+-----------------+ | Local | | Applications or | |Management (e.g.,| | for aggregates) | +-----------------+ ^ V V +----------+ +----------+ +---------+ | QoS NSLP | | Resource | | Policy | |Processing|<<<<<<>>>>>>>|Management|<<<>>>| Control | +----------+ +----------+ +---------+ . ^ | * ^ | V . * ^ +----------+ * ^ | NTLP | * ^ |Processing| * V +----------+ * V | | * V ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ . . * V | | * ............................. . . * . Traffic Control . | | * . +---------+. . . * . |Admission|. | | * . | Control |. +----------+ +------------+ . +---------+. <-.-| Input | | Outgoing |-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-> | Packet | | Interface | .+----------+ +---------+. ===>|Processing|====| Selection |===.| Packet |====| Packet |.==> | | |(Forwarding)| .|Classifier| Scheduler|. +----------+ +------------+ .+----------+ +---------+. ............................. <.-.-> = signaling flow =====> = data flow (sender --> receiver) <<<>>> = control and configuration operations ****** = routing table manipulation Figure 2: QoS NSLP in a Node This diagram shows an example implementation scenario where QoS conditioning is performed on the output interface. However, this does not limit the possible implementations. For example, in some cases, traffic conditioning may be performed on the incoming interface, or it may be split over the input and output interfaces.
Also, the interactions with the Policy Control component may be more complex, involving interaction with the Resource Management Function, and the AAA infrastructure. From the perspective of a single node, the request for QoS may result from a local application request or from processing an incoming QoS NSLP message. The request from a local application includes not only user applications but also network management and the policy control module. For example, a request could come from multimedia applications, initiate a tunnel to handle an aggregate, interwork with some other reservation protocol (such as RSVP), and contain an explicit teardown triggered by a AAA policy control module. In this sense, the model does not distinguish between hosts and routers. Incoming messages are captured during input packet processing and handled by GIST. Only messages related to QoS are passed to the QoS NSLP. GIST may also generate triggers to the QoS NSLP (e.g., indications that a route change has occurred). The QoS request is handled by the RMF, which coordinates the activities required to grant and configure the resource. It also handles policy-specific aspects of QoS signaling. The grant processing involves two local decision modules, 'policy control' and 'admission control'. Policy control determines whether the user is authorized to make the reservation. Admission control determines whether the network of the node has sufficient available resources to supply the requested QoS. If both checks succeed, parameters are set in the packet classifier and in the link-layer interface (e.g., in the packet scheduler) to obtain the desired QoS. Error notifications are passed back to the request originator. The Resource Management Function may also manipulate the forwarding tables at this stage to select (or at least pin) a route; this must be done before interface-dependent actions are carried out (including sending outgoing messages over any new route), and is in any case invisible to the operation of the protocol. Policy control is expected to make use of the authentication infrastructure or the authentication protocols external to the node itself. Some discussion can be found in a separate document on authorization issues [qos-auth]. More generally, the processing of policy and Resource Management Functions may be outsourced to an external node, leaving only 'stubs' co-located with the NSLP node; this is not visible to the protocol operation. A more detailed discussion of authentication and authorization can be found in Section 3.1.3.
Admission control, packet scheduling, and any part of policy control beyond simple authorization have to be implemented using specific definitions for types and levels of QoS. A key assumption is made that the QoS NSLP is independent of the QoS parameters (e.g., IntServ service elements). These are captured in a QoS model and interpreted only by the resource management and associated functions, and are opaque to the QoS NSLP itself. QoS models are discussed further in Section 3.1.2. The final stage of processing for a resource request is to indicate to the QoS NSLP protocol processing that the required resources have been configured. The QoS NSLP may generate an acknowledgment message in one direction, and may forward the resource request in the other. Message routing is carried out by the GIST module. Note that while Figure 2 shows a unidirectional data flow, the signaling messages can pass in both directions through the node, depending on the particular message and orientation of the reservation. 3.1.1. Protocol Messages The QoS NSLP uses four message types: RESERVE: The RESERVE message is the only message that manipulates QoS NSLP reservation state. It is used to create, refresh, modify, and remove such state. The result of a RESERVE message is the same whether a message is received once or many times. QUERY: A QUERY message is used to request information about the data path without making a reservation. This functionality can be used to make reservations or to support certain QoS models. The information obtained from a QUERY may be used in the admission control process of a QNE (e.g., in case of measurement-based admission control). Note that a QUERY does not change existing reservation state. RESPONSE: The RESPONSE message is used to provide information about the result of a previous QoS NSLP message. This includes explicit confirmation of the state manipulation signaled in the RESERVE message, and the response to a QUERY message or an error code if the QNE or QNR is unable to provide the requested information or if the response is negative. The RESPONSE message does not cause any reservation state to be installed or modified. NOTIFY: NOTIFY messages are used to convey information to a QNE. They differ from RESPONSE messages in that they are sent asynchronously and need not refer to any particular state or previously received message. The information conveyed by a NOTIFY
message is typically related to error conditions. Examples would be notification to an upstream peer about state being torn down or notification when a reservation has been preempted. QoS NSLP messages are sent peer-to-peer. This means that a QNE considers its adjacent upstream or downstream peer to be the source of each message. Each protocol message has a common header which indicates the message type and contains various flag bits. Message formats are defined in Section 5.1.2. Message processing rules are defined in Section 5.4. QoS NSLP messages contain three types of objects: 1. Control Information: Control information objects carry general information for the QoS NSLP processing, such as sequence numbers or whether a response is required. 2. QoS specifications (QSPECs): QSPEC objects describe the actual resources that are required and depend on the QoS model being used. Besides any resource description, they may also contain other control information used by the RMF's processing. 3. Policy objects: Policy objects contain data used to authorize the reservation of resources. Object formats are defined in Section 5.1.3. Object processing rules are defined in Section 5.3. 3.1.2. QoS Models and QoS Specifications The QoS NSLP provides flexibility over the exact patterns of signaling messages that are exchanged. The decoupling of QoS NSLP and QSPEC allows the QoS NSLP to be ignorant about the ways in which traffic, resources, etc., are described, and it can treat the QSPEC as an opaque object. Various QoS models can be designed, and these do not affect the specification of the QoS NSLP protocol. Only the RMF specific to a given QoS model will need to interpret the QSPEC. The Resource Management Function (RMF) reserves resources for each flow. The QSPEC fulfills a similar purpose to the TSpec, RSpec, and AdSpec objects used with RSVP and specified in RFC 2205 [RFC2205] and RFC 2210 [RFC2210]. At each QNE, the content of the QSPEC is interpreted by the Resource Management Function and the Policy Control Function for the purposes of traffic and policy control (including admission control and configuration of the packet classifier and scheduler).
The QoS NSLP does not mandate any particular behavior for the RMF, instead providing interoperability at the signaling-protocol level whilst leaving the validation of RMF behavior to contracts external to the protocol itself. The RMF may make use of various elements from the QoS NSLP message, not only the QSPEC object. Still, this specification assumes that resource sharing is possible between flows with the same SESSION-ID that originate from the same QNI or between flows with a different SESSION-ID that are related through the BOUND-SESSION-ID object. For flows with the same SESSION-ID, resource sharing is only applicable when the existing reservation is not just replaced (which is indicated by the REPLACE flag in the common header). We assume that the QoS model supports resource sharing between flows. A QoS Model may elect to implement a more general behavior of supporting relative operations on existing reservations, such as ADDING or SUBTRACTING a certain amount of resources from the current reservation. A QoS Model may also elect to allow resource sharing more generally, e.g., between all flows with the same Differentiated Service Code Point (DSCP). The QSPEC carries a collection of objects that can describe QoS specifications in a number of different ways. A generic template is defined in [RFC5975] and contains object formats for generally useful elements of the QoS description, which is designed to ensure interoperability when using the basic set of objects. A QSPEC describing the resources requested will usually contain objects that need to be understood by all implementations, and it can also be enhanced with additional objects specific to a QoS model to provide a more exact definition to the RMF, which may be better able to use its specific resource management mechanisms (which may, e.g., be link specific) as a result. A QoS Model defines the behavior of the RMF, including inputs and outputs, and how QSPEC information is used to describe resources available, resources required, traffic descriptions, and control information required by the RMF. A QoS Model also describes the minimum set of parameters QNEs should use in the QSPEC when signaling about this QoS Model. QoS Models may be local (private to one network), implementation/ vendor specific, or global (implementable by different networks and vendors). All QSPECs should follow the design of the QSPEC template. The definition of a QoS model may also have implications on how local behavior should be implemented in the areas where the QoS NSLP gives freedom to implementers. For example, it may be useful to identify recommended behavior for how a forwarded RESERVE message relates to a received one, or for when additional signaling sessions should be
started based on existing sessions, such as required for aggregate reservations. In some cases, suggestions may be made on whether state that may optionally be retained should be held in particular scenarios. A QoS model may specify reservation preemption, e.g., an incoming resource request may cause removal of an earlier established reservation. 3.1.3. Policy Control Getting access to network resources, e.g., network access in general or access to QoS, typically involves some kind of policy control. One example of this is authorization of the resource requester. Policy control for QoS NSLP resource reservation signaling is conceptually organized as illustrated below in Figure 3. +-------------+ | Policy | | Decision | | Point (PDP) | +------+------+ | /-\-----+-----/\ //// \\\\ || || | Policy transport | || || \\\\ //// \-------+------/ | +-------------+ QoS signaling +------+------+ | Entity |<==============>| QNE = Policy|<=========> | requesting | Data Flow | Enforcement | | resource |----------------|-Point (PEP)-|----------> +-------------+ +-------------+ Figure 3: Policy Control with the QoS NSLP Signaling From the QoS NSLP point of view, the policy control model is essentially a two-party model between neighboring QNEs. The actual policy decision may depend on the involvement of a third entity (the Policy Decision Point, PDP), but this happens outside of the QoS NSLP protocol by means of existing policy infrastructure (Common Open Policy Service (COPS), Diameter, etc.). The policy control model for the entire end-to-end chain of QNEs is therefore one of transitivity, where each of the QNEs exchanges policy information with its QoS NSLP policy peer.
The authorization of a resource request often depends on the identity of the entity making the request. Authentication may be required. The GIST channel security mechanisms provide one way of authenticating the QoS NSLP peer that sent the request, and so may be used in making the authorization decision. Additional information might also be provided in order to assist in making the authorization decision. This might include alternative methods of authenticating the request. The QoS NSLP does not currently contain objects to carry authorization information. At the time of writing, there exists a separate individual work [NSIS-AUTH] that defines this functionality for the QoS NSLP and the NAT and firewall (NATFW) NSLP. It is generally assumed that policy enforcement is likely to concentrate on border nodes between administrative domains. This may mean that nodes within the domain are "Policy-Ignorant Nodes" that perform no per-request authentication or authorization, relying on the border nodes to perform the enforcement. In such cases, the policy management between ingress and egress edge of a domain relies on the internal chain of trust between the nodes in the domain. If this is not acceptable, a separate signaling session can be set up between the ingress and egress edge nodes in order to exchange policy information. 3.2. Design Background This section presents some of the key functionality behind the specification of the QoS NSLP. 3.2.1. Soft States The NSIS protocol suite takes a soft-state approach to state management. This means that reservation state in QNEs must be periodically refreshed. The frequency with which state installation is refreshed is expressed in the REFRESH-PERIOD object. This object contains a value in milliseconds indicating how long the state that is signaled for remains valid. Maintaining the reservation beyond this lifetime can be done by sending a RESERVE message periodically. 3.2.2. Sender and Receiver Initiation The QoS NSLP supports both sender-initiated and receiver-initiated reservations. For a sender-initiated reservation, RESERVE messages travel in the same direction as the data flow that is being signaled for (the QNI is at the side of the source of the data flow). For a
receiver-initiated reservation, RESERVE messages travel in the opposite direction (the QNI is at the side of the receiver of the data flow). Note: these definitions follow the definitions in Section 3.3.1 of RFC 4080 [RFC4080]. The main issue is about which node is in charge of requesting and maintaining the resource reservation. In a receiver-initiated reservation, even though the sender sends the initial QUERY, the receiver is still in charge of making the actual resource request and maintaining the reservation. 3.2.3. Protection against Message Re-ordering and Duplication RESERVE messages affect the installed reservation state. Unlike NOTIFY, QUERY, and RESPONSE messages, the order in which RESERVE messages are received influences the eventual reservation state that will be stored at a QNE; that is, the most recent RESERVE message replaces the current reservation. Therefore, in order to protect against RESERVE message re-ordering or duplication, the QoS NSLP uses a Reservation Sequence Number (RSN). The RSN has local significance only, i.e., between a QNE and its downstream peers. 3.2.4. Explicit Confirmations A QNE may require a confirmation that the end-to-end reservation is in place, or a reply to a query along the path. For such requests, it must be able to keep track of which request each response refers to. This is supported by including a Request Identification Information (RII) object in a QoS NSLP message. 3.2.5. Reduced Refreshes For scalability, the QoS NSLP supports an abbreviated form of refresh RESERVE message. In this case, the refresh RESERVE references the reservation using the RSN and the SESSION-ID, and does not include the full reservation specification (including QSPEC). By default, state refresh should be performed with reduced refreshes in order to save bytes during transmission. Stateless QNEs will require full refresh since they do not store the whole reservation information. If the stateful QNE does not support reduced refreshes, or there is a mismatch between the local and received RSN, the stateful QNE must reply with a RESPONSE carrying an INFO-SPEC indicating the error. Furthermore, the QNE must stop sending reduced refreshes to this peer if the error indicates that support for this feature is lacking.
3.2.6. Summary Refreshes and Summary Tear For limiting the number of individual messages, the QoS NSLP supports summary refresh and summary tear messages. When sending a refreshing RESERVE for a certain (primary) session, a QNE may include a SESSION- ID-LIST object where the QNE indicates (secondary) sessions that are also refreshed. An RSN-LIST object must also be added. The SESSION- IDs and RSNs are stacked in the objects such that the index in both stacks refer to the same reservation state, i.e., the SESSION-ID and RSN at index i in both objects refers to the same session. If the receiving stateful QNE notices unknown SESSION-IDs or a mismatch with RSNs for a session, it will reply back to the upstream stateful QNE with an error. In order to tear down several sessions at once, a QNE may include SESSION-ID-LIST and RSN-LIST objects in a tearing reserve. The downstream stateful QNE must then also tear down the other sessions indicated. The downstream stateful QNE must silently ignore any unknown SESSION-IDs. GIST provides a SII-Handle for every downstream session. The SII- Handle identifies a peer and should be the same for all sessions whose downstream peer is the same. The QoS NSLP uses this information to decide whether summary refresh messages can be sent or when a summary tear is possible. 3.2.7. Message Scoping A QNE may use local policy when deciding whether to propagate a message or not. For example, the local policy can define/configure that a QNE is, for a particular session, a QNI and/or a QNR. The QoS NSLP also includes an explicit mechanism to restrict message propagation by means of a scoping mechanism. For a RESERVE or a QUERY message, two scoping flags limit the part of the path on which state is installed on the downstream nodes that can respond. When the SCOPING flag is set to zero, it indicates that the scope is "whole path" (default). When set to one, the scope is "single hop". When the PROXY scope flag is set, the path is terminated at a pre-defined Proxy QNE (P-QNE). This is similar to the Localized RSVP [lrsvp]. The propagation of a RESPONSE message is limited by the RII object, which ensures that it is not forwarded back along the path further than the node that requested the RESPONSE.
3.2.8. Session Binding Session binding is defined as the enforcement of a relation between different QoS NSLP sessions (i.e., signaling flows with different SESSION-IDs (SIDs) as defined in GIST [RFC5971]). Session binding indicates a unidirectional dependency relation between two or more sessions by including a BOUND-SESSION-ID object. A session with SID_A (the binding session) can express its unidirectional dependency relation to another session with SID_B (the bound session) by including a BOUND-SESSION-ID object containing SID_B in its messages. The concept of session binding is used to indicate the unidirectional dependency relation between the end-to-end session and the aggregate session in case of aggregate reservations. In case of bidirectional reservations, it is used to express the unidirectional dependency relation between the sessions used for forward and reverse reservation. Typically, the dependency relation indicated by session binding is purely informative in nature and does not automatically trigger any implicit action in a QNE. A QNE may use the dependency relation information for local resource optimization or to explicitly tear down reservations that are no longer useful. However, by using an explicit binding code (see Section 188.8.131.52), it is possible to formalize this dependency relation, meaning that if the bound session (e.g., session with SID_B) is terminated, the binding session (e.g., the session with SID_A) must be terminated also. A message may include more than one BOUND-SESSION-ID object. This may happen, e.g., in certain aggregation and bidirectional reservation scenarios, where an end-to-end session has a unidirectional dependency relation with an aggregate session and at the same time it has a unidirectional dependency relation with another session used for the reverse path. 3.2.9. Message Binding QoS NSLP supports binding of messages in order to allow for expressing dependencies between different messages. The message binding can indicate either a unidirectional or bidirectional dependency relation between two messages by including the MSG-ID object in one message ("binding message") and the BOUND-MSG-ID object in the other message ("bound message"). The unidirectional dependency means that only RESERVE messages are bound to each other whereas a bidirectional dependency means that there is also a dependency for the related RESPONSE messages. The message binding can be used to speed up signaling by starting two signaling exchanges simultaneously that are synchronized later by using message IDs.
This can be used as an optimization technique, for example, in scenarios where aggregate reservations are used. Section 4.6 provides more details. 3.2.10. Layering The QoS NSLP supports layered reservations. Layered reservations may occur when certain parts of the network (domains) implement one or more local QoS models or when they locally apply specific transport characteristics (e.g., GIST unreliable transfer mode instead of reliable transfer mode). They may also occur when several per-flow reservations are locally combined into an aggregate reservation. 184.108.40.206. Local QoS Models A domain may have local policies regarding QoS model implementation, i.e., it may map incoming traffic to its own locally defined QoS models. The QSPEC allows this functionality, and the operation is transparent to the QoS NSLP. The use of local QoS models within a domain is performed in the RMF. 220.127.116.11. Local Control Plane Properties The way signaling messages are handled is mainly determined by the parameters that are sent over the GIST-NSLP API and by the domain internal configuration. A domain may have a policy to implement local transport behavior. It may, for instance, elect to use an unreliable transport locally in the domain while still keeping end- to-end reliability intact. The QoS NSLP supports this situation by allowing two sessions to be set up for the same reservation. The local session has the desired local transport properties and is interpreted in internal QNEs. This solution poses two requirements: the end-to-end session must be able to bypass intermediate nodes, and the egress QNE needs to bind both sessions together. Bypassing intermediate nodes is achieved with GIST. The local session and the end-to-end session are bound at the egress QNE by means of the BOUND-SESSION-ID object. 18.104.22.168. Aggregate Reservations In some cases, it is desirable to create reservations for an aggregate, rather than on a per-flow basis, in order to reduce the amount of reservation state needed as well as the processing load for signaling messages. Note that the QoS NSLP does not specify how reservations need to be combined in an aggregate or how end-to-end properties need to be computed, but only provides signaling support for aggregate reservations.
The essential difference with the layering approaches described in Sections 22.214.171.124 and 126.96.36.199 is that the aggregate reservation needs a MRI that describes all traffic carried in the aggregate (e.g., a DSCP in case of IntServ over Diffserv). The need for a different MRI mandates the use of two different sessions, as described in Section 188.8.131.52 and in the RSVP aggregation solution in RFC 3175 [RFC3175]. Edge QNEs of the aggregation domain that want to maintain some end- to-end properties may establish a peering relation by sending the end-to-end message transparently over the domain (using the intermediate node bypass capability described above). Updating the end-to-end properties in this message may require some knowledge of the aggregated session (e.g., for updating delay values). For this purpose, the end-to-end session contains a BOUND-SESSION-ID carrying the SESSION-ID of the aggregate session. 3.2.11. Support for Request Priorities This specification acknowledges the fact that in some situations, some messages or reservations may be more important than others, and therefore it foresees mechanisms to give these messages or reservations priority. Priority of certain signaling messages over others may be required in mobile scenarios when a message loss during call setup is less harmful than during handover. This situation only occurs when GIST or QoS NSLP processing is the congested part or scarce resource. Priority of certain reservations over others may be required when QoS resources are oversubscribed. In that case, existing reservations may be preempted in order to make room for new higher-priority reservations. A typical approach to deal with priority and preemption is through the specification of a setup priority and holding priority for each reservation. The Resource Management Function at each QNE then keeps track of the resource consumption at each priority level. Reservations are established when resources, at their setup priority level, are still available. They may cause preemption of reservations with a lower holding priority than their setup priority. Support of reservation priority is a QSPEC parameter and therefore outside the scope of this specification. The GIST specification provides a mechanism to support a number of levels of message priority that can be requested over the NSLP-GIST API.
3.2.12. Rerouting The QoS NSLP needs to adapt to route changes in the data path. This assumes the capability to detect rerouting events, create a QoS reservation on the new path, and optionally tear down reservations on the old path. From an NSLP perspective, rerouting detection can be performed in two ways. It can either come through NetworkNotification from GIST, or from information seen at the NSLP. In the latter case, the QoS NSLP node is able to detect changes in its QoS NSLP peers by keeping track of a Source Identification Information (SII) handle that provides information similar in nature to the RSVP_HOP object described in RFC 2205 [RFC2205]. When a RESERVE message with an existing SESSION-ID and a different SII is received, the QNE knows its upstream or downstream peer has changed, for sender-oriented and receiver- oriented reservations, respectively. Reservation on the new path happens when a RESERVE message arrives at the QNE beyond the point where the old and new paths diverge. If the QoS NSLP suspects that a reroute has occurred, then a full RESERVE message (including the QSPEC) would be sent. A refreshing RESERVE (with no QSPEC) will be identified as an error by a QNE on the new path, which does not have the reservation installed (i.e., it was not on the old path) or which previously had a different previous-hop QNE. It will send back an error message that results in a full RESERVE message being sent. Rapid recovery at the NSLP layer therefore requires short refresh periods. Detection before the next RESERVE message arrives is only possible at the IP layer or through monitoring of GIST peering relations (e.g., by monitoring the Time to Live (TTL), i.e., the number of GIST hops between NSLP peers, or observing the changes in the outgoing interface towards GIST peer). These mechanisms can provide implementation-specific optimizations and are outside the scope of this specification. When the QoS NSLP is aware of the route change, it needs to set up the reservation on the new path. This is done by sending a new RESERVE message. If the next QNE is in fact unchanged, then this will be used to refresh/update the existing reservation. Otherwise, it will lead to the reservation being installed on the new path. Note that the operation for a receiver-initiated reservation session differs a bit from the above description. If the routing changes in the middle of the path, at some point (i.e., the divergence point) the QNE that notices that its downstream path has changed (indicated by a NetworkNotification from GIST), and it must send a QUERY with the R-flag downstream. The QUERY will be processed as above, and at some point hits a QNE for which the path downstream towards the QNI
remains (i.e., the convergence point). This node must then send a full RESERVE upstream to set up the reservation state along the new path. It should not send the QUERY further downstream, since this would have no real use. Similarly, when the QNE that sent the QUERY receives the RESERVE, it should not send the RESERVE further upstream. After the reservation on the new path is set up, the branching node may want to tear down the reservation on the old path (sooner than would result from normal soft-state timeout). This functionality is supported by keeping track of the old SII-Handle provided over the GIST API. This handle can be used by the QoS NSLP to route messages explicitly to the next node. If the old path is downstream, the QNE can send a tearing RESERVE using the old SII-Handle. If the old path is upstream, the QNE can send a NOTIFY with the code for "Route Change". This is forwarded upstream until it hits a QNE that can issue a tearing RESERVE downstream. A separate document discusses in detail the effect of mobility on the QoS NSLP signaling [NSIS-MOB]. A QNI or a branch node may wish to keep the reservation on the old branch. For instance, this could be the case when a mobile node has experienced a mobility event and wishes to keep reservation to its old attachment point in case it moves back there. For this purpose, a REPLACE flag is provided in the QoS NSLP common header, which, when not set, indicates that the reservation on the old branch should be kept. Note that keeping old reservations affects the resources available to other nodes. Thus, the operator of the access network must make the final decision on whether this behavior is allowed. Also, the QNEs in the access network may add this flag even if the mobile node has not used the flag initially. The latency in detecting that a new downstream peer exists or that truncation has happened depends on GIST. The default QUERY message transmission interval is 30 seconds. More details on how NSLPs are able to affect the discovery of new peers or rerouting can be found in the GIST specification. 184.108.40.206. Last Node Behavior The design of the QoS NSLP allows reservations to be installed at a subset of the nodes along a path. In particular, usage scenarios include cases where the data flow endpoints do not support the QoS NSLP.
In the case where the data flow receiver does not support the QoS NSLP, some particular considerations must be given to node discovery and rerouting at the end of the signaling path. There are three cases for the last node on the signaling path: 1) the last node is the data receiver, 2) the last node is a configured proxy for the data receiver, or 3) the last node is not the data receiver and is not explicitly configured to act as a signaling proxy on behalf of the data receiver. Cases (1) and (2) can be handled by the QoS NSLP itself during the initial path setup, since the QNE knows that it should terminate the signaling. Case (3) requires some assistance from GIST, which provides messages across the API to indicate that no further GIST nodes that support QoS NSLP are present downstream, and that probing of the downstream route change needs to continue once the reservation is installed to detect any changes in this situation. Two particular scenarios need to be considered in this third case. In the first, referred to as "Path Extension", rerouting occurs such that an additional QNE is inserted into the signaling path between the old last node and the data receiver, as shown in Figure 4. /-------\ Initial route / v /-\ /--|B|--\ +-+ / \-/ \ |x| = QoS NSLP aware +-+ /-\ +-+ ----|A| |D| +-+ \-/ /-\ \ +-+ / |x| = QoS NSLP unaware \--|C|--/ \-/ +-+ \ ^ \-------/ Updated route Figure 4: Path Extension When rerouting occurs, the data path changes from A-B-D to A-C-D. Initially the signaling path ends at A. Despite initially being the last node, node A needs to continue to attempt to send messages downstream to probe for path changes, unless it has been explicitly
configured as a signaling proxy for the data flow receiver. This is required so that the signaling path change is detected, and C will become the new last QNE. In a second case, referred to as "Path Truncation", rerouting occurs such that the QNE that was the last node on the signaling path is no longer on the data path. This is shown in Figure 5. /-------\ Initial route / v +-+ /--|B|--\ +-+ / +-+ \ |x| = QoS NSLP aware +-+ /-\ +-+ ----|A| |D| +-+ \-/ /-\ \ /-\ / |x| = QoS NSLP unaware \--|C|--/ \-/ \-/ \ ^ \-------/ Updated route Figure 5: Path Truncation When rerouting occurs, the data path again changes from A-B-D to A-C-D. The signaling path initially ends at B, but this node is not on the new path. In this case, the normal GIST path change detection procedures at A will detect the path change and notify the QoS NSLP. GIST will also notify the signaling application that no downstream GIST nodes supporting the QoS NSLP are present. Node A will take over as the last node on the signaling path. 220.127.116.11. Handling Spurious Route Change Notifications The QoS NSLP is notified by GIST (with the NetworkNotification primitive) when GIST believes that a rerouting event may have occurred. In some cases, events that are detected as possible route changes will turn out not to be. The QoS NSLP will not always be able to detect this, even after receiving messages from the 'new' peer. As part of the RecvMessage API primitive, GIST provides an SII-Handle that can be used by the NSLP to direct a signaling message to a particular peer. The current SII-Handle will change if the signaling peer changes. However, it is not guaranteed to remain the same after a rerouting event where the peer does not change. Therefore, the QoS NSLP mechanism for reservation maintenance after a route change
includes robustness mechanisms to avoid accidentally tearing down a reservation in situations where the peer QNE has remained the same after a 'route change' notification from GIST. A simple example that illustrates the problem is shown in Figure 6 below. (1) +-+ /-----\ |x| = QoS NSLP aware +-+ /-\ (3) +-+ +-+ ----|A| |B|-----|C|---- +-+ \-/ +-+ /-\ \-----/ |x| = QoS NSLP unaware (2) \-/ Figure 6: Spurious Reroute Alerting In this example, the initial route A-B-C uses links (1) and (3). After link (1) fails, the path is rerouted using links (2) and (3). The set of QNEs along the path is unchanged (it is A-C in both cases, since B does not support the QoS NSLP). When the outgoing interface at A has changes, GIST may signal across its API to the NSLP with a NetworkNotification. The QoS NSLP at A will then attempt to repair the path by installing the reservation on the path (2),(3). In this case, however, the old and new paths are the same. To install the new reservation, A will send a RESERVE message, which GIST will transport to C (discovering the new next peer as appropriate). The RESERVE also requests a RESPONSE from the QNR. When this RESERVE message is received through the RecvMessage API call from GIST at the QoS NSLP at C, the SII-Handle will be unchanged from its previous communications from A. A RESPONSE message will be sent by the QNR, and be forwarded from C to A. This confirms that the reservation was installed on the new path. The SII-Handle passed with the RecvMessage call from GIST to the QoS NSLP will be different to that seen previously, since the interface being used on A has changed. At this point, A can attempt to tear down the reservation on the old path. The RESERVE message with the TEAR flag set is sent down the old path by using the GIST explicit routing mechanism and specifying the SII-Handle relating to the 'old' peer QNE.
If RSNs were being incremented for each of these RESERVE and RESERVE- with-TEAR messages, the reservation would be torn down at C and any QNEs further along the path. To avoid this, the RSN is used in a special way. The RESERVE down the new path is sent with the new current RSN set to the old RSN plus 2. The RESERVE-with-TEAR down the old path is sent with an RSN set to the new current RSN minus 1. This is the peer from which it was receiving RESERVE messages (see for more details). 3.2.13. Preemption The QoS NSLP provides building blocks to implement preemption. This specification does not define how preemption should work, but only provides signaling mechanisms that can be used by QoS models. For example, an INFO-SPEC object can be added to messages to indicate that the signaled session was preempted. A BOUND-SESSION-ID object can carry the Session ID of the flow that caused the preemption of the signaled session. How these are used by QoS models is out of scope of the QoS NSLP specification. 3.3. GIST Interactions The QoS NSLP uses GIST for delivery of all its messages. Messages are passed from the NSLP to GIST via an API (defined in Appendix B of [RFC5971]), which also specifies additional information, including an identifier for the signaling application (e.g., 'QoS NSLP'), session identifier, MRI, and an indication of the intended direction (towards data sender or receiver). On reception, GIST provides the same information to the QoS NSLP. In addition to the NSLP message data itself, other meta-data (e.g., session identifier and MRI) can be transferred across this interface. The QoS NSLP keeps message and reservation state per session. A session is identified by a Session Identifier (SESSION-ID). The SESSION-ID is the primary index for stored NSLP state and needs to be constant and unique (with a sufficiently high probability) along a path through the network. The QoS NSLP picks a value for Session-ID. This value is subsequently used by GIST and the QoS NSLP to refer to this session. Currently, the QoS NSLP specification considers mainly the path- coupled MRM. However, extensions may specify how other types of MRMs may be applied in combination with the QoS NSLP. When GIST passes the QoS NSLP data to the NSLP for processing, it must also indicate the value of the 'D' (Direction) flag for that message in the MRI.
The QoS NSLP does not provide any method of interacting with firewalls or Network Address Translators (NATs). It assumes that a basic NAT traversal service is provided by GIST. 3.3.1. Support for Bypassing Intermediate Nodes The QoS NSLP may want to restrict the handling of its messages to specific nodes. This functionality is needed to support layering (explained in Section 3.2.10), when only the edge QNEs of a domain process the message. This requires a mechanism at the GIST level (which can be invoked by the QoS NSLP) to bypass intermediate nodes between the edges of the domain. The intermediate nodes are bypassed using multiple levels of the router alert option. In that case, internal routers are configured to handle only certain levels of router alerts. This is accomplished by marking this message at the ingress, i.e., modifying the QoS NSLP default NSLPID value to an NSLPID predefined value (see Section 6.6). The egress stops this marking process by reassigning the QoS NSLP default NSLPID value to the original RESERVE message. The exact operation of modifying the NSLPID must be specified in the relevant QoS model specification. 3.3.2. Support for Peer Change Identification There are several circumstances where it is necessary for a QNE to identify the adjacent QNE peer, which is the source of a signaling application message. For example, it may be to apply the policy that "state can only be modified by messages from the node that created it" or it might be that keeping track of peer identity is used as a (fallback) mechanism for rerouting detection at the NSLP layer. This functionality is implemented in the GIST service interface with SII-handle. As shown in the above example, we assume the SII- handling will support both its own SII and its peer's SII. Keeping track of the SII of a certain reservation also provides a means for the QoS NSLP to detect route changes. When a QNE receives a RESERVE referring to existing state but with a different SII, it knows that its upstream peer has changed. It can then use the old SII to initiate a teardown along the old section of the path. This functionality is supported in the GIST service interface when the peer's SII (which is stored on message reception) is passed to GIST upon message transmission.
3.3.3. Support for Stateless Operation Stateless or reduced-state QoS NSLP operation makes the most sense when some nodes are able to operate in a stateless way at the GIST level as well. Such nodes should not worry about keeping reverse state, message fragmentation and reassembly (at GIST), congestion control, or security associations. A stateless or reduced-state QNE will be able to inform the underlying GIST of this situation. GIST service interface supports this functionality with the Retain-State attribute in the MessageReceived primitive. 3.3.4. Priority of Signaling Messages The QoS NSLP will generate messages with a range of performance requirements for GIST. These requirements may result from a prioritization at the QoS NSLP (Section 3.2.11) or from the responsiveness expected by certain applications supported by the QoS NSLP. GIST service interface supports this with the 'priority' transfer attribute. 3.3.5. Knowledge of Intermediate QoS-NSLP-Unaware Nodes In some cases, it is useful to know that there are routers along the path where QoS cannot be provided. The GIST service interface supports this by keeping track of IP-TTL and Original-TTL in the RecvMessage primitive. A difference between the two indicates the number of QoS-NSLP-unaware nodes. In this case, the QNE that detects this difference should set the "B" (BREAK) flag. If a QNE receives a QUERY or RESERVE message with the BREAK flag set, and then generates a QUERY, RESERVE, or RESPONSE message, it can set the BREAK flag in those messages. There are however, situations where the egress QNE in a local domain may have some other means to provide QoS [RFC5975]. For example, in a local domain that is aware of RMD-QOSM [RFC5977] (or a similar QoS Model) and that uses either NTLP stateless nodes or NSIS-unaware nodes, the end-to-end RESERVE or QUERY message bypasses these NTLP stateless or NSIS-unaware nodes. However, the reservation within the local domain can be signaled by the RMD-QOSM (or a similar QoS Model). In such situations, the "B" (BREAK) flag in the end-to- end RESERVE or QUERY message should not be set by the edges of the local domain.