RMT Working Group Generic Router Assist (GRA) Brad Cain INTERNET-DRAFT for Multicast Transport Protocols Nortel Tony Speakman Expires September 2000 cisco Don Towsley UMASS March 2000 Generic Router Assist (GRA) Building Block Motivation and Architecture Status of this Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of RFC2026. 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." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract Generic Router Assist (GRA) is a network-based service that enables end-to-end multicast transport protocols to take advantage of infor- mation distributed across the network elements in a given multicast distribution tree. The service consists of a canonical set of simple functions which network elements may apply to selected packets in the transport session as they traverse the distribution tree. The choice of function and the packet parameters to which it is applied can be defined and customized for a given transport session in a highly con- trolled fashion that still permits enough flexibility for GRA to be used to address a wide range of multicast transport problems not Cain/Speakman/Towsley [Page 1] INTERNET-DRAFT Generic Router Assist March 2000 amenable to end-to-end solution. This document provides the motiva- tion and an architecture for GRA. Cain/Speakman/Towsley [Page 2] INTERNET-DRAFT Generic Router Assist March 2000 Table of Contents 1. Introduction ............................................... 4 2. Scope of Generic Router Assist ............................. 6 3. Canonical Services and Functional Models/Examples .......... 11 4. Implementation Considerations .............................. 17 Abbreviations .................................................. 19 References ..................................................... 20 Revision History ............................................... 21 Authors' Addresses ............................................. 21 Cain/Speakman/Towsley [Page 3] INTERNET-DRAFT Generic Router Assist March 2000 1. Introduction The development of scalable end-to-end multicast protocols poses a tremendous challenge to network protocol designers. For example the development of reliable multicast protocols has received considerable attention in recent years. Most protocols are based on an end-to-end solution [SRM, RMTP, TKP] and have found the problem of scaling to 1000s or even 100s of receivers daunting. The primary obstacles to the development of scalable protocols have been "feedback implosion" and "transmission isolation". The first of these concerns the diffi- culty of a large multicast application to limit feedback from receivers to a data source or to each other. The second concerns the difficulty of limiting the transmission of data to the subset of a multicast group that requires it. Several proposals have been made to add functionality to routers for the purpose of improving the performance of multicast applications, particularly reliable multicast. Papadopoulos and Parulkar [LSM] introduced additional forwarding functionality to a router which would allow each router to identify a ``special outgoing interface'' over which to transmit a particular class of packets. They showed how this ``turning point functionality'' could be used to improve the performance of reliable multicast protocols. Levine and Garcia-Luna- Aceves [LABEL] proposed the addition of ``routing labels'' to routing tables which could be used to direct packets over specific inter- faces. One of these, called a distance label, was shown to be quite useful in reliable multicast for directing requests for repairs to nearby repair servers. The third and, perhaps most relevant proposal is the PGM protocol [PGM]. Briefly, PGM is a reliable multicast pro- tocol which uses negative acknowledgements (NAKs). The PGM protocol is an end-to-end transport protocol that contains a router component which performs NAK suppression and retransmission subcasting func- tionality. This proposal, like others [GMTS, BFS], is primarily motivated by PGM and the recognition of the benefits of exporting a set of flexible, simple router-based functionality for the purpose of multicast protocol design. Such functionality can significantly sim- plify the design of a large class of scalable multicast transport protocols. In this draft, we present Generic Router Assist (GRA) functionality intended to help protocol designers deal with the two problems of feedback implosion and transmission isolation. This functionality is designed to assist in the scaling of receiver feedback information and in providing subcasting for large multicast groups. It consists of simple filtering and aggregation functions that reside within routers. Signaling protocols are used by hosts to set up and invoke this Cain/Speakman/Towsley [Page 4] INTERNET-DRAFT Generic Router Assist March 2000 functionality. Briefly, a data source first initializes one or more desired services on its multicast tree using GRA setup messages. The GRA-capable routers on the tree then aggregate feedback from receivers and/or isolate transmissions through the use of filters set by either the sender or the receivers. For robustness, periodic transmissions of setup messages on the multicast tree are used to refresh GRA state in the face of routing changes and other possible errors. It should be stressed that GRA services are only invoked for certain packets; data packets are usually not treated any differently and will not cause any additional processing in routers. GRA is not intended to provide sophisticated services which are dif- ficult or impossible to implement in routers. GRA functionality is implemented at the IP layer and provides unreliable ``best-effort'' services. Transport protocols which make use of GRA must be robust in the face of failures and the absence of GRA-capable routers in the network. Before describing the details of GRA, we present a simple example in the context of a PGM-like reliable multicast protocol. Consider a NAK-based reliable multicast protocol which places the responsibility of packet loss detection on each receiver. Each time that a receiver detects a loss (based on a gap in the sequence numbers of the packets that it receives), it unicasts a request for a repair (NAK) to the sender. Upon receipt of a NAK for a specific packet, the sender retransmits the packet to all receivers. This protocol faces considerable challenges in dealing with multiple NAKs for the same packet. First, there is the problem of the sender having to process many NAKs. Second, there is the problem of limiting the number of retransmissions to the same packet. GRA can be used to (partially) solve these two problems. Prior to the transfer of any data, the application sets up a NAK aggregation filter at each GRA- capable router using a setup message. This filter is set up to suppress NAKs generated for the same packet. In addition, the router maintains information regarding the interfaces over which it has received NAKs so that it can subcast the retransmission on the por- tion of the multicast tree that contains receivers requiring a retransmission of the packet. In Figure 1, we show how GRA can be used to aggregate feedback infor- mation in a reliable multicast transport protocol. In this figure, a multicast source (Src 1) transmits to two receivers (Rec 1 and Rec 2). The data packets from Src 1 are treated as regular multicast packets and forwarded accordingly. On the link between router R1 and router R2, a data packet is lost. Assuming a NAK based reliable mul- ticast protocol, this loss causes the receivers to each send a NAK to Cain/Speakman/Towsley [Page 5] INTERNET-DRAFT Generic Router Assist March 2000 the source for the packet that was lost. In the example, receivers invoke GRA to send the NAKs to the source. Router R2 treats these NAKs in a special manner, removing the redundant NAKs to the source. Therefore, only one NAK arrives at the source. We see from this exam- ple that only certain types of packets require additional processing at GRA routers and that the majority of end-to-end packets are for- warded according to normal multicast forwarding rules (i.e. without additional router processing). Src 1 ^ | | | | | data packet | | | R1 | ^ | | | | V R2 suppresses | | X data packet dropped duplicate +--->R2 <---+ NAK | --- --- | | | | | | | | | | | | | R3 R4 GRA ^ | | ^ GRA NAK | | | | NAK | | | | Rec 1 Rec 2 Figure 1. Example of network support for transport protocols. We conclude the introduction by commenting on the relationship of GRA to the multicast routing algorithm. GRA works on all types of multi- cast forwarding trees. However, GRA state is per session state and requires per session state in routers. If a source-based tree routing protocol is used to forward multicast packets, then this per session state will already exist in routers in the form of multicast forward- ing entries. If a shared tree type multicast routing protocol is being used, then GRA per session state must be maintained on the shared tree. This is simply because GRA provides per session func- tionality. 2. Scope of Generic Router Assist The types of services implemented with GRA are bounded by constraints and limitations of routing devices. In this section we explicitly describe the limitations and constraints of routing devices and Cain/Speakman/Towsley [Page 6] INTERNET-DRAFT Generic Router Assist March 2000 explain what we believe to be reasonable services to implement in routers. We specifically describe router limitations in order to limit the scope of GRA. We believe that a small set of GRA services in routers can assist in scaling many of the problems in end-to-end multicast. Previous proposals[ACTIVE] have proposed complex elements that reside in routers to provide sophisticated capabilities. We feel that these are unreasonable for current generation routers. The approach of GRA is to provide a simple fixed set of services which give the maximum benefit with the least cost. 2.1. Service Properties GRA services are performed on subset of packets sent between end-to- end transport protocols on the multicast distribution tree between a GRA Source and the set of GRA receivers. Only routers on the distri- bution tree for a particular GRA source act upon GRA packets. The advantages of GRA type services can only be realized when the actions are performed on packets that are directly on the forwarding tree of a multicast group. In order to describe the requirements of GRA, we first describe the properties of what we feel are appropriate services. Fixed: by fixed services we mean those of which are statically part of router software or hardware. We DO NOT mean dynamically loadable modules. We feel that a fixed set of simple services will suffice for most of the scaling issues in transport proto- cols. Simple: We wish only to include those services that we feel are reasonable to be implemented in routers. These are services which can be performed with minimum CPU and memory overhead. Short Term: We wish to provide services for which state and pro- cessing overhead is short lived. GRA makes use of soft-state design principles. 2.2. GRA Requirements When considering the service and architecture of GRA, we adhere to the following principles: 1. GRA services should be simple and fixed. They should not require excessive processing in routing devices nor should they buffer messages. Cain/Speakman/Towsley [Page 7] INTERNET-DRAFT Generic Router Assist March 2000 2. GRA services are not substitutes for well-engineered end-to-end protocol designs. We support the end-to-end design principles of transport protocols. GRA is an "assist" service which is designed to assist protocols in scaling aspects. 3. GRA services will not take on active networking attributes such as dynamically uploadable modules or programming language propo- sals. 4. GRA services should not directly participate in transport pro- tocols. GRA should not be required for any transport protocols nor should any GRA services directly support a particular proto- col. 5. GRA services should be those which may assist all or a reason- able subset of transport protocols. 6. GRA services should be used for assisting in *control* packet operations. GRA services should not be for the majority of pack- ets in a multicast group. 2.3. Constraints of networking devices Current generation routers perform processing of packets and execute routing and signaling protocols. Routers perform fast packet for- warding on the "forwarding path". This is usually performed by hardware and software specifically designed for the forwarding of packets (and may include other functions such as policing, shaping, etc). This is in contrast to the "control plane" of a router where control protocols are run under an operating system. Examples of these types of protocols are routing, management and signaling proto- cols. We now describe the role and impact of GRA services on these two "planes" of a router. Forwarding path: A router forwarding path usually consists of spe- cialized hardware and software which is designed specifically for the purpose of forwarding packets. Newer routers also have abili- ties to perform other actions such as marking or policing for ser- vices such as QoS. In general, the forwarding path of routers is very limited in the amount of state and complexity of processing that can be performed. Although we do not rule out GRA services being implemented in the forwarding path of routers, we do not see this as feasible at this point in time. This is because of the more complicated processing rules for GRA packets (in comparison with basic forwarding operations) and the amount of state that is sometimes involved in performing GRA operations. Cain/Speakman/Towsley [Page 8] INTERNET-DRAFT Generic Router Assist March 2000 Control plane: Router control planes run embedded or general operating systems. On top of these operating systems are imple- mentations of IP control protocols such as routing, signaling, and network management. The processing power and memory of a control plane depends highly on the hardware design of the router. Most routers use general purpose CISC or RISC processors for running the control plane operating system and protocols. The control plane is limited in processing by its hardware and the load gen- erated from the other processes or tasks running. We expect that most GRA operations will be performed in the control plane where state and processing power are more readily available. However, we do stress that routers generally have fairly slow control plane hardware. This is to keep the cost low and because the processing required for control plane protocols is usually low. 2.4. State Constraints As we evaluate particular services which are candidates for inclusion in GRA, constraints in router state are an important consideration. We wish to select services which will not create substantial or long-lived state. 2.4.1. Session State Routers which perform multicast forwarding contain per tree forward- ing state. For trees rooted at multicast sources, this amounts to per source per group forwarding entries. This state must be kept in both the forwarding as well as the control plane. GRA services are per session, or per source. This is simply because GRA services are per transport session. The session state of GRA does not impose much additional state to routers. Each session requires a state block describing the desired services. Other types of state may also be created during the course of a GRA session. The GRA session state is the only state required for the length of a session; other state is set up and torn down in smaller intervals. 2.4.2. Packet State The implementation of particular services may require per packet state in GRA routers. Services which require per packet state should use short-lived timers to tear down this state so as to avoid an explosion in the amount of state at routers. An example of a service which causes per packet state is NAK elimination. We feel the use of small timers will minimize problems in state growth. Cain/Speakman/Towsley [Page 9] INTERNET-DRAFT Generic Router Assist March 2000 2.4.3. Buffering GRA services are not allowed to buffer packets. GRA services may drop or modify packets in transit, but they will never buffer pack- ets. We feel that the buffering of packets in routing devices is unaccept- able due to the unpredictable behavior of such a service. In addi- tion we feel that this is an unreasonable service for routers to sup- port without a significant payback in end-to-end scaling. 2.5. Processing Constraints GRA services may require differing amount of computation. As a gen- eral rule, GRA services should require minimal computation and packet manipulation. 2.5.1. Computation We expect most GRA services to require minimal computation. Services which require minimal computation are reasonable to implement in routing devices and minimize security risk. Examples of operations which we feel are appropriate are: Comparison Operations: comparison operations may be performed on GRA packets against the GRA session state in the router in order to determine whether a service should be invoked. Examples of comparison operations are: equal to, less than, greater than, etc. Update Operations: When receiving a GRA packet, a router may be required to update its GRA session state. The operations required for updating the state should remain simple. Example of simple operations are addition, subtraction, etc. In order to limit the computational complexity of GRA services, we recommend that GRA operations remain singular. The combination of predicates and operations creates problems in both router processing and in GRA specifications themselves. We wish to avoid problems regarding operator and action precedence. 2.5.2. Buffer Operations One can define services requiring extensive packet manipulation by GRA routers. We believe this to be expensive and therefore unreason- able for routers. GRA services should be invoked without extensive manipulation of packets. Services which update or overwrite fields are acceptable. Services which require the formation of new packets or aggregate information into new packets are unacceptable. Cain/Speakman/Towsley [Page 10] INTERNET-DRAFT Generic Router Assist March 2000 2.6. Examples of "Reasonable" Services In this section we briefly describe two GRA services which we feel are appropriate and why we feel they meet the above requirements. The two example services are those that provide general functions which do not incur large amounts of state or processing. Elimination: Elimination is the selected dropping of redundant signaling. An example of elimination is NAK suppression in a reliable multicast protocol. Elimination is a service which requires little computational overhead. It does however, require per packet (or per sequence block) state. This state can be con- trolled by the use of short soft state timers. Subcasting: Subcasting is the forwarding of a multicast packet to a subset of the multicast tree. Subcasting is useful for a variety of multicast protocols. Subcasting does not require a large amount of processing. The state required is an identifier for the subtree and a list of outgoing interfaces. This state is similar to multicast forwarding tree state. 3. Canonical Services and Functional Models/Examples While a variety of mechanisms must come together to enable a specific GRA service in a distribution tree (session path messages to estab- lish session parameters and neighbour information, a control protocol to define, enable, and disable specific filtering services, etc.), the basic mechanism of GRA consists of router based services that can be described in the language of filters, keys, and conditional func- tions (binary predicates and their outcomes). Using the example of reliable multicast presented earlier, this sec- tion expands the model and terminology of filters to describe the full flexibility of GRA. The intent here is to generalize the mechanism fully enough to be able to accommodate currently unspeci- fied requirements for multicast transport services as they emerge. Revisiting the example of predicate elimination, note that the receipt of a loss report at a router implies several things. The packet type implies that a type of filter should be established for the transport session, that the filter key is (a sequence number) of a particular length at a particular offset, that the loss report in hand should be forwarded, that the interface upon which the loss report was received should be recorded and associated with the key in the filter, and that subsequent matching loss reports on any inter- face should be eliminated. The filter itself has other implied characteristics. It eliminates only for a certain interval after forwarding any subsequent loss report, and it has an implied lifetime Cain/Speakman/Towsley [Page 11] INTERNET-DRAFT Generic Router Assist March 2000 after which it is locally expired by the router. Similarly in the example of sub-casting, the receipt of a retransmis- sion at a router implies several things. The packet type implies that the packet should be matched against an existing filter type based on a key of a particular length and at a particular offset, that the packet itself should be forwarded on all interfaces for which a loss report was recorded associated with the key, and that the key's state should be discarded. By implication, unmatched retransmissions should not be forwarded. From this example some generalities emerge which, once they are extricated from their specific semantics, can be re-assembled in a variety of useful ways to provide router-based assistance for a broader class of transport services. 3.1. General Model The general model is one in which packets carry one of a tightly con- strained set of signals that alert the router to apply a pre-defined filtering service. The packet-borne signal conveys a filter type, an associated action, a key, and possible packet variables Canonical filtering services in routers can be defined by a filter type, associated supported actions, predicated on specific conditions, and three functions gated on that predicate whether TRUE or FALSE: f(p): how to dispose of the packet, f(s): how to transform the key's state, and f(v): how to transform the outgoing interface (OIF) list associated with the key. All of these as well as the offsets and lengths of the key and any packet variables for each supported action constitute the definition of the filtering service. 3.2. An Example Using the General Model Given this model, the handling of retransmissions in PGM can be Cain/Speakman/Towsley [Page 12] INTERNET-DRAFT Generic Router Assist March 2000 described as a predicate eliminating and subcasting filter. Let IIF be the interface on which a packet is received. Let RETX denote whether a packet retransmission has been requested either in a loss report (RQST RETX), as interface state (OIF RETX), or as key state (KEY RETX). Let ET be the elimination timer for a particular key's state, and LT be the life timer for a particular key's state. The filtering service in the router supports two actions, one inbound (RCVR_UPDATE) and one outbound (FORWARD). 3.2.1. RCVR_UPDATE For RCVR_UPDATE, in addition to a transport session identifier, the following are defined for the signal in the packet: ELIM_SUBCis the filter type RCVR_UPDATE is the action SQNis the key RETX is a packet-borne variable (value of one in the PGM example) For RCVR_UPDATE, the following are defined for the filtering service in the router: In case the key fails to match existing key state: predicate: NOOP f(v):OIF RETX = 1 for IIF f(s):start ET, start LT, KEY RETX = RQST RETX f(p):reverse forward to upstream neighbour In case the key matches existing key state: f(v): OIF RETX = 1 for IIF f(s): restart LT f(p): discard FORWARD For FORWARD, in addition to a transport session identifier, the fol- lowing are defined for the signal in the packet: ELIM_SUBCis the filter type FORWARD is the action SQNis the key For FORWARD, the following are defined for the filtering service in the router: Cain/Speakman/Towsley [Page 13] INTERNET-DRAFT Generic Router Assist March 2000 In case the key fails to match existing key state: predicate: NOOP f(v):NOOP f(s):NOOP f(p):discard In case the key matches existing key state: predicate: for all OIF RETXs, OIF RETX NE 0 In case the predicate is TRUE: f(v): decrement OIF RETX f(s): NOOP f(p): forward on OIF In case the predicate is FALSE: f(v): NOOP f(s): NOOP f(p): discard Associated with the filtering service would be an additional house- keeping function which would discard a key's state if either LT expired or all the OIF COUNTs were zero. The point of this and subsequent examples is to demonstrate that this model for GRA can accommodate a highly functional set of router-based services which, given a transport-layer- independent implementation, could be provided by routers for general deployment by transport pro- tocols engineered to signal the routers in a generic way. We now present a refinement of the PGM example adding forward error correc- tion. In this example, a packet variable is used to carry a count of parity packets requested with the additional implication that the interface upon which the loss report was received along with the count of par- ity packets requested should be recorded and associated with the key in the filter, and that subsequent matching loss reports on any interface should be eliminated unless they request a larger number of parity packets than has been requested on any interface. Similarly upon forwarding parity packets implies that the packet itself should be forwarded on all interfaces with non-zero counts recorded as state associated with the key, that the corresponding count should be decremented by 1 per interface until it reaches 0, and that the key's state should be discarded once all counts are Cain/Speakman/Towsley [Page 14] INTERNET-DRAFT Generic Router Assist March 2000 zero. 3.3. Another Example Using the General Model The handling of parity NAKs and parity retransmissions in PGM can be described as a predicate eliminating and subcasting filter this time augmented by a packet variable, the number of parity packets requested. Let IIF be the interface on which a packet is received. Let COUNT be the number of parity packets requested and recorded either in the loss report (RQST COUNT), as interface state (OIF COUNT), or as key state (HIGH COUNT). Let ET be the elimination timer for a particular key's state, and LT be the life timer for a particular key's state. The filtering service in the router supports two actions, one inbound (RCVR_UPDATE) and one outbound (FORWARD). 3.3.1. RCVR_UPDATE For RCVR_UPDATE, in addition to a transport session identifier, the following are defined for the signal in the packet: ELIM_SUBC is the filter type RCVR_UPDATE is the action SQN is the key COUNT is a packet-borne variable For RCVR_UPDATE, the following are defined for the filtering service in the router: In case the key fails to match existing key state: predicate: NOOP f(v): OIF COUNT for IIF = MAX(RQST COUNT, 0) f(s): start ET, start LT, HIGH COUNT = RQST COUNT f(p): reverse forward to upstream neighbour In case the key matches existing key state: predicate: ET is running or RQST COUNT LEQ HIGH COUNT In case the predicate is TRUE: f(v): OIF COUNT for IIF = MAX(RQST COUNT, OIF COUNT for IIF) f(s): restart LT f(p): discard Cain/Speakman/Towsley [Page 15] INTERNET-DRAFT Generic Router Assist March 2000 In case the predicate is FALSE: f(v): OIF COUNT for IIF = MAX(RQST COUNT, OIF COUNT for IIF) f(s): restart ET, HIGH COUNT = RQST COUNT f(p): reverse forward to upstream neighbour 3.3.2. FORWARD For FORWARD, in addition to a transport session identifier, the fol- lowing are defined for the signal in the packet: ELIM_SUBC is the filter type FORWARD is the action SQN is the key For FORWARD, the following are defined for the filtering service in the router: In case the key fails to match existing key state: predicate: NOOP f(v): NOOP f(s): NOOP f(p): discard In case the key matches existing key state: predicate: for all OIF COUNTs, OIF COUNT NE 0 In case the predicate is TRUE: f(v): decrement OIF COUNT f(s): NOOP f(p): forward on OIF In case the predicate is FALSE: f(v): NOOP f(s): NOOP f(p): discard Associated with the filtering service would be an additional house- keeping function which would discard a key's state if either LT expired or all the OIF COUNTs were zero. 3.4. Summary The point of this example (eventually, "these examples") is to Cain/Speakman/Towsley [Page 16] INTERNET-DRAFT Generic Router Assist March 2000 demonstrate that this model for GRA can accommodate a highly func- tional set of router-based services which, given a transport-layer- independent implementation, could be provided by routers for general deployment by transport protocols engineered to signal the routers in a generic way. Now that we have established the context and a model for GRA, the next section discusses implementation considerations in the interest of making the mechanisms of GRA more concrete, and highlighting their practical and performance consequences to tradi- tional multicast forwarding. 4. Implementation Considerations 4.1. Signalling Protocol - GRA service indicators A consideration of the implementation issues attending GRA strongly determines the class of functions that may be realized with this mechanism. These considerations relate to both security of the router and performance in the forwarding path. The former will be dealt with in another section. This section outlines the time and space scaling consequences of GRA to the forwarding path. To be a generic network layer service, it's clear that some minimal indicator is required in the network layer to signal the presence of GRA signal on a packet. As has been previously noted, remember that the GRA indicator is typically NOT borne by basic data packets. These are switched without exception in the usual forwarding path. In this section, packets bearing a GRA indicator will be referred to as GRA packets. A network layer indicator frees forwarding engines from the burden of having to walk into and parse transport headers on every switched packet. It's highly efficient to encode the GRA indicator on the network layer header since that header is typically already within the grasp of the forwarding engine. In PGM, this indicator is imple- mented with an IP Router Alert option, but a single bit in the basic IP header would function just as well (even better, actually, for legacy routers). Once GRA packets can be detected in the forwarding path, the next step is to locate and parse the GRA parameters: the filter type, the action, the key, and the variables from above. These should be encoded as TLVs located somewhere between the end of the network layer header and the beginning of the transport layer payload or SDU. It's tempting in the case of the key and the packet variables to con- sider encoding their offsets and lengths in a TLV that could be used to locate the actual values themselves in the transport header or, more wildly, in the SDU, but this would amount to providing a near Cain/Speakman/Towsley [Page 17] INTERNET-DRAFT Generic Router Assist March 2000 programmatic directive to the network element which is fraught with risk. So note that, in the example of elimination above, the number of loss reports requested would, in this model, be encoded both in its natural place in the transport header, and also as a GRA-specific variable in a separate GRA TLV. An alternative is to establish the location and length of keys and packet variables as attributes of the filtering service defined in the network element itself so that the only vulnerability to the net- work elements would derive from abuse of the setup protocol in the control plane in place of vulnerability in the forwarding path or data plane. The location of GRA TLVs before, inside, or after the transport layer header is at the crux of whether GRA is regarded as integral to a specific layer or as a shim layer between layers. The answer to the question determines not just where to locate the TLVs but also where to implement the service in host protocol stacks. As for forwarding-time implementation of GRA services, we take it as a requirement for this specification that some services must be light-weight enough to be candidates for optimized implementations in hardware or firmware, while others may be sufficiently complex to warrant software implementations. Soft-state-based services that do not maintain per-packet state may be in the first class; services that maintain per-packet state and therefor require a sorted/searchable data structure may be in the second class. Beyond these very GRA-specific issues are the more general control protocol issues of how to interoperate network elements with disparate GRA capabilities, and all of the specifics of defining, enabling, and disabling filters themselves. These issues are best treated once a solid model of the GRA mechanism itself is esta- blished. 4.2. Control Protocol - Filter definition, enabling, and disabling 4.3. Control Protocol - Session path messages and neighbour informa- tion Cain/Speakman/Towsley [Page 18] INTERNET-DRAFT Generic Router Assist March 2000 Abbreviations IIF Incoming Interface NAK Negative Acknowledgement NOOP No Operation OIF Outgoing Interface SDU Service Data Unit SQN Sequence Number TLV Type Length Value Cain/Speakman/Towsley [Page 19] INTERNET-DRAFT Generic Router Assist March 2000 References [ACTIVE] L.H. Lehman, S.J. Garland, D. L. Tennenhouse. ``Active Reli- able Multicast'', Proc. INFOCOM'98, 1998. [BFS] Koichi Yano, Steven McCanne, "The Breadcrumb Forwarding Service and the Digital Fountain Rainbow: Toward a TCP-Friendly Reliable Mul- ticast", UCB/CSD Technical Report No. UCB//CSD-99-1068 [GMTS] Brad Cain, Don Towsley, "Generic Multicast Transport Services (GMTS)", Nortel Networks Technical Report, December 1998. [KKT] S. Kasera, J. Kurose, D. Towsley. ``A Comparison of Server- Based and Receiver-Based Local Recovery Approaches for Scalable Reli- able Multicast'', Proc. INFOCOM'98, 1998. [LABEL] B.N. Levine, J.J. Garcia-Luna-Aceves. ``Improving Internet Multicast with Routing Labels'', Proc. ICNP-97, pp. 241-250, Oct. 1997. [LSM] C. Papadopoulos, G. Parulkar. ``An Error Control Scheme for Large-Scale Multicast Applications'', Proc. INFOCOM'98. [PGM] T. Speakman, D. Farinacci, S. Lin, A. Tweedly. "PGM Reliable Transport Protocol", IETF draft-speakman-pgm-spec-03.txt, June, 1999. [RMTP] J. Lin, S. Paul. ``RMTP: A reliable multicast transport proto- col'', Proc. of IEEE INFOCOM'95, 1995. [SRM] S. Floyd, V. Jacobson, S. McCanne, C. Lin, L. Zhang. ``A Reli- able Multicast Framework for Light-weight Sessions and Application Level Framing'', IEEE/ACM Trans on Networking, vol. 5, 784-803, Dec. 1997. [TKP] D. Towsley, J. Kurose, S. Pingali. ``A comparison of sender- initiated and receiver-initiated reliable multicast protocols'' IEEE JSAC, April 1997. Cain/Speakman/Towsley [Page 20] INTERNET-DRAFT Generic Router Assist March 2000 Revision History draft-ietf-rmt-gra-00.txt October 1999 Original draft. Authors' Addresses Brad Cain bcain@nortelnetworks.com Tony Speakman speakman@cisco.com Don Towsley towsley@cs.umass.edu Cain/Speakman/Towsley [Page 21]