Internet Draft Mick Seaman Expires January 1998 3Com draft-ietf-issll-is802-svc-mapping-00.txt Andrew Smith Extreme Networks Eric Crawley Gigapacket Networks July 1997 Integrated Service Mappings on IEEE 802 Networks Status of this Memo This document is an Internet Draft. Internet Drafts are working documents of the Internet Engineering Task Force (IETF), its Areas, and its Working Groups. Note that other groups may also distribute working documents as Internet Drafts. Internet Drafts are draft documents valid for a maximum of six months. Internet Drafts may be updated, replaced, or obsoleted by other documents at any time. It is not appropriate to use Internet Drafts as reference material or to cite them other than as a "working draft" or "work in progress." Please check the I-D abstract listing contained in each Internet Draft directory to learn the current status of this or any other Internet Draft. Abstract This document describes the support of IETF Integrated Services over LANs built from IEEE 802 network segments which may be interconnected by IEEE 802.1 MAC Bridges (switches) [1]. It describes the practical capabilities and limitations of this technology for supporting Controlled Load [8] and Guaranteed Service [9] using the inherent capabilities of the relevant 802 technologies [5],[6],[15],[16] etc. and the proposed 802.1p queuing features in switches. IEEE P802.1p [2] is a superset of the existing IEEE 802.1D bridging specification. This document provides a functional model for the layer 3 to layer 2 and user-to-network dialogue which supports admission control and defines requirements for interoperability between switches. The special case of such networks where the sender and receiver are located on the same segment is also discussed. This scheme expands on the ISSLL over 802 LANs framework described in Seaman, Smith, Crawley Expires January 1998 [Page 1] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 [7]. It makes reference to a signaling protocol for admission control developed by the ISSLL WG which is known as the "Subnet Bandwidth Manager". This is an extension to the IETF's RSVP protocol [4] and is described in a separate document [10]. 1. Introduction The IEEE 802.1 Interworking Task Group is currently enhancing the basic MAC Service provided in Bridged Local Area Networks (a.k.a. "switched LANs"). As a supplement to the original IEEE MAC Bridges standard [1], the update P802.1p [2] proposes differential traffic class queuing and access to media on the basis of a "user_priority" signaled in frames. In this document we * review the meaning and use of user_priority in LANs and the frame forwarding capabilities of a standard LAN switch. * examine alternatives for identifying layer 2 traffic flows for admission control. * review the options available for policing traffic flows. * derive requirements for consistent traffic class handling in a network of switches and use these requirements to discuss queue handling alternatives for 802.1p and the way in which these meet administrative and interoperability goals. * consider the benefits and limitations of this switched-based approach, contrasting it with full router based RSVP implementation in terms of complexity, utilisation of transmission resources and administrative controls. The model used is outlined in the "framework document" [7] which in summary: * partitions the admission control process into two separable operations: * an interaction between the user of the integrated service and the local network elements ("provision of the service" in the terms of 802.1D) to confirm the availability of transmission resources for traffic to be introduced. * selection of an appropriate user_priority for that traffic on the basis of the service and service parameters to be supported. * distinguishes between the user to network interface above and the mechanisms used by the switches ("support of the service"): these include communication between the switches (network to network signaling). * describes a simple architecture for the provision and support of these services, broken down into components with functional and interface descriptions: * "user" components: a layer-3 to layer-2 negotiation and translation Seaman, Smith, Crawley Expires January 1998 [Page 2] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 component for sending and receiving, with interfaces to other components residing in the station. * processes residing in a bridge/switch to handle admission control and mapping requests, including proposals for actual traffic mappings to user_priority values. * identifies the requirements of a signaling protocol to carry admission control requests between devices. It will be noted that this document is written from the pragmatic viewpoint that there will be a widely deployed network technology and we are evaluating it for its ability to support some or all of the defined IETF integrated services: this approach is intended to ensure development of a system which can provide useful new capabilities in existing (and soon to be deployed) network infrastructures. 2. Goals and Assumptions It is assumed that typical subnetworks that are concerned about quality-of-service will be "switch-rich": that is to say most communication between end stations using integrated services support will pass through at least one switch. The mechanisms and protocols described will be trivially extensible to communicating systems on the same shared media, but it is important not to allow problem generalisation to complicate the practical application that we target: the access characteristics of Ethernet and Token-Ring LANs are forcing a trend to switch-rich topologies. In addition, there have been developments in the area of MAC enhancements to ensure delay- deterministic access on network links e.g. IEEE 802.12 [15] and other proprietary schemes. Note that we illustrate most examples in this document using RSVP as an "upper-layer" QoS signaling protocol but there are actually no real dependencies on this protocol: RSVP could be replaced by some other dynamic protocol or else the requests could be made by network management or other policy entities. In particular, the SBM signaling protocol [10], which is based upon RSVP, is designed to work seamlessly in the service-mapping architecture described in this document and the "Integrated Services over IEEE 802" framework [7]. There may be a heterogeneous mixture of switches with different capabilities, all compliant with IEEE 802.1p, but implementing queuing and forwarding mechanisms in a range from simple 2-queue per port, strict priority, up to more complex multi-queue (maybe even one per- flow) WFQ or other algorithms. The problem is broken down into smaller independent pieces: this may lead to sub-optimal usage of the network resources but we contend that Seaman, Smith, Crawley Expires January 1998 [Page 3] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 such benefits are often equivalent to very small improvements in network efficiency in a LAN environment. Therefore, it is a goal that the switches in the network operate using a much simpler set of information than the RSVP engine in a router. In particular, it is assumed that such switches do not need to implement per-flow queuing and policing (although they might do so). It is a fundamental assumption of the int-serv model that flows are isolated from each other throughout their transit across a network. Intermediate queueing nodes are expected to police the traffic to ensure that it conforms to the pre-agreed traffic flow specification. In the architecture proposed here for mapping to layer-2, we diverge from that assumption in the interests of simplicity: the policing function is assumed to be implemented in the transmit schedulers of the layer-3 devices (end stations, routers). In the LAN environments envisioned, it is reasonable to assume that end stations are "trusted" to adhere to their agreed contracts at the inputs to the network and that we can afford to over-allocate resources at admission -control time to compensate for the inevitable extra jitter/bunching introduced by the switched network itself. These divergences have some implications on the types of receiver heterogeneity that can be supported and the statistical multiplexing gains that might have been exploited, especially for Controlled Load flows: this is discussed in a later section of this document. 3. Non-Goals This document describes service mappings onto existing IEEE- and ANSI- defined standard MAC layers and uses standard MAC-layer services as in IEEE 802.1 bridging. It does not attempt to make use of or describe the capabilities of other proprietary or standard MAC-layer protocols although it should be noted that there exists published work regarding MAC layers suitable for QoS mappings: these are outside the scope of the IETF ISSLL working group charter. 4. User Priority and Frame Forwarding in IEEE 802 Networks 4.1 General IEEE 802 Service Model User_priority is a value associated with the transmission and reception of all frames in the IEEE 802 service model: it is supplied by the sender which is using the MAC service. It is provided along with the data to a receiver using the MAC service. It may or may not be actually carried over the network: Token- Ring/802.5 carries this value (encoded in its FC octet), basic Ethernet/802.3 does not, 802.12 may or may not depending on the frame format in use. 802.1p defines a consistent way to carry this value over the bridged network on Ethernet, Token Ring, Seaman, Smith, Crawley Expires January 1998 [Page 4] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 Demand- Priority, FDDI or other MAC-layer media using an extended frame format. The usage of user_priority is summarised below but is more fully described in section 2.5 of 802.1D [1] and 802.1p [2] "Support of the Internal Layer Service by Specific MAC Procedures" and readers are referred to these documents for further information. If the "user_priority" is carried explicitly in packets, its utility is as a simple label in the data stream enabling packets in different classes to be discriminated easily by downstream nodes without their having to parse the packet in more detail. Apart from making the job of desktop or wiring-closet switches easier, an explicit field means they do not have to change hardware or software as the rules for classifying packets evolve (e.g. based on new protocols or new policies). More sophisticated layer-3 switches, perhaps deployed towards the core of a network, can provide added value here by performing the classification more accurately and, hence, utilising network resources more efficiently or providing better protection of flows from one another: this appears to be a good economic choice since there are likely to be very many more desktop/wiring closet switches in a network than switches requiring layer-3 functionality. The IEEE 802 specifications make no assumptions about how user_priority is to be used by end stations or by the network. In particular it can only be considered a "priority" in a loose sense: although the current 802.1p draft defines static priority queuing as the default mode of operation of switches that implement multiple queues (user_priority is defined as a 3-bit quantity so strict priority queueing would give value 7 = high priority, 0 = low priority). The general switch algorithm is as follows: packets are placed onto a particular queue based on the received user_priority (from the packet if a 802.1p header or 802.5 network was used, invented according to some local policy if not). The selection of queue is based on a mapping from user_priority [0,1,2,3,4,5,6 or 7] onto the number of available queues. Note that switches may implement any number of queues from 1 upwards and it may not be visible externally, except through any advertised int- serv parameters and the switch's admission control behaviour, which user_priority values get mapped internally onto the same vs. different queues. Other algorithms that a switch might implement might include e.g. weighted fair queueuing, round robin. In particular, IEEE makes no recommendations about how a sender should select the value for user_priority: one of the main purposes of this current document is to propose such usage rules and how to communicate the semantics of the values between switches, end- stations and routers. In the remainder of this document we use the term "traffic class" synonymously with user_priority. Seaman, Smith, Crawley Expires January 1998 [Page 5] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 4.2 Ethernet/802.3 There is no explicit traffic class or user_priority field carried in Ethernet packets. This means that user_priority must be regenerated at a downstream receiver or switch according to some defaults or by parsing further into higher-layer protocol fields in the packet. Alternatively, the IEEE 802.1Q encapsulation [11] may be used which provides an explicit traffic class field on top of an basic MAC format. For the different IP packet encapsulations used over Ethernet/802.3, it will be necessary to adjust any admission- control calculations according to the framing and to the padding requirements: Encapsulation Framing Overhead IP MTU bytes/pkt bytes IP EtherType (ip_len<=46 bytes) 64-ip_len 1500 (1500>=ip_len>=46 bytes) 18 1500 IP EtherType over 802.1p/Q (ip_len<=42) 64-ip_len 1500* (1500>=ip_len>=42 bytes) 22 1500* IP EtherType over LLC/SNAP (ip_len<=40) 64-ip_len 1492 (1500>=ip_len>=40 bytes) 24 1492 * note that the draft IEEE 802.1Q specification exceeds the current IEEE 802.3 maximum packet length values by 4 bytes although work is proceeding within IEEE to address this issue. 4.3 Token-Ring/802.5 The token ring standard [6] provides a priority mechanism that can be used to control both the queuing of packets for transmission and the access of packets to the shared media. The priority mechanisms are implemented using bits within the Access Control (AC) and the Frame Control (FC) fields of a LLC frame. The first three bits of the AC field, the Token Priority bits, together with the last three bits of the AC field, the Reservation bits, regulate which stations get access to the ring. The last three bits of the FC field of an LLC frame, the User Priority bits, are obtained from the higher layer in the user_priority parameter when it requests transmission of a packet. This parameter also establishes the Access Priority used by the MAC. The user_priority value is conveyed end-to-end by the User Priority bits in the FC field and is typically preserved through Token-Ring bridges of all types. In all cases, 0 is the lowest priority. Token-Ring also uses a concept of Reserved Priority: this relates to the Seaman, Smith, Crawley Expires January 1998 [Page 6] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 value of priority which a station uses to reserve the token for the next transmission on the ring. When a free token is circulating, only a station having an Access Priority greater than or equal to the Reserved Priority in the token will be allowed to seize the token for transmission. Readers are referred to [14] for further discussion of this topic. A token ring station is theoretically capable of separately queuing each of the eight levels of requested user priority and then transmitting frames in order of priority. A station sets Reservation bits according to the user priority of frames that are queued for transmission in the highest priority queue. This allows the access mechanism to ensure that the frame with the highest priority throughout the entire ring will be transmitted before any lower priority frame. Annex I to the IEEE 802.5 token ring standard recommends that stations send/relay frames as follows: Application user_priority non-time-critical data 0 - 1 - 2 - 3 LAN management 4 time-sensitive data 5 real-time-critical data 6 MAC frames 7 To reduce frame jitter associated with high-priority traffic, the annex also recommends that only one frame be transmitted per token and that the maximum information field size be 4399 octets whenever delay- sensitive traffic is traversing the ring. Most existing implementations of token ring bridges forward all LLC frames with a default access priority of 4. Annex I recommends that bridges forward LLC frames that have a user priorities greater that 4 with a reservation equal to the user priority (although the draft IEEE P802.1p [2] permits network management override this behaviour). The capabilities provided by token ring's user and reservation priorities and by IEEE 802.1p can provide effective support for Integrated Services flows that request QoS using RSVP. These mechanisms can provide, with few or no additions to the token ring architecture, bandwidth guarantees with the network flow control necessary to support such guarantees. For the different IP packet encapsulations used over Token Ring/802.5, it will be necessary to adjust any admission-control calculations according to the framing requirements: Encapsulation Framing Overhead IP MTU Seaman, Smith, Crawley Expires January 1998 [Page 7] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 bytes/pkt bytes IP EtherType over 802.1p/802.1Q 29 4370* IP EtherType over LLC/SNAP 25 4370* *the suggested MTU from RFC 1042 [13] is 4464 bytes but there are issues related to discovering what the maximum supported MTU between any two points both within and between Token Ring subnets. We recommend here an MTU consistent with the 802.5 Annex I recommendation. 4.4 FDDI The Fiber Distributed Data Interface standard [16] provides a priority mechanism that can be used to control both the queuing of packets for transmission and the access of packets to the shared media. The priority mechanisms are implemented using similar mechanisms to Token-Ring described above. The standard also makes provision for "Synchronous" data traffic with strict media access and delay guarantees - this mode of operation is not discussed further here: this is an area within the scope of the ISSLL WG that requires further work. In the remainder of this document we treat FDDI as a 100Mbps Token Ring (which it is) using a service interface compatible with IEEE 802 networks. 4.5 Demand-Priority/802.12 IEEE 802.12 [15] is a standard for a shared 100Mbit/s LAN. Data packets are transmitted using either 803.3 or 802.5 frame formats. The MAC protocol is called Demand Priority. Its main characteristics in respect to QoS are the support of two service priority levels (normal- and high-priority) and the service order: data packets from all network nodes (e.g. end-hosts and bridges/switches) are served using a simple round robin algorithm. If the 802.3 frame format is used for data transmission then user_priority is encoded in the starting delimiter of the 802.12 data packet. If the 802.5 frame format is used then the priority is additionally encoded in the YYY bits of the AC field in the 802.5 packet header (see also section 4.3). Furthermore, the 802.1p/Q encapsulation may also be applied in 802.12 networks with its own user_priority field. Thus, in all cases, switches are able to recover any user_priority supplied by a sender. The same rules apply for 802.12 user_priority mapping through a bridge as with other media types: the only additional information is that "normal" priority is used by default for user_priority values 0 through 4 inclusive and "high" priority is used for user_priority levels 5 through 7: this ensures that the default Token-Ring user_priority level of 4 for 802.5 bridges is mapped to "normal" on 802.12 segments. Seaman, Smith, Crawley Expires January 1998 [Page 8] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 The medium access in 802.12 LANs is deterministic: the demand priority mechanism ensures that, once the normal priority service has been pre- empted, all high priority packets have strict priority over packets with normal priority. In the abnormal situation that a normal-priority packet has been waiting at the front of a MAC transmit queue for a time period longer than PACKET_PROMOTION (200 - 300 ms [15]),its priority is automatically 'promoted' to high priority. Thus, even normal-priority packets have a maximum guaranteed access time to the medium. Integrated Services can be built on top of the 802.12 medium access mechanism. When combined with admission control and bandwidth enforcement mechanisms, delay guarantees as required for a Guaranteed Service can be provided without any changes to the existing 802.12 MAC protocol. Since the 802.12 standard supports the 802.3 and 802.5 frame formats, the same framing overhead as reported in sections 4.2 and 4.3 must be considered in the admission control equations for 802.12 links. 5. Integrated services through layer-2 switches 5.1 Summary of switch characteristics For the sake of illustration, we divide layer-2 bridges/switches into several categories, based on the level of sophistication of their QoS and software protocol capabilities: these categories are not intended to represent all possible implementation choices but, instead, to aid discussion of what QoS capabilities can be expected from a network made of these devices (the basic "class 0" device is included for completeness but cannot really provide useful integrated service). Class 0 - 802.1D MAC bridging - single queue per output port, no separation of traffic classes - Spanning-Tree to remove topology loops (single active path) Class I - 802.1p priority queueuing between traffic classes. - No multicast heterogeneity. - 802.1p GARP/GMRP pruning of individual multicast addresses. Class II As (I) plus: - can map received user_priority on a per-input-port basis to some internal set of canonical values. - can map internal canonical values onto transmitted user_priority on a per-output-port basis giving some limited form of multicast heterogeneity. Seaman, Smith, Crawley Expires January 1998 [Page 9] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 - maybe implements IGMP snooping for pruning. Class III As (II) plus: - per-flow classification - maybe per-flow policing and/or reshaping - more complex transmit scheduling (probably not per-flow) 5.2 Queueing Connectionless packet-based networks in general, and LAN-switched networks in particular, work today because of scaling choices in network provisioning. Consciously or (more usually) unconsciously, enough excess bandwidth and buffering is provisioned in the network to absorb the traffic sourced by higher-layer protocols or cause their transmission windows to run out, on a statistical basis, so that the network is only overloaded for a short duration and the average expected loading is less than 60% (usually much less). With the advent of time-critical traffic such over-provisioning has become far less easy to achieve. Time critical frames may find themselves queued for annoyingly long periods of time behind temporary bursts of file transfer traffic, particularly at network bottleneck points, e.g. at the 100 Mb/s to 10 Mb/s transition that might occur between the riser to the wiring closet and the final link to the user from a desktop switch. In this case, however, if it is known (guaranteed by application design, merely expected on the basis of statistics, or just that this is all that the network guarantees to support) that the time critical traffic is a small fraction of the total bandwidth, it suffices to give it strict priority over the "normal" traffic. The worst case delay experienced by the time critical traffic is roughly the maximum transmission time of a maximum length non-time-critical frame - less than a millisecond for 10 Mb/s Ethernet, and well below an end to end budget based on human perception times. When more than one "priority" service is to be offered by a network element e.g. it supports Controlled-Load as well as Guaranteed Service, the queuing discipline becomes more complex. In order to provide the required isolation between the service classes, it will probably be necessary to queue them separately. There is then an issue of how to service the queues - a combination of admission control and more intelligent queueing disciplines e.g. weighted fair queuing, may be required in such cases. As with the service specifications themselves, it is not the place for this document to specify queuing algorithms, merely to observe that the external behaviour meet the services' requirements. 5.3 Multicast Heterogeneity Seaman, Smith, Crawley Expires January 1998 [Page 10] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 At layer-3, the int-serv model allows heterogeneous multicast flows where different branches of a tree can have different types of reservations for a given multicast destination. It also supports the notion that trees may have some branches with reserved flows and some using best effort (default) service. If we were to treat a layer-2 subnet as a single "network element", as defined in [3], then all of the branches of the distribution tree that lie within the subnet could be assumed to require the same QoS treatment and be treated as an atomic unit as regards admission control etc. with this assumption, the model and protocols already defined by int- serv and RSVP already provide sufficient support for multicast heterogeneity. Note, though, that an admission control request may well be rejected because just one link in the subnet has reached its traffic limit and that this will lead to rejection of the request for the whole subnet. The above approach would, therefore, provide very sub-optimal utilisation of resources given the size and complexity of the layer-2 subnets envisioned by this document. Therefore, it is desirable to support the ability of layer-2 switches to apply QoS differently on different egress branches of a tree that divides at that switch: this is discussed in the following paragraphs. IEEE 802.1D and 802.1p specify a basic model for multicast whereby a switch performs multicast routing decisions based on the destination address: this would produce a list of output ports to which the packet should be forwarded. In its default mode, such a switch would use the user_priority value in received packets (or a value regenerated on a per-input-port basis in the absence of an explicit value) to enqueue the packets at each output port. All of the classes of switch identified above can support this operation. If a switch is selecting per-port output queues based only on the incoming user_priority, as described by 802.1p, it must treat all branches of all multicast sessions within that user_priority class with the same queuing mechanism: no heterogeneity is then possible and this could well lead to the failure of an admission control request for the whole multicast session due to a single link being at its maximum allocation, as described above. Note that, in the layer-2 case as distinct from the layer-3 case with RSVP/int-serv, the option of having some receivers getting the session with the requested QoS and some getting it best effort does not exist as the Class I switches are unable to re-map the user_priority on a per- link basis: this could well become an issue with heavy use of dynamic multicast sessions. If a switch were to implement a separate user_priority mapping at each output port, as described under "Class II switch" above, then some limited form of receiver heterogeneity can be supported e.g. forwarding of traffic as user_priority 4 on one branch where receivers have performed admission control reservations and as user_priority 0 on one where they have not. Seaman, Smith, Crawley Expires January 1998 [Page 11] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 We assume that per-user_priority queuing without taking account of input or output ports is the minimum standard functionality for switches in a LAN environment (Class I switch, as defined above) but that more functional layer-2 or even layer-3 switches (a.k.a. routers) can be used if even more flexible forms of heterogeneity are considered necessary to achieve more efficient resource utilisation: note that the behaviour of layer-3 switches in this context is already well standardised by IETF. 5.4 Override of incoming user_priority In some cases, a network administrator may not trust the user_priority values contained in packets from a source and may wish to map these into some more suitable set of values. Alternatively, due perhaps to equipment limitations or transition periods, values may need to be mapped to/from different regions of a network. Some switches may implement such a function on input that maps received user_priority into some internal set of values (this table is known in 802.1p as the "user_priority regeneration table"). These values can then be mapped using the output table described above onto outgoing user_priority values: these same mappings must also be used when applying admission control to requests that use the user_priority values (see e.g. [10]). More sophisticated approaches may also be envisioned where a device polices traffic flows and adjusts their onward user_priority based on their conformance to the admitted traffic flow specifications. 5.5 Remapping of non-conformant aggregated flows One other topic under discussion in the int-serv context is how to handle the traffic for data flows from sources that are exceeding their currently agreed traffic contract with the network. An approach that shows some promise is to treat such traffic with "somewhat less than best effort" service in order to protect traffic that is normally given "best effort" service from having to back off (such traffic is often "adaptive" using TCP or other congestion control algorithms and it would be unfair to penalise it due to badly behaved traffic from reserved flows which are often set up by non-adaptive applications). One solution here might be to assign normal best effort traffic to one user_priority and to label excess non-conformant traffic as a "lower" user_priority although the re-ordering problems that might arise from doing this may make this solution undesirable, particularly if the flows are using TCP: for this reason the controlled load service recommends dropping excess traffic, rather than re-mapping to a lower priority. This topic is further discussed below. Seaman, Smith, Crawley Expires January 1998 [Page 12] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 6. Selecting traffic classes One fundamental question is "who gets to decide what the classes mean and who gets access to them?" One approach would be for the meanings of the classes to be "well-known": we would then need to standardise a set of classes e.g. 1 = best effort, 2 = controlled- load, 3 = guaranteed (loose delay bound, high bandwidth), 4 = guaranteed (slightly tighter delay) etc. The values to encode in such a table in end stations, in isolation from the network to which they are connected, is problematical: one approach could be to define one user_priority value per int-serv service and leave it at that (reserving the rest of the combinations for future traffic classes - there are sure to be plenty!). We propose here a more flexible mapping: clients ask "the network" which user_priority traffic class to use for a given traffic flow, as categorised by its flow-spec and layer-2 endpoints. The network provides a value back to the requester which is appropriate to the current network topology, load conditions, other admitted flows etc. The task of configuring switches with this mapping (e.g. through network management, a switch-switch protocol or via some network-wide QoS-mapping directory service) is an order of magnitude less complex than performing the same function in end stations. Also, when new services (or other network reconfigurations) are added to such a network, the network elements will typically be the ones to be upgraded with new queuing algorithms etc. and can be provided with new mappings at this time. Given the need for a new session or "flow" requiring some QoS support, a client then needs answers to the following questions: 1. which traffic class do I add this flow to? The client needs to know how to label the packets of the flow as it places them into the network. 2. who do I ask/tell? The proposed model is that a client ask "the network" which user_priority traffic class to use for a given traffic flow. This has several benefits as compared to a model which allows clients to select a class for themselves. 3. how do I ask/tell them? A request/response protocol is needed between client and network: in fact, the request can be piggy-backed onto an admission control request and the response can be piggy-backed onto an admission control acknowledgment: this "one pass" assignment has the benefit of completing the admission control in a timely way and reducing the exposure to changing conditions which could occur if clients cached the knowledge for extensive periods. Seaman, Smith, Crawley Expires January 1998 [Page 13] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 The network (i.e. the first network element encountered downstream from the client) must then answer the following questions: 1. which traffic class do I add this flow to? This is a packing problem, difficult to solve in general, but many simplifying assumptions can be made: presumably some simple form of allocation can be done without a more complex scheme able to dynamically shift flows around between classes. 2. which traffic class has worst-case parameters which meet the needs of this flow? This might be an ordering/comparison problem: which of two service classes is "better" than another? Again, we can make this tractable by observing that all of the current int-serv classes can be ranked (best effort <= Controlled Load <= Guaranteed Service) in a simple manner. If any classes are implemented in the future that cannot be simply ranked then the issue can be finessed by either a priori knowledge about what classes are supported or by configuration. and return the chosen user_priority value to the client. Note that the client may be either an end station, router or a first switch which may be acting as a proxy for a client which does not participate in these protocols for whatever reason. Note also that a device e.g. a server or router, may choose to implement both the "client" as well as the "network" portion of this model so that it can select its own user_priority values: such an implementation would, however, be discouraged unless the device really does have a close tie- in with the network topology and resource allocation policies but would work in some cases where there is known over- provisioning of resources. 7. Flow Identification Some models for int-serv over lower-layers treat layer-2 switches very much as a special case of routers: in particular, that switches along the data path will make packet handling decisions based on the RSVP flow and filter specifications and use them to classify the corresponding data packets. However, filtering to the per-flow level becomes difficult with increasing switch speed: devices with such filtering capabilities are unlikely to have a very different implementation complexity from IP routers and there already exist protocol specifications for those devices. This document argues that "aggregated flow" identification based on user_priority is a useful intermediate point between no QoS and full router-type integrated services and that this be the minimum flow classification capability required of switches. Seaman, Smith, Crawley Expires January 1998 [Page 14] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 8. Reserving Network Resources - Admission Control So far we have not discussed admission control. In fact, without admission control it is possible to assemble a layer-2 LAN of some size capable of supporting real-time services, providing that the traffic fits within certain scaling constraints (relative link speeds, numbers of ports etc. - see below). This is not surprising since it is possible to run a fair approximation to real time services on small LANs today with no admission control or help from encoded priority bits. As an example, imagine a campus network providing dedicated 10 Mbps connections to each user. Each floor of each building supports up to 96 users, organized into groups of 24, with each group being supported by a 100 Mbps downlink to a basement switch which concentrates 5 floors (20 x 100 Mbps) and a data center (4 x 100 Mbps) to a 1 Gbps link to an 8 Gbps central campus switch, which in turn hooks 6 buildings together (with 2 x 1 Gbps full duplex links to support a corporate server farm). Such a network could support 1.5 Mb/s of voice/video from every user to any other user or (for half the population) the server farm, provided the video ran high priority: this gives 3000 users, all with desktop video conferencing running along with file transfer/email etc. In such a network, a discussion as to the best service policy to apply to high and low priority queues may prove academic: while it is true that "normal" traffic may be delayed by bunches of high priority frames, queuing theory tells us that the average queue occupancy in the high priority queue at any switch port will be somewhat less than 1 (with real user behaviour, i.e. not all watching video conferences all the time) it should be far less. A cheaper alternative to buying equipment with a fancy queue service policy may be to buy equipment with more bandwidth to lower the average link utilisation by a few per cent. In practice a number of objections can be made to such a simple solution. There may be long established expensive equipment in the network which does not provide all the bandwidth required. There will be considerable concern over who is allowed to say what traffic is high priority. There may be a wish to give some form of "prioritised" service to crucial business applications, above that given to experimental video-conferencing: in this context, admission control needs to provide administrative control to some level, without making that control so elaborate to implement that it is not simply rejected in favor of providing yet more bandwidth instead. The proposed admission control mechanism requires a query-response interaction with the network returning a "YES/NO" answer and, if successful, a user_priority value with which to tag the data frames of this flow. Seaman, Smith, Crawley Expires January 1998 [Page 15] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 The relevant int-serv specifications describe the parameters which need to be considered when making an admission control decision at each node in the network path between sender and receiver. We discuss how to calculate these parameters for different network technologies below but we do not specify admission control algorithms or mechanisms as to how to progress the admission control process across the network. The proposed IETF protocol for this purpose "Subnet Bandwidth Manager" (SBM) is defined in [10]. Where there are multiple mechanisms in use for allocating resources e.g. some combination of SBM and network management, it will be necessary to ensure that network resources are partitioned amongst the different mechanisms in some way: this could be by configuration or maybe by having the mechanisms allocate from a common resource pool within any device. 9. Mapping of integrated services to layer-2 in layer-3 devices 9.1 Layer-3 Client Model We assume the same client model as int-serv and RSVP where we use the term "client" to mean the entity handling QoS in the layer-3 device at each end of a layer-2 hop (e.g. end-station, router). The sending client itself is responsible for local admission control and scheduling packets onto its link in accordance with the service agreed. As with the current int-serv model, this involves per-flow scheduling (a.k.a. traffic shaping) in every such originating source. The client is running an RSVP process which presents a session establishment interface to applications, signals over the network, programs a scheduler and classifier in the driver and interfaces to a policy control module. In particular, RSVP also interfaces to a local admission control module: it is this entity that we focus on here. Seaman, Smith, Crawley Expires January 1998 [Page 16] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 The following diagram is taken from the RSVP specification[4]: _____________________________ | _______ | | | | _______ | | |Appli- | | | | RSVP | | cation| | RSVP <--------------------> | | <--> | | | | | |process| _____ | | |_._____| | -->Polcy|| | | |__.__._| |Cntrl|| | |data | | |_____|| |===|===========|==|==========| | | --------| | _____ | | | | | ---->Admis|| | _V__V_ ___V____ |Cntrl|| | | | | | |_____|| | |Class-| | Packet | | | | ifier|==>Schedulr|====================> | |______| |________| | data | | |_____________________________| Figure 1 - RSVP in Sending Hosts Note that we illustrate examples in this document using RSVP as the "upper-layer" signaling protocol but there are no actual dependencies on this protocol: RSVP could be replaced by some other dynamic protocol or else the requests could be made by network management or other policy entities. 9.2 Requests to layer-2 ISSLL The local admission control entity within a client is responsible for mapping these layer-3 session-establishment requests into layer-2 language. The upper-layer entity makes a request, in generalised terms, to ISSLL of the form: "May I reserve for traffic with with from to and how should I label it?" where = Sender Tspec (e.g. bandwidth, burstiness, MTU) = FlowSpec Seaman, Smith, Crawley Expires January 1998 [Page 17] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 (e.g. latency, jitter bounds) = IP address(es) = IP address(es) - may be multicast 9.3 At the Layer-3 Sender The ISSLL functionality in the sender is illustrated below and the functions of the box labeled "SBM client" may be summarised as: * maps the endpoints of the conversation to layer-2 addresses in the LAN, so that the client can figure out what traffic is really going where (probably makes reference to the ARP protocol cache for unicast or an algorithmic mapping for multicast destinations). * applies local admission control on outgoing link and driver * formats a SBM request to the network with the mapped addresses and filter/flow specs * receives response from the network and reports the YES/NO admission control answer back to the upper layer entity, along with any negotiated modifications to the session parameters. * saves any returned user_priority to be associated with this session in a "802 header" table: this will be used when adding layer-2 header before sending any future data packet belonging to this session. This table might, for example, be indexed by the RSVP flow identifier. from IP from RSVP ____|____________|____________ | | | | | __V____ ___V___ | | | | | | | | | Addr |<->| | | SBM signaling | |mapping| | SBM |<------------------------> | |_______| |Client | | | ___|___ | | | | | |<->| | | | | 802 | |_______| | | | header| / | | | | |_______| / | | | | | / | | _____ | | | +-----/ | +->|Local| | | __V_V_ _____V__ |Admis| | | | | | | |Cntrl| | | |Class-| | Packet | |_____| | | | ifier|==>Schedulr|======================> | |______| |________| | data |______________________________| Figure 2 - ISSLL in End-station Sender Seaman, Smith, Crawley Expires January 1998 [Page 18] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 9.4 At the Layer-3 Receiver The ISSLL functionality in the receiver is a good deal simpler. It is summarised below and is illustrated by the following picture: * handles any received SBM protocol indications. * applies local admission control to see if a request can be supported with appropriate local receive resources. * passes indications up to RSVP if OK. * accepts confirmations from RSVP and relays them back via SBM signaling towards the requester. * may program a receive classifier and scheduler, if any is used, to identify traffic classes of received packets and accord them appropriate treatment e.g. reserve some buffers for particular traffic classes. * programs receiver to strip any 802 header information from received packets. to RSVP to IP ^ ^ ____|____________|___________ | | | | | __|____ | | | | | | | SBM signaling | | SBM | ___|___ | <-----------------> |Client | | Strip | | | |_______| |802 hdr| | | | \ |_______| | | __v___ \ ^ | | | Local |\ | | | | Admis | \ | | | | Cntrl | \ | | | |_______| \ | | | ______ v___|____ | | |Class-| | Packet | | ===================>| ifier|==>|Scheduler| | data | |______| |_________| | |_____________________________| Figure 3 - ISSLL in End-station Receiver 10. Layer-2 Switch Functions 10.1 Switch Model The model of layer-2 switch behaviour described here uses the terminology of the SBM protocol [10] as an example of an admission control protocol: the model is equally applicable when other mechanisms Seaman, Smith, Crawley Expires January 1998 [Page 19] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 e.g. static configuration, network management are in use for admission control. We define the following entities within the switch: * Local admission control - one of these on each port accounts for the available bandwidth on the link attached to that port. For half-duplex links, this involves taking account of the resources allocated to both transmit and receive flows. For full-duplex, the input port accountant's task is trivial. * Input SBM module: one instance on each port, performs the "network" side of the signaling protocol for peering with clients or other switches. Also holds knowledge of the mappings of int-serv classes to user_priority. * SBM propagation - relays requests that have passed admission control at the input port to the relevant output ports' SBM modules. This will require access to the switch's forwarding table (layer-2 "routing table" cf. RSVP model) and port spanning-tree states. * Output SBM module - forwards requests to the next layer-2 or -3 network hop. * Classifier, Queueing and Scheduler - these functions are basically as described by the Forwarding Process of IEEE 802.1p (see section 3.7 of [2]). The Classifier module identifies the relevant QoS information from incoming packets and uses this, together with the normal bridge forwarding database, to decide to which output queue of which output port to enqueue the packet. In Class I switches, this information is the "regenerated user_priority" parameter which has already been decoded by the receiving MAC service and potentially re-mapped by the 802.1p forwarding process (see description in section 3.7.3 of [2]). This does not preclude more sophisticated classification rules which may be applied in more complex Class III switches e.g. matching on individual int-serv flows. The Queueing and Scheduler module holds the output queues for ports and provides the algorithm for servicing the queues for transmission onto the output link in order to provide the promised int-serv service. Switches will implement one or more output queues per port and all will implement at least a basic strict priority dequeueing algorithm as their default, in accordance with 802.1p. * Ingress traffic class mapper and policing - as described in 802.1p section 3.7. This optional module may check on whether the data within traffic classes are conforming to the patterns currently agreed: switches may police this and discard or re-map packets. The default behaviour is to pass things through unchanged. Seaman, Smith, Crawley Expires January 1998 [Page 20] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 * Egress traffic class mapper - as described in 802.1p section 3.7. This optional module may apply re-mapping of traffic classes e.g. on a per- output port basis. The default behaviour is to pass things through unchanged. These are shown by the following diagram which is a superset of the IEEE 802.1D/802.1p bridge model: _______________________________ | _____ ______ ______ | SBM signaling | | | | | | | | SBM signaling <------------------>| IN |<->| SBM |<->| OUT |<----------------> | | SBM | | prop.| | SBM | | | |_____| |______| |______| | | / | ^ / | | ______________| / | | | | |_____________ | \ / __V__ | | __V__ / | | \ ____/ |Local| | | |Local| / | | \ / |Admis| | | |Admis| / | | \/ |Cntrl| | | |Cntrl| / | | _____V \ |_____| | | |_____| / _____ | | |traff | \ ___|__ V_______ / |egrss| | | |class | \ |Filter| |Queue & | / |traff| | | |map & |=====|==========>|Data- |=| Packet |=|===>|class| | | |police| | | base| |Schedule| | |map | | | |______| | |______| |________| | |_____| | |____^_________|_______________________________|______|______| data in | |data out ========+ +========> Figure 4 - ISSLL in Switches 10.2 Admission Control On reception of an admission control request, a switch performs the following actions, again using SBM as an example: the behaviour is different depending on whether the "Designated SBM" for this segment is within this switch or not - see [10] for a more detailed specification of the DSBM/SBM actions: * if the ingress SBM is the "Designated SBM" for this link/segment, it translates any received user_priority or else selects a layer-2 traffic class which appears compatible with the request and whose use does not violate any administrative policies in force. In effect, it matches up the requested service with those available in each of the user_priority classes and chooses the "best" one. It ensures that, if this reservation is successful, the selected value is passed back to the client. * ingress DSBM observes the current state of allocation of resources on the input port/link and then determines whether the new resource allocation from the mapped traffic class would be excessive. The request Seaman, Smith, Crawley Expires January 1998 [Page 21] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 is passed to the reservation propagator if accepted so far. * if the ingress SBM is not the "Designated SBM" for this link/segment then it passes the request on directly to the reservation propagator * reservation propagator relays the request to the bandwidth accountants on each of the switch's outbound links to which this reservation would apply (implied interface to routing/forwarding database). * egress bandwidth accountant observes the current state of allocation of queueing resources on its outbound port and bandwidth on the link itself and determines whether the new allocation would be excessive. Note that this is only the local decision of this switch hop: each further layer-2 hop through the network gets a chance to veto the request as it passes along. * the request, if accepted by this switch, is then passed on down the line on each output link selected. Any user_priority described in the forwarded request must be translated according to any egress mapping table. * if accepted, the switch must notify the client of the user_priority to use for packets belonging to this flow. Note that this is a "provisional YES" - we assume an optimistic approach here: later switches can still say "NO" later. * if this switch wishes to reject the request, it can do so by notifying the original client (by means of its layer-2 address). 11. Mappings from int-serv service models to IEEE 802 It is assumed that admission control will be applied when deciding whether or not to admit a new flow through a given network element and that a device sending onto a link will be proxying the parameters and admission control decisions on behalf of that link: this process will require the device to be able to determine (by estimation, measurement or calculation) several parameters. It is assumed that details of the potential flow are provided to the device by some means (e.g. a signaling protocol, network management). The service definition specifications themselves provide some implementation guidance as to how to calculate some of these quantities. The accuracy of calculation of these parameters may not be very critical: indeed it is an assumption of this model's being used with relatively simple Class I switches that they merely provide values to describe the device and admit flows conservatively. 11.1 General characterisation parameters There are some general parameters that a device will need to use and/or supply for all service types: * Ingress link Seaman, Smith, Crawley Expires January 1998 [Page 22] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 * Egress links and their MTUs, framing overheads and minimum packet sizes (see media-specific information presented above). * available path bandwidth: updated hop-by-hop by any device along the path of the flow. * minimum latency 11.2 Parameters to implement Guaranteed Service A network element must be able to determine the following parameters: * Constant delay bound through this device (in addition to any value provided by "minimum latency" above) and up to the receiver at the next network element for the packets of this flow if it were to be admitted: this would include any access latency bound to the outgoing link as well as propagation delay across that link. * Rate-proportional delay bound through this device and up to the receiver at the next network element for the packets of this flow if it were to be admitted. * Receive resources that would need to be associated with this flow (e.g. buffering, bandwidth) if it were to be admitted and not suffer packet loss if it kept within its supplied Tspec/Rspec. * Transmit resources that would need to be associated with this flow (e.g. buffering, bandwidth, constant- and rate-proportional delay bounds) if it were to be admitted. 11.3 Parameters to implement Controlled Load A network element must be able to determine the following parameters which can be extracted from [8]: * Receive resources that would need to be associated with this flow (e.g. buffering) if it were to be admitted. * Transmit resources that would need to be associated with this flow (e.g. buffering) if it were to be admitted. 11.4 Parameters to implement Best Effort For a network element to implement best effort service there are no explicit parameters that need to be characterised. 11.5 Mapping to IEEE 802 user_priority There are many options available for mapping aggregations of flows described by int-serv service models (Best Effort, Controlled Load, and Guaranteed are the services considered here) onto user_priority classes. There currently exists very little practical experience with particular mappings to help make a determination as to the "best" mapping. In that spirit, the following options are presented in order to stimulate Seaman, Smith, Crawley Expires January 1998 [Page 23] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 experimentation in this area. Note, this does not dictate what mechanisms/algorithms a network element (e.g. an Ethernet switch) needs to perform to implement these mappings: this is an implementation choice and does not matter so long as the requirements for the particular service model are met. Having said that, we do explore below the ability of a switch implementing strict priority queueing to support some or all of the service types under discussion: this is worthwhile because this is likely to be the most widely deployed dequeueing algorithm in simple switches as it is the default specified in 802.1p. In order to reduce the administrative problems, such a mapping table is held by *switches* (and routers if desired) but generally not by end- station hosts and is a read-write table. The values proposed below are defaults and can be overridden by management control so long as all switches agree to some extent (the required level of agreement requires further analysis). It is possible that some form of network-wide lookup service could be implemented that serviced requests from clients e.g. traffic_class = getQoSbyName("H.323 video") and notified switches of what sorts of traffic categories they were likely to encounter and how to allocate those requests into traffic classes: such mechanisms are for further study. Example: A Simple Scheme user_priority Service 0 "less than" Best Effort 1 Best Effort 2 reserved 3 reserved 4 Controlled Load 5 Guaranteed Service, 100ms bound 6 Guaranteed Service, 10ms bound 7 reserved Table 1 - Example user_priority to service mappings In this proposal, all traffic that uses the controlled load service is mapped to a single 802.1p user_priority whilst that for guaranteed service is placed into one of two user_priority classes with different delay bounds. Unreserved best effort traffic is mapped to another. The use of classes 4, 5 and 6 for Controlled Load and Guaranteed Service is somewhat arbitrary as long as they are increasing. Any two classes greater than Best Effort can be used as long as GS is "greater" than CL although those proposed here have the advantage that, for transit Seaman, Smith, Crawley Expires January 1998 [Page 24] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 through 802.1p switches with only two-level strict priority queuing, they both get "high priority" treatment (the current 802.1p default split is 0-3 and 4-7 for a device with 2 queues). The choice of delay bound is also arbitrary but potentially very significant: this can lead to a much more efficient allocation of resources as well as greater (though still not very good) isolation between flows. The "less than best effort" class might be useful for devices that wish to tag packets that are exceeding a committed network capacity and can be optionally discarded by a downstream device. Note, this is not *required* by any current int-serv models but is under study. The advantage to this approach is that it puts some real delay bounds on the Guaranteed Service without adding any additional complexity to the other services. It still ignores the amount of *bandwidth* available for each class. This should behave reasonably well as long as all traffic for CL and GS flows does not exceed any resource capacities in the device. Some isolation between very delay-critical GS and less critical GS flows is provided but there is still an overall assumption that flows will in general be well- behaved. In addition, this mapping still leaves room for future service models. Expanding the number of classes for CL service is not as appealing since there is no need to map to a particular delay bound. There may be cases where an administrator might map CL onto more classes for particular bandwidths or policy levels. It may also be desirable to further subdivide CL traffic in cases where the it is frequently non-conformant for certain applications. 12. Network Topology Scenarios 12.1 Switched networks using priority scheduling algorithms In general, the int-serv standards work has tried to avoid any specification of scheduling algorithms, instead relying on implementers to deduce appropriate algorithms from the service definitions and on users to apply measurable benchmarks to check for conformance. However, since one standards' body has chosen to specify a single default scheduling algorithm for switches [2], it seems appropriate to examine to some degree, how well this "implementation" might actually support some or all of the int-serv services. If the mappings of Proposal A above are applied in a switch implementing strict priority queueing between the 8 traffic classes (7 = highest) then the result will be that all Guaranteed Service packets will be transmitted in preference to any other service. Controlled Load packets will be transmitted next, with everything else waiting until both of Seaman, Smith, Crawley Expires January 1998 [Page 25] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 these queues are empty. If the admission control algorithms in use on the switch ensure that the sum of the "promised" bandwidth of all of the GS and CL sessions are never allowed to exceed the available link bandwidth then the promised service can be maintained. 12.2 Full-duplex switched networks We have up to now ignored the MAC access protocol. On a full-duplex switched LAN (of either Ethernet or Token-Ring types - the MAC algorithm is, by definition, unimportant) this can be factored in to the characterisation parameters advertised by the device since the access latency is well controlled (jitter = one largest packet time). Some example characteristics (approximate): Type Speed Max Pkt Max Access Length Latency Ethernet 10Mbps 1.2ms 1.2ms 100Mbps 120us 120us 1Gbps 12us 12us Token-Ring 4Mbps 9ms 9ms 16Mbps 9ms 9ms FDDI 100Mbps 360us 8.4ms Demand-Priority 100Mbps 120us 253us Table 2 - Full-duplex switched media access latency These delays should be also be considered in the context of speed- of- light delays of e.g. ~400ns for typical 100m UTP links and ~7us for typical 2km multimode fibre links. Therefore we see Full-Duplex switched network topologies as offering good QoS capabilities for both Controlled Load and Guaranteed Service when supported by suitable queueing strategies in the switch nodes. 12.3 Shared-media Ethernet networks We have not mentioned the difficulty of dealing with allocation on a single shared CSMA/CD segment: as soon as any CSMA/CD algorithm is introduced then the ability to provide any form of Guaranteed Service is seriously compromised in the absence of any tight coupling between the multiple senders on the link. There are a number of reasons for not offering a better solution for this issue. Firstly, we do not believe this is a truly solvable problem: it would seem to require a new MAC protocol. There have been proposals for enhancements to the MAC layer protocols e.g. BLAM and enhanced flow- control in IEEE 802.3; IEEE 802.1 has examined research showing Seaman, Smith, Crawley Expires January 1998 [Page 26] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 disappointing simulation results for performance guarantees on shared CSMA/CD Ethernet without MAC enhancements. However, any solution involving a new "software MAC" running above the traditional 802.3 MAC or other proprietary MAC protocols is clearly outside the scope of the work of the ISSLL WG and this document. Secondly, we are not convinced that it is really an interesting problem. While not everyone in the world is buying desktop switches today and there will be end stations living on repeated segments for some time to come, the number of switches is going up and the number of stations on repeated segments is going down. This trend is proceeding to the point that we may be happy with a solution which assumes that any network conversation requiring resource reservations will take place through at least one switch (be it layer-2 or layer-3). Put another way, the easiest QoS upgrade to a layer-2 network is to install segment switching: only when this has been done is it worthwhile to investigate more complex solutions involving admission control. Thirdly, in the core of the network (as opposed to at the edges), there does not seem to be wide deployment of repeated segments as opposed to switched solutions. There may be special circumstances in the future (e.g. Gigabit buffered repeaters) but these have differing characteristics to existing CSMA/CD repeaters anyway. Type Speed Max Pkt Max Access Length Latency Etherne 10Mbps 1.2ms unbounded 100Mbps 120us unbounded 1Gbps 12us unbounded Table 3 - Shared Ethernet media access latency 12.4 Half-duplex switched Ethernet networks Many of the same arguments for sub-optimal support of Guaranteed Service apply to half-duplex switched Ethernet as to shared media: in essence, this topology is a medium that *is* shared between at least two senders contending for each packet transmission opportunity. Unless these are tightly coupled and cooperative then there is always the chance that the best-effort traffic of one will interfere with the important traffic of the other. Such coupling would seem to need some form of modifications to the MAC protocol (see above). Notwithstanding the above, half-duplex switched topologies do seem to offer the chance to provide Controlled Load service: with the knowledge that there are only a small limited number (e.g. two) of potential senders that are both using prioritisation for their CL traffic (with admission control for those CL flows based on the knowledge of the Seaman, Smith, Crawley Expires January 1998 [Page 27] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 number of potential senders) over best effort, the media access characteristics, whilst not deterministic in the true mathematical sense, are somewhat predictable. This is probably a close enough approximation to CL to be useful. Type Speed Max Pkt Max Access Length Latency Ethernet 10Mbps 1.2ms unbounded 100Mbps 120us unbounded 1Gbps 12us unbounded Table 4 - Half-duplex switched Ethernet media access latency 12.5 Half-duplex and shared Token Ring networks In a shared Token Ring network, the network access time for high priority traffic at any station is bounded and is given by (N+1)*THTmax, where N is the number of stations sending high priority traffic and THTmax is the maximum token holding time [14]. This assumes that network adapters have priority queues so that reservation of the token is done for traffic with the highest priority currently queued in the adapter. It is easy to see that access times can be improved by reducing N or THTmax. The recommended default for THTmax is 10 ms [6]. N is an integer from 2 to 256 for a shared ring and 2 for a switched half duplex topology. A similar analysis applies for FDDI. Using default values gives: Type Speed Max Pkt Max Access Length Latency Token-Ring 4/16Mbps shared 9ms 2570ms 4/16Mbps switched 9ms 30ms FDDI 100Mbps 360us 8ms Table 5 - Half-duplex and shared Token-Ring media access latency Given that access time is bounded, it is possible to provide an upper bound for end-to-end delays as required by Guaranteed Service assuming that traffic of this class uses the highest priority allowable for user traffic. The actual number of stations that send traffic mapped into the same traffic class as GS may vary over time but, from an admission control standpoint, this value is needed a priori. The admission control entity must therefore use a fixed value for N, which may be the total number of stations on the ring or some lower value if it is desired to keep the offered delay guarantees smaller. If the value of N used is lower than the total number of stations on the ring, admission control must ensure that the number of stations sending high priority Seaman, Smith, Crawley Expires January 1998 [Page 28] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 traffic never exceeds this number. This approach allows admission control to estimate worst case access delays assuming that all of the N stations are sending high priority data even though, in most cases, this will mean that delays are significantly overestimated. Assuming that Controlled Load flows use a traffic class lower than that used by GS, no upper-bound on access latency can be provided for CL flows. However, CL flows will receive better service than best effort flows. Note that, on many existing shared token rings, bridges will transmit frames using an Access Priority (see section 4.3) value 4 irrespective of the user_priority carried in the frame control field of the frame. Therefore, existing bridges would need to be reconfigured or modified before the above access time bounds can actually be used. 12.6 Half-duplex and shared Demand-Priority networks In 802.12 networks, communication between end-nodes and hubs and between the hubs themselves is based on the exchange of link control signals. These signals are used to control the shared medium access. If a hub, for example, receives a high-priority request while another hub is in the process of serving normal- priority requests, then the service of the latter hub can effectively be pre-empted in order to serve the high-priority request first. After the network has processed all high- priority requests, it resumes the normal-priority service at the point in the network at which it was interrupted. The time needed to preempt normal-priority network service (the high- priority network access time) is bounded: the bound depends on the physical layer and on the topology of the shared network. The physical layer has a significant impact when operating in half- duplex mode as e.g. used across unshielded twisted-pair cabling (UTP) links, because link control signals cannot be exchanged while a packet is transmitted over the link. Therefore the network topology has to be considered since, in larger shared networks, the link control signals must potentially traverse several links (and hubs) before they can reach the hub which possesses the network control. This may delay the preemption of the normal priority service and hence increase the upper bound that may be guaranteed. Upper bounds on the high-priority access time are given below for a UTP physical layer and a cable length of 100 m between all end- nodes and hubs using a maximum propagation delay of 570ns as defined in [15]. These values consider the worst case signaling overhead and assume the transmission of maximum-sized normal- priority data packets while the normal-priority service is being pre-empted. Seaman, Smith, Crawley Expires January 1998 [Page 29] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 Type Speed Max Pkt Max Access Length Latency Demand Priority 100Mbps, 802.3pkt, UTP 120us 253us 802.5pkt, UTP 360us 733us Table 6 - Half-duplex switched Demand-Priority UTP access latency Shared 802.12 topologies can be classified using the hub cascading level "N". The simplest topology is the single hub network (N = 1). For a UTP physical layer, a maximum cascading level of N = 5 is supported by the standard. Large shared networks with many hundreds nodes can however already be built with a level 2 topology. The bandwidth manager could be informed about the actual cascading level by using network management mechanisms and use this information in its admission control algorithms. Type Speed Max Pkt Max Access Topology Length Latency Demand Priority 100Mbps, 802.3pkt 120us 262us N=1 120us 554us N=2 120us 878us N=3 120us 1.24ms N=4 120us 1.63ms N=5 Demand Priority 100Mbps, 802.5pkt 360us 722us N=1 360us 1.41ms N=2 360us 2.32ms N=3 360us 3.16ms N=4 360us 4.03ms N=5 Table 7 - Shared Demand-Priority UTP access latency In contrast to UTP, the fibre-optic physical layer operates in dual simplex mode: Upper bounds for the high-priority access time are given below for 2 km multimode fibre links with a propagation delay of 10 us. Type Speed Max Pkt Max Access Length Latency Demand Priority 100Mbps,802.3pkt,Fibre 120us 139us 802.5pkt,Fibre 360us 379us Table 8 - Half-duplex switched Demand-Priority Fibre access latency For shared-media with distances of 2km between all end-nodes and hubs, the 802.12 standard allows a maximum cascading level of 2. Higher levels of cascaded topologies are supported but require a reduction of the Seaman, Smith, Crawley Expires January 1998 [Page 30] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 distances [15]. Type Speed Max Pkt Max Access Topology Length Latency Demand Priority 100Mbps,802.3pkt 120us 160us N=1 120us 202us N=2 Demand Priority 100Mbps,802.5pkt 360us 400us N=1 360us 682us N=2 Table 9 - Shared Demand-Priority Fibre access latency The bounded access delay and deterministic network access allow the support of service commitments required for Guaranteed Service and Controlled Load, even on shared-media topologies. The support of just two priority levels in 802.12, however, limits the number of services that can simultaneously be implemented across the network. 13. Signaling protocol The mechanisms described in this document make use of a signaling protocol for devices to communicate their admission control requests across the network: the service definitions to be provided by such a protocol e.g. [10] are described below. Below, we illustrate the primitives and information that need to be exchanged with such a signaling protocol entity - in all these examples, appropriate delete/cleanup mechanisms will also have to be provided for when sessions are torn down. 13.1 Client service definitions The following interfaces can be identified from Figures 2 and 3: * SBM <-> Address mapping This is a simple lookup function which may cause ARP protocol interactions, may be just a lookup of an existing ARP cache entry or may be an algorithmic mapping. The layer-2 addresses are needed by SBM for inclusion in its signaling messages to/from switches which avoids the switches having to perform the mapping and, hence, have knowledge of layer-3 information for the complete subnet: l2_addr = map_address( ip_addr ) * SBM <-> Session/802 header Seaman, Smith, Crawley Expires January 1998 [Page 31] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 This is for notifying the transmit path of how to add layer-2 header information e.g. user_priority values to the traffic of each outgoing flow: the transmit path will provide the user_priority value when it requests a MAC-layer transmit operation for each packet (user_priority is one of the parameters passed in the packet transmit primitive defined by the IEEE 802 service model): bind_l2_header( flow_id, user_priority ) * SBM <-> Classifier/Scheduler This is for notifying transmit classifier/scheduler of any additional layer-2 information associated with scheduling the transmission of a flow packets: this primitive may be unused in some implementations or it may be used, for example, to provide information to a transmit scheduler that is performing per- traffic_class scheduling in addition to the per-flow scheduling required by int-serv: the l2_header may be a pattern (additional to the FilterSpec) to be used to identify the flow's traffic. bind_l2schedulerinfo( flow_id, , l2_header, traffic_class ) * SBM <-> Local Admission Control For applying local admission control for a session e.g. is there enough transmit bandwidth still uncommitted for this potential new session? Are there sufficient receive buffers? This should commit the necessary resources if OK: it will be necessary to release these resources at a later stage if the session setup process fails. This call would be made by a segment's Designated SBM for example: status = admit_l2session( flow_id, Tspec, FlowSpec ) * SBM <-> RSVP - this is outlined above in section 9.2 and fully described in [10]. * Management Interfaces Some or all of the modules described by this model will also require configuration management: it is expected that details of the manageable objects will be specified by future work in the ISSLL WG. 13.2 Switch service definitions The following interfaces are identified from Figure 4: * SBM <-> Classifier Seaman, Smith, Crawley Expires January 1998 [Page 32] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 This is for notifying receive classifier of how to match up incoming layer-2 information with the associated traffic class: it may in some cases consist of a set of read-only default mappings: bind_l2classifierinfo( flow_id, l2_header, traffic_class ) * SBM <-> Queue and Packet Scheduler This is for notifying transmit scheduler of additional layer-2 information associated with a given traffic class (it may be unused in some cases - see discussion in previous section): bind_l2schedulerinfo( flow_id, l2_header, traffic_class ) * SBM <-> Local Admission Control As for host above. * SBM <-> Traffic Class Map and Police Optional configuration of any user_priority remapping that might be implemented on ingress to and egress from the ports of a switch (note that, for Class I switches, it is likely that these mappings will have to be consistent across all ports): bind_l2ingressprimap( inport, in_user_pri, internal_priority ) bind_l2egressprimap( outport, internal_priority, out_user_pri ) Optional configuration of any layer-2 policing function to be applied on a per-class basis to traffic matching the l2_header. If the switch is capable of per-flow policing then existing int- serv/RSVP models will provide a service definition for that configuration: bind_l2policing( flow_id, l2_header, Tspec, FlowSpec ) * SBM <-> Filtering Database SBM propagation rules need access to the layer-2 forwarding database to determine where to forward SBM messages (analogous to RSRR interface in L3 RSVP): output_portlist = lookup_l2dest( l2_addr ) * Management Interfaces Some or all of the modules described by this model will also require configuration management: it is expected that details of the manageable objects will be specified by future work in the ISSLL WG. Seaman, Smith, Crawley Expires January 1998 [Page 33] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 14. Compatibility and Interoperability with existing equipment Switches using layer-2-only standards (e.g. 802.1p) will have to cooperate with routers and layer-3 switches. Wide deployment of such 802.1p switches will occur in a number of roles in the network: "desktop switches" provide dedicated 10/100 Mbps links to end stations and high speed core switches will act as central campus switching points for layer-3 devices. Layer-2 devices will have to operate in all of the following scenarios: * every device along a network path is layer-3 capable and intrusive into the full data stream * only the edge devices are pure layer-2 * every alternate device lacks layer-3 functionality * most devices lack layer-3 functionality except for some key control points such as router firewalls, for example. Of course, where int-serv flows pass through equipment which is ignorant of priority queuing and which places all packets through the same queuing/overload-dropping path, it is obvious that some of the characteristics of the flow get more difficult to support. Suitable courses of action in the cases where sufficient bandwidth or buffering is not available are of the form: * buy more (and bigger) routers * buy more capable switches * rearrange the network topology: 802.1Q VLANs [11] may help here. * buy more bandwidth It would also be possible to pass more information between switches about the capabilities of their neighbours and to route around non- QoS-capable switches: such methods are for further study. 15. Justification An obvious comment is that this is all too complex, it's what RSVP is doing already, why do we think we can do better by reinventing the solution to this problem at layer-2? The key is that there are a number of simple layer-2 scenarios that cover a considerable proportion of the real QoS problems that will occur and a solution that covers nearly all of the problems at significantly lower cost is beneficial: full RSVP/int-serv with per-flow queueing in strategically-positioned high-function switches or routers may be needed to completely solve all issues but devices implementing the architecture Seaman, Smith, Crawley Expires January 1998 [Page 34] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 described in this document will allow a significantly simpler network. 16. References [1] ISO/IEC 10038, ANSI/IEEE Std 802.1D-1993 "MAC Bridges" [2] "Supplement to MAC Bridges: Traffic Class Expediting and Dynamic Multicast Filtering", May 1997, IEEE P802.1p/D6 [3] "Integrated Services in the Internet Architecture: an Overview" RFC1633, June 1994 [4] "Resource Reservation Protocol (RSVP) - Version 1 Functional Specification", Internet Draft, June 1997 [5] "Carrier Sense Multiple Access with Collision Detection (CSMA/CD) Access Method and Physical Layer Specifications" ANSI/IEEE Std 802.3-1985. [6] "Token-Ring Access Method and Physical Layer Specifications" ANSI/IEEE Std 802.5-1995 [7] "A Framework for Providing Integrated Services Over Shared and Switched LAN Technologies", Internet Draft, May 1997 [8] "Specification of the Controlled-Load Network Element Service", Internet Draft, May 1997, [9] "Specification of Guaranteed Quality of Service", Internet Draft, February 1997, [10] "SBM (Subnet Bandwidth Manager): A Proposal for Admission Control over Ethernet", Internet Draft, July 1997 [11] "Draft Standard for Virtual Bridged Local Area Networks", May 1997, IEEE P802.1Q/D6 [12] "General Characterization Parameters for Integrated Service Network Elements", Internet Draft, July 1997 Seaman, Smith, Crawley Expires January 1998 [Page 35] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 [13] "A Standard for the Transmission of IP Datagrams over IEEE 802 Networks", RFC 1042, February 1988 [14] C. Bisdikian, B. V. Patel, F. Schaffa, and M Willebeek-LeMair, The Use of Priorities on Token-Ring Networks for Multimedia Traffic, IEEE Network, Nov/Dec 1995. [15] "Demand Priority Access Method, Physical Layer and Repeater Specification for 100Mbit/s", IEEE Std. 802.12-1995. [16] "Fiber Distributed Data Interface", ANSI Std. X3.139-1987 17. Security Considerations Implementation of the model described in this memo creates no known new avenues for malicious attack on the network infrastructure although readers are referred to section 2.8 of the RSVP specification for a discussion of the impact of the use of admission control signaling protocols on network security. 18. Acknowledgments This document draws heavily on the work of the ISSLL WG of the IETF and the IEEE P802.1 Interworking Task Group. In particular, it relies on previous work on Token-Ring/802.5 from Anoop Ghanwani, Wayne Pace and Vijay Srinivasan and on Demand-Priority/802.12 from Peter Kim. 19. Authors' addresses Mick Seaman 3Com Corp. 5400 Bayfront Plaza Santa Clara CA 95052-8145 USA +1 (408) 764 5000 mick_seaman@3com.com Andrew Smith Extreme Networks 10460 Bandley Drive Cupertino CA 95014 USA Seaman, Smith, Crawley Expires January 1998 [Page 36] INTERNET DRAFT Int-Serv over IEEE 802.1D/p July 1997 +1 (408) 863 2821 andrew@extremenetworks.com Eric Crawley Gigapacket Networks 25 Porter Rd. Littleton MA 01460 USA +1 (508) 486 0665 esc@gigapacket.com Seaman, Smith, Crawley Expires January 1998 [Page 37]