furthermore, future versions of LDAP
  promise to make it more suitable for the exchange of dynamic information.
  As these modifications are introduced the role of LDAP in QoS policy con-
  trol could be reevaluated.
  6.2.3. SNMP
  SNMP was designed for the specific purpose of monitoring network devices
  from a central network management station (NMS). Our investigation of SNMP
  led us to believe that it does not provide adequate support for the pur-
  pose of communication between a PEP and its PDP for the following reasons:
  *    Lack of support for large messages: Policy information carried in
       messages exchanged between a PEP and a PDP typically consists of
       several policy elements. Each policy element itself may contain siz-
       able amount of information. For example, a Kerberos-based credential
       (a single policy element) constitutes as many as 300-plus bytes of
       data. A single request for policy decision will contain several such
                                                                   [Page 21]

                A Framework for Policy-based Admission Control     Nov. 1998
       elements and some  additional information such as flowspecs and,
       thus, will add up to a large message. Similarly, a report message
       containing monitoring information will also contain a large amount of
       information. Therefore, the protocol must support reliable delivery
       of large messages whereas SNMP only supports delivery of small mes-
  *    Use of UDP as a transport protocol: SNMP uses UDP as the transport
       protocol and, as discussed earlier, UDP is not appropriate as a tran-
       sport protocol for this purpose and requires an additional protocol
       to ensure reliable delivery of large messages.
  *    Master-Slave relationship: SNMP is designed for a Master-slave rela-
       tionship in which a Network Management Station (NMS) acts as a master
       for a group of devices. NMS is always in charge and initiates any
       communication for retrieval of monitoring information except for
       traps that can be asynchronously generated by the slave devices. The
       relationship between a PEP and its PDP, however, is not master-slave;
       PEP always initiates communication, requests decisions, or  sends
       reports, or re-negotiates a decision. In addition, the exchange
       between a PEP and PDP is stateful and establishes shared state that
       is referred to in subsequent communication. SNMP does not support
       this style of communication.
  *    Lack of support for policy-specific errors:  SNMP provides support
       for generic errors whereas we need support for reporting errors
       specific to policy data content (authentication failure) as well as
       errors specific to RSVP semantics.
  7. Security Considerations
  The communication tunnel between policy clients and policy servers should
  be secured by the use of an IPSEC [4] channel. It is advisable that this
  tunnel makes use of both the AH (Authentication Header) and ESP (Encapsu-
  lating Security Payload) protocols, in order to provide confidentiality,
  data origin authentication, integrity and replay prevention.
  In the case of the RSVP signaling mechanism, RSVP MD5 [2] message authen-
  tication can be used to secure communications between network elements.
  8. References
                                                                   [Page 22]

                A Framework for Policy-based Admission Control     Nov. 1998
  [1] R. Braden, L. Zhang, S. Berson, S. Herzog, S. Jamin, "Resource ReSer-
  Vation Protocol (RSVP) -- Version 1 Functional Specification ", RFC 2205,
  September 1997.
  [2] F. Baker., "RSVP Cryptographic Authentication", draft-ietf-rsvp-md5-
  05.txt, August 1997.
  [3] S. Herzog., "RSVP Extensions for Policy Control",  Internet Draft},
  draft-ietf-rsvp-policy-ext-03.[ps,txt], August 1998.
  [4] R. Atkinson. Security Architecture for the Internet Protocol. RFC1825,
  Aug. 1995.
  [5] C. Rigney, A Rubens, W. Simpson and S. Willens. Remote Authentication
  Dial In User Service (RADIUS). RFC 2138.
  [6] R. Rajan et al. Schema for Differentiated Services and Integrated Ser-
  vices in Networks, draft-rajan-policy-qosschema-00.txt, October 1998.
  [7] S. Herzog, "RSVP Preemption Priority Policy", Internet Draft, draft-
  ietf-rap-priority-00.txt, Nov. 1998.
  [8] S. Herzog, "COPS Usage for RSVP", Internet Draft, draft-ietf-rap-
  cops-rsvp-00.txt, August 1998.
  8. Acknowledgements
  This is a result of an ongoing discussion among many members of the RAP
  group including Jim Boyle, Ron Cohen, Laura Cunningham, Dave Durham, Shai
  Herzog, Tim O'Malley, Raju Rajan, and Arun Sastry.
  9.  Authors` Addresses
          Raj Yavatkar
          Intel Corporation
          2111 N.E. 25th Avenue,
          Hillsboro, OR 97124
          phone: +1 503-264-9077
          Dimitrios Pendarakis
          IBM T.J. Watson Research Center
          P.O. Box 704
          Yorktown Heights
          NY 10598
          phone: +1 914-784-7536
                                                                   [Page 23]

                A Framework for Policy-based Admission Control     Nov. 1998
          Roch Guerin
          University of Pennsylvania
          Dept. of Electrical Engineering
          200 South 33rd Street
          Philadelphia, PA  19104
          phone: +1 215 898-9351
                                                                   [Page 24]