Architecture for Scheduled Use of Resources

Traffic Engineering Label Switched Paths (TE-LSPs) are connection oriented tunnels in packet and non-packet networks , . TE-LSPs may reserve network resources for use by the traffic they carry, thus providing some guarantees of service delivery and allowing a network operator to plan the use of the resources across the whole network. In some technologies (such as wavelength switched optical networks) the resource is synonymous with the label that is switched on the path of the LSP so that it is not possible to establish an LSP that can carry traffic without assigning a concrete resource to the LSP. In other technologies (such as packet switched networks) the resources assigned to an LSP are a measure of the capacity of a link that is dedicated for use by the traffic on the LSP. In all cases, network planning consists of selecting paths for LSPs through the network so that there will be no contention for resources; LSP establishment is the act of setting up an LSP and reserving resources within the network; and network optimization or re-optimization is the process of re-positioning LSPs in the network to make the unreserved network resources more useful for potential future LSPs while ensuring that the established LSPs continue to fulfill their objectives. It is often the case that it is known that an LSP will be needed at some time in the future. While a path for that LSP could be computed using knowledge of the currently established LSPs and the currently available resources, this does not give any degree of certainty that the necessary resources will be available when it is time to set up the new LSP. Yet setting up the LSP ahead of the time when it is needed (which would guarantee the availability of the resources) is wasteful since the network resources could be used for some other purpose in the meantime. Similarly, it may be known that an LSP will no longer be needed after some future time and that it will be torn down releasing the network resources that were assigned to it. This information can be helpful in planning how a future LSP is placed in the network. Time-Scheduled (TS) reservation of TE resources can be used to provide resource booking for TE-LSPs so as to better guarantee services for customers and to improve the efficiency of network resource usage into the future. This document provides a framework that describes and discusses the architecture for the scheduled reservation of TE resources. This document does not describe specific protocols or protocol extensions needed to realize this service.

TE-LSPs in existing networks are provisioned using RSVP-TE as a signaling protocol , by direct control of network elements such as in the Software Defined Networking (SDN) paradigm, and using the PCE Communication Protocol (PCEP) as a control protocol. TE resources are reserved at the point of use. That is, the resources (wavelengths, timeslots, bandwidth, etc.) are reserved for use on a specific link and are tracked by the Label Switching Routers (LSRs) at the end points of the link. Those LSRs learn which resources to reserve during the LSP setup process. The use of TE resources can be varied by changing the parameters of the LSP that uses them, and the resources can be released by tearing down the LSP.

Although TE-LSPs can determine their paths hop-by-hop using the shortest path toward the destination to route the signaling protocol messages , in practice this option is not applied because it does not look far enough ahead into the network to verify that the desired resources are available. Instead, the full length of the path of an LSP is computed ahead of time either by the head-end LSR of a signaled LSP, or by Path Computation Element (PCE) functionality in a dedicated server or built into network management software . Such full-path computation is applied in order that an end-to-end view of the available resources in the network can be used to determine the best likelihood of establishing a viable LSP that meets the service requirements. Even in this situation, however, it is possible that two LSPs being set up at the same time will compete for scarce network resources meaning that one or both of them will fail to be established. This situation is avoided by using a centralized PCE that is aware of the LSP setup requests that are in progress.

LSPs may be established "on demand" when the requester determines that a new LSP is needed. In this case, the path of the LSP is computed as described in . However, in many situations, the requester knows in advance that an LSP will be needed at a particular time in the future. For example, the requester may be aware of a large traffic flow that will start at a well-known time, perhaps for a database synchronzation or for the exchange of content between streamng sites. Furthermore, the requester may also know for how long the LSP is required before it can be torn down. The set of requests for future LSPs could be collected and held in a central database (such as at a Network Management System - NMS): when the time comes for each LSP to be set up the NMS can ask the PCE to compute a path and can then requst the LSP to be provisioned. This approach has a number of drawbacks because it is not possible to determine in advance whether it will be possible to deliver the LSP since the resources it needs might be used by other LSPs in the network. Thus, at the time the requester asks for the future LSP, the NMS can only make a best-effort guarantee that the LSP will be set up at the desired time. A better solution, therefore, is for the requests for future LSPs to be serviced at once. The paths of the LSPs can be computed ahead of time and converted into reservations of network resources during specific windows in the future.

While path computation as described in takes account of the currently available network resources, and can act to place LSPs in the network so that there is the best possibility of future LSPs being accommodated, it cannot handle all eventualities. It is simple to construct scenarios where LSPs that are placed one at a time lead to future LSPs being blocked, but where foreknowledge of all of the LSPs would have made it possible for them all to be set up. If, therefore, we were able to know in advance what LSPs were going to be requested we could plan for them and ensure resources were available. Furthermore, such an approach enables a commitment to be made to a service user that an LSP will be set up and available at a specific time. This service can be achieved by tracking the current use of network resources and also a future view of the resource usage. We call this time-scheduled TE (TS-TE) resource reservation.

In order to achieve the TS-TE resource reservation, the use of resources on the path needs to be scheduled. Scheduling state is used to indicate when resources are reserved and when they are available for use. A simple information model for one piece of scheduling state is as follows:

{ link id; resource id or reserved capacity; reservation start time; reservation end time } The resource that is scheduled can be link capacity, physical resources on a link, CPU utilization, memory, buffers on an interfaces, etc. The resource might also be the maximal unreserved bandwidth of the link over a time intervals. For any one resource there could be multiple pieces of scheduling state, and for any one link, the timing windows might overlap. There are multiple ways to realize this information model and different ways to store the data. The resource state could be expressed as a start time and and end time as shown above, or could be expressed as a start time and a duration. Multiple periods, possibly of different lengths, may be associated with one reservation request, and a reservation might repeat on a regular cycle. Furthermore, the current state of network reservation could be kept separate from the scheduled usage, or everything could be merged into a single TS databasae. This document does not spend any more time on discussion of encoding of state information except to discuss the location of storage of the state information and the recovery of the information after failure events. This scheduling state information can be used by applications to book resources for future or now, so as to maximize chance of services being delivered. Also, it can avoid contention for resources of LSPs. Note that it is also to store the information about future LSPs. This information is held to allow the LSPs to be instantiated when they are due and using the paths/resources that have been computed for them, but also to provide correlation with the TS-TE resource reservations so that it is clear why resources were reserved allowing pre-emption and handling release of reserved resources in the event of cancelation of future LSPs.

This section examines several important architectural concepts that lead to design decisions that will influence how networks can achieve TS-TE in a scalable and robust manner.

The scheduling state information described in has to be held somewhere. There are two places where this makes sense: In the network nodes where the resources exist; In a central scheduling controller where decisions about resource allocation are made. The first of these makes policing of resource allocation easier. It means that many points in the network can request immediate or scheduled LSPs with the associated resource reservation and that all such requests can be correlated at the point where the resources are allocated. However, this approach has some scaling and technical problems: The most obvious issue is that each network node must retain the full time-based state for all of its resources. In a busy network with a high arrival rate of new LSPs and a low hold time for each LSP, this could be a lot of state. Yet network nodes are normally implemented with minimal spare memory. In order that path computation can be performed, the computing entity normally known as a Path Computation Element (PCE) needs access to a database of available links and nodes in the network, and of the TE properties of the links. This database is known as the Traffic Engineering Database (TED) and is usually populated from information advertised in the IGP by each of the network nodes or exported using BGP-LS . To be able to compute a path for a future LSP the PCE needs to populate the TED with all of the future resource availability: if this information is held on the network nodes it must also be advertised in the IGP. This could be a significant scaling issue for the IGP and the network nodes as all of the advertised information is held at every network node and must be periodically refreshed by the IGP. When a normal node restarts it can recover resource reservation state from the forwarding hardware, from Non-volatile random-access memory (NVRAM), or from adjacent nodes through the signaling protocol . If scheduling state is held at the network nodes it must also be recovered after the restart of a network node. This cannot be achieved from the forwarding hardware because the reservation will not have been made, could require additional expensive NVRAM, or might require that all adjacent nodes also have the scheduling state in order to reinstall it on the restarting node. This is potentially complex processing with scaling and cost implications. Conversely, if the scheduling state is held centrally it is easily available at the point of use. That is, the PCE can utilize the state to plan future LSPs and can update that stored information with the scheduled reservation of resources for those future LSPs. This approach also has several issues: If there are multiple controllers then they must synchronise their stored scheduling state as they each plan future LSPs, and must have a mechanism to resolve resource contention. This is relatively simple and is mitigated by the fact that there is ample processing time to replan future LSPs in the case of resource contention. If other sources of immediate LSPs are allowed (for example, other controllers or autonomous action by head-end LSRs) then the changes in resource availability caused by the setup or teardown of these LSPs must be reflected in the TED (by use of the IGP as currently) and may have an impact of planned future LSPs. This impact can be mitigated by replanning future LSPs or through LSP preemption. If other sources of planned LSPs are allowed, they can request path computation and resource reservation from the centralized PCE using PCEP . If the scheduling state is held centrally at a PCE, the state must be held and restored after a system restart. This is relatively easy to achieve on a central server that can have access to non-volatile storage. The PCE could also synchronize the scheduling state with other PCEs after restart. See for details. Of course, a centralized system must store informaton about all of the resources in the network. In a busy network with a high arrival rate of new LSPs and a low hold time for each LSP, this could be a lot of state. This is multiplied by the size of the network measured both by the number of links and nodes, and by the number of trackable resources on each link or at each node. The challenge may be mitigated by the centralized server being dedicated hardware, but the problem of collecting the information from the network is only solved if the central server has full control of the booking of resources and the estblshment of new LSPs. Thus the architectural conclusion is that scheduling state should be held centrally at the point of use and not in the network devices.

As already described, the PCE needs access to an enhanced, time-based TED. It stores the traffic engineering (TE) information such as bandwidth for every link for a series of time intervals. There are a few ways to store the TE information in the TED. For example, suppose that the amount of the unreserved bandwidth at a priority level for a link is Bj in a time interval from time Tj to Tk (k = j+1), where j = 0, 1, 2, ....

Time |T0 T1 T2 T3 T4 ]]> The unreserved bandwidth for the link can be represented and stored in the TED as [T0, B0], [T1, B1], [T2, B2], [T3, B3], ... as shown in . But it must be noted that service requests for future LSPs are known in terms of the LSPs whose paths are computed and for which resources are scheduled. For example, if the requester of a future LSP decides to cancel the request or to modify the request, the PCE must be able to map this to the resources that were reserved. When the LSP or the request for the LSP with a number of time intervals is cancelled, the PCE must release the resources that were reserved on each of the links along the path of the LSP in every time intervals from the TED. If the bandwidth reserved on a link for the LSP is B from time T2 to T3 and the unreserved bandwidth on the link is B2 from T2 to T3, B is added to the link for the time interval from T2 to T3 and the unreserved bandwidth on the link from T2 to T3 will be B2 + B. This suggests that the PCE needs an LSP Database (LSP-DB) that contains information not only about LSPs that are active in the network, but also those that are planned. The information for an LSP stored in the LSP-DB includes for each time interval that applies to the LSP: the time interval, the paths computed for the LSP satisfying the constraints in the time interval, and the resources such as bandwidth reserved for the LSP in the time interval. See also It is an implementation choice how the TED and LSP-DB are stored both for dynamic use and for recovery after failure or restart, but it may be noted that all of the information in the scheduled TED can be recovered from the active network state and from the scheduled LSP-DB.

The architectural considerations and conclusions described in the previous section lead to the architecture described in this section.

| LSP-DB | | | -------- | PCE | | | c ----- | |<---->| TED | ------- ----- ^ ^ | | d| |e | | ------+-----+-------------------- | | Network | -------- | | Router | v -------- ----- ----- | LSR |<------>| LSR | ----- f ----- ]]>

As shown in , some component in the network requests a service. This may be an application, an NMS, an LSR, or any component that qualifies as a Path Computation Client (PCC). We show this on the figure as the "Service Requester" and it sends a request to the PCE for an LSP to be set up at some time (either now or in the future). The request, indicated on by the arrow (a) includes all of the parameters of the LSP that the requester wishes to supply such as bandwidth, start time, and end time. Note that the requester in this case may be the same LSR shown in the figure or may be a distinct system. The PCE enters the LSP request in its LSP-DB (b), and uses information from its TED (c) to compute a path that satisfies constraints such as bandwidth constraint for the LSP in the time interval from a start time to an end time. It updates the future resource availability in the TED so that further path computations can take account of the scheduled resource usage. It stores the path for the LSP into the LSP-DB (b). When it is time such as at a start time for the LSP to be set up, the PCE sends a PCEP Initiate request to the head end LSR (d) providing the path to be signaled as well as other parameters such as the bandwidth of the LSP. As the LSP is signaled between LSRs (f) the use of resources in the network is updated and distributed using the IGP. This information is shared with the PCE either through the IGP or using BGP-LS (e), and the PCE updates the information stored in its TED (c). After the LSP is set up, the head end LSR sends a PCEP LSP State Report (PCRpt message) to the PCE (d). The report contains the resources such as bandwidth usage for the LSP. The PCE updates the status of the LSP in the LSPDB according to the report. When an LSP is no longer required (either because the Service Requester has cancelled the request, or because the LSP's scheduled lifetime has expired) the PCE can remove it. If the LSP is currently active, the PCE instructs the head-end LSR to tear it down (d), and the network resource usage will be updated by the IGP and advertised back to the PCE through the IGP or BGP-LS (e). Once the LSP is no longer active, the PCE can remove it from the LSP-DB (b).

When a PCE in the architecture shown in is initialized, it must learn state from the network, from its stored databases, and potentially from other PCEs in the network. The first step is to get an accurate view of the topology and resource availability in the network. This would normally involve reading the state direct from the network via the IGP or BGP-LS (e), but might include receiving a copy of the TED from another PCE. Note that a TED stored from a previous instantiation of the PCE is unlikely to be valid. Next, the PCE must construct a time-based TED to show scheduled resource usage. How it does this is implementation specific and this document does not dictate any particular mechanism: it may recover a time-based TED previously saved to non-volatile storage, or it may reconstruct the time-based TED from information retrieved from the LSP-DB previously saved to non-volatile storage. If there is more than one PCE active in the network, the recovering PCE will need to synchronize the LSP-DB and time-based TED with other PCEs (see ).

If there is more than one PCE active in the network which supports scheduling, it is important to achieve some consistency between the scheduled TED and scheduled LSP-DB between the PCEs. answers various questions around synchronization between the PCEs. It should be noted that the time-based "scheduled" information adds another dimension to it. It should be noted that the deployment may use a primary PCE and the other PCEs as backup, where the backup PCE can take over only in the event of a failure of the primary PCE. Or the PCEs may share the load at all times. The choice of the synchronization technique is largely dependent on the deployment of PCEs in the network. One option for ensuring that multiple PCEs use the same scheduled information is simply to have the PCEs driven from the same shared database, but it is likely to be inefficient and inter-operation between multiple implementation harder. Or the PCEs might be responsible for its own scheduled database and utilize some distributed database synchronization mechanism to have a consistent database. Based on the implementation, this could be efficient but the inter-operation between heterogeneous implementation is still hard. Another approach would be to utilize PCEP messages to synchronize the scheduled state between PCEs. This approach would work well if the number of PCEs which support scheduling are less, but as the number increases considerable message exchange needs to happen to keep the scheduled database in sync. Future solution could also utilize some synchronization optimization techniques for efficiency. Another variation would be to request information from other PCEs for a particular time slice but this might have impact on the optimization algorithm.

TBD

This work has benefited from the discussions of resource scheduling over the years. In particular the DRAGON project and both of which provide approaches to auto-bandwidth services in GMPLS networks. Mehmet Toy, Lei Liu and Khuzema Pithewan contributed the earlier version of . We would like to thank authors of that draft on Temporal Tunnel Services andfor help inspire discussion in the TEAS WG and get this work solid. Thanks to Michael Scharf and Daniele Ceccarelli for useful comments on this work.

The following people contributed to discussions that led to the development of this document: