Classifications

• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/52—Queue scheduling by attributing bandwidth to queues
  • H04L47/522—Dynamic queue service slot or variable bandwidth allocation
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/52—Queue scheduling by attributing bandwidth to queues
  • H04L47/521—Static queue service slot or fixed bandwidth allocation
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/52—Queue scheduling by attributing bandwidth to queues
  • H04L47/525—Queue scheduling by attributing bandwidth to queues by redistribution of residual bandwidth
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/56—Queue scheduling implementing delay-aware scheduling
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/62—Queue scheduling characterised by scheduling criteria
  • H04L47/6215—Individual queue per QOS, rate or priority
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/62—Queue scheduling characterised by scheduling criteria
  • H04L47/625—Queue scheduling characterised by scheduling criteria for service slots or service orders
  • H04L47/627—Queue scheduling characterised by scheduling criteria for service slots or service orders policing
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
  • H04L47/00—Traffic control in data switching networks
  • H04L47/50—Queue scheduling
  • H04L47/62—Queue scheduling characterised by scheduling criteria
  • H04L47/6285—Provisions for avoiding starvation of low priority queues

Definitions

• the present invention relates to a packet network and to elements and hosts therein.
• a user Prior to sending any information across a traditional connection-oriented network, a user is allocated a circuit, either by provision or by on-demand signalling. During the allocation phase it can be arranged for the circuit to meet user specified performance criteria and this is typically controlled by the connection admission rules. Consequently, once an appropriate circuit has been established information can be sent between host machines to a specified Quality of Service (QoS) .
• QoS Quality of Service
• connection admission control In contrast a traditional connectionless network has no requirement for circuit allocation and such networks do not incorporate connection admission control. This has the result that, during periods of network congestion, the meeting of performance criteria, and thus the satisfaction of Quality of Service requirements (such as end to end delay and delay variation) cannot be guaranteed.
• ATM Asynchronous Transfer Mode
• ATM standards have tended to concentrate on service class optimisations, rather than taking a broader network view. For instance optimising bearer utilisation does not necessarily require the bandwidth utilisation to be optimised for each individual service class supported. More broadly it seems likely that applications are likely to evolve at a more rapid pace than is feasible for networks to track.
• Buffer size needed to achieve zero cell loss is indeterminate when using statistical multiplexing. • The fixed ATM cell size is unlikely to suit all services.
• Cell header size is a further bandwidth overhead in addition to that already imposed by any higher layer protocols (for AA 5 adaptation the cell header wastes ten percent of available bandwidth, for other AAL's it is higher).
• IS integrated services
• the IS Internet will support a number of service classes with particular QoS provisions.
• Guaranteed Service which supports a guaranteed bandwidth and well defined delay bound
• Controlled Load which supports a more loose guarantee of bandwidth
• BE Best Effort
• flow is used to denote a stream of one or more packets produced by a given application and transmitted with a particular QoS requirement.
• the IS Internet will require flow state information in all the network routers.
• Guaranteed Service for future networks include
• WFQ delay control offered by WFQ may become negligible when GS is operating under conditions of time contention rather than bandwidth contention (i e. no statistical multiplexing) •
• a packet network element comprising at least one input for receiving flow based packets; at least one output of predetermined bandwidth; wherein a received packet is associable with a first or second class of service; means for directing each received packet on the basis of its class to a first or a second corresponding packet buffer, said first packet buffer being allocated a predetermined portion of the output bandwidth; said second packet buffer being allocated the remaining portion of the output bandwidth; bandwidth requirement determination means for determining a bandwidth requirement associated with at least said first class flows; means for allowing admission of the first class flow packets to the first packet buffer if said bandwidth requirement can be met; and means for directing packets from the first and second packet buffers to an output.
• said means for allowing admission is operable to apply a peak rate test, allowing admission to the first buffer if the currently unused portion of said predetermined portion of the output bandwidth is able to meet said peak rate bandwidth requirement
• this allows the formulation of a fixed delay bound whilst avoiding the potential processing bottlenecks that may limit the throughput of more complex schemes - first class flow packets admitted to the first packet buffer will be provided the guaranteed delay bound. Furthermore this is achieved without per flow scheduling and hence provides for scalability which is of immediate and recognised concern in building large networks.
• said means for allowing admission is operable to apply said peak rate test and to apply a buffer-fill test, allowing admission if said first packet buffer has sufficient space to accept another flow.
• Such embodiments have the advantage that they can provide a guaranteed delay bound even when only a small number of first class flows are handled by the element.
• the buffer-fill test is likely to be unnecessary in high-speed core network elements that handle a large number of first class flows.
• said admission allowing means is arranged to allow admission of the first class flow packets to the first packet buffer only if the number of flows the first packet buffer is sized to accommodate minus the number of flows currently accommodated is at least unity and if the free portion of the output bandwidth allocated to the first packet buffer minus the proposed peak rate flow bandwidth requirement is greater than or equal to zero.
• Some embodiments have a class determining means for determining whether a received packet is associated with a first or second class of service. This is required in elements where the association of a packet with a given class is not determmable from, say, the interface on which the packet arrives.
• said class determining means is arranged to read a class identifying portion from a said received packet.
• bandwidth requirement determining means is arranged to read a peak rate flow bandwidth requirement information portion from a said received packet
• bandwidth requirement determining means is arranged to determine particular peak rate flow bandwidth requirement values for respective single packet flows.
• a host element for use in association with a packet network comprising: means for generating packet based flows and for associating each flow with a respective selected associated first or second class of service; a first packet buffer arranged to receive packets associated with the first class of service; means for controlling the first packet buffer size; a second packet buffer arranged to receive packets associated with the second class of service; and means for directing packets from the first and second packet buffers to an output arranged to ensure that the first class packet flow rate does not exceed a selected peak rate bandwidth.
• a host architecture advantageously provides for the creation of flows with a strictly bounded peak rate bandwidth removing the necessity for more complex traffic descriptors.
• packets may be created such as to provide for 'signalling on the fly', advantageously removing the need for a separate network signalling phase.
• packets may be created such as to provide for 'signalling on the fly' , advantageously removing the need for a separate network signalling phase.
• FIG. 1 illustrates different architectural configurations of network elements
• Figure 2 illustrates different network architectures including the network elements illustrated in Figure 1 ;
• Figure 3 illustrates relative delays in a network between a sending host and a receiving host as a function of differing network implementation;
• Figure 4 illustrates an exemplary packet structure for use in conjunction with the above
• Figure 5 illustrates an exemplary host architectural configuration
• Figure 6 illustrates a comparison of two traffic shaping schemes
• Figure 7 represents an exemplary flow profile.
• a sending host and a receiving host are connected by means of a network comprising elements or nodes and interconnecting links.
• An application running on the sending host produces data packets or datagrams, the term 'flow' being applied to denote a stream of one or more packets produced by the application and transmitted with a particular Quality of Service (QoS) requirement.
• QoS Quality of Service
• the service requirement of each respective flow is chosen for the purposes of transmission across the network as one from at least two defined service classes with associated Quality of Service (QoS) levels, typically one of either a so called bounded delay class or a so called best effort class.
• QoS Quality of Service
• the network must be able to deliver the constituent packets end to end in a time bounded by the maximum stated delay If the flow is to be transmitted across the network in best effort mode then the network will deliver the constituent packets without a maximum delay bound being specified.
• FIG. 1 (a) to 1 (d) A number of network elements or nodes of alternative architecture are illustrated in Figures 1 (a) to 1 (d) . These elements or nodes may, for example, be implemented using well known router technology, a typical example being those available from Cisco Inc.
• the network element ( 1 ) illustrated in Figure 1 (a) has one or more packet input ports (2) connected to input links (not shown) to receive packet based flows.
• a mode identifier (3) is provided to determine for each packet received from an input (2) with which one of the two modes the packet is associated and then to direct the packet to an appropriate packet buffer.
• a single mode identifier (3) may be provided in respect of all the input ports or alternatively more than one may be provided
• This packet identification may then initiate further processing in a processing element (6) .
• an admission decision typically a connection admission control (CAC) decision may be made as to whether or not to admit the packets of the new flow to one of the first buffers (4) .
• CAC connection admission control
• 'Bounded delay' packets may also be directed to the second packet buffer if they have been refused admission to the first packet buffer as will be discussed below.
• Each of the first and second buffers (4, 7) have buffer playout elements (8, 9) which are controlled respectively by one or more scheduling elements ( 1 0).
• Each scheduling element ( 10) is controlled by means of an appropriate scheduling algorithm.
• Each buffer playout element is connected to an output port ( 1 1 , 1 2) which thence connects to an output link (not shown) .
• the output ports ( 1 1 , 1 2) associated with the one or more bounded delay buffers (4) and the one or more best effort buffers (7) respectively are allocated predetermined shares of the total output bandwidth.
• each such network element ( 1 ) may have multiple inputs (2) as to receive flows from several hosts or other network elements.
• the network element ( 1 ) may have multiple outputs ( 1 1 , 1 2) , whereby a single bounded delay buffer (4) will exist for each bounded delay mode output port ( 1 1 ) and similarly for the best effort buffer (7) . If the associated output links (not shown) are physically distinct then scheduling will be completely independent. If, however, the output links for the bounded delay and best effort traffic are merely virtually separate then, depending on the technology, a range of multiplexing options may present itself.
• the bounded delay and best effort buffers (4, 7) may also be virtually rather than physically separated.
• the network element ( 1 3) illustrated in Figure Kb) is similar to that shown in Figure 1 (a) but is provided with one or more first and second input ports (14, 1 5). Consequently the explicit mode identifier (3) of the Figure 1 (a) depiction need not be utilised as the sets of inputs ( 14, 1 5) themselves may now be thought of as associated with the bounded delay and best effort modes respectively. However an explicit mode identifier (3) of the Figure 1 (a) form may still be useful to act in a policing function to ensure that all incoming packets are of the correct mode.
• Each of the first input ports ( 1 4) , associated with the bounded delay mode, is therefore directly connected with a packet identifier ( 1 6) with associated additional processing element ( 1 7) as discussed above with regard to Figure 1 (a) before connection to a bounded delay buffer ( 1 8)
• Each of the second input ports ( 1 5) associated with the best effort mode is connected directly with a best effort buffer ( 1 9) .
• the network element (25) illustrated in Figure 1 (c) is similar to that shown in Figure Kb) in terms of input, packet identifier, additional processing, bounded delay and best effort buffer elements (26, 27, 28, 29, 30, 31 ) but is only provided with a single set of output ports (32) associated with both the bounded delay and best effort modes.
• a dual buffer structure (30, 31 ) will be provided Consequently for each output port (32) the respective bounded delay and best effort buffers are played out by a single buffer playout element (33), under the control of a scheduling element (34), which is then connected to this single output port (32) .
• Simple priority scheduling always selects the next packet from the first buffer (30), if the first buffer (30) contains at least one packet. In this way the best possible service is given to the first buffer (30) in respect of complete packets. Where the bounded delay traffic is a small proportion of the total, the effect of this on the best effort traffic may be relatively minor.
• the network element (35) illustrated in Figure 1 (d) is provided with the one or more first inputs (36) and hence mode identifiers (37) and associated packet identifier, additional processing, bounded delay and best effort buffer structure (38, 39, 40, 41 ) of Figure 1 (a) and the single set of outputs (42) and hence single buffer playout elements (43) and scheduling element (44) of Figure Kc).
• the ability of the network element ( 1 , 1 3, 25, 35) to offer the bounded delay mode of operation will depend on ensuring that the sum of the peak-rates of all contending flows does not exceed the fraction of the output bandwidth allocated to bounded-delay traffic.
• Specifying peak-rate bandwidth for a flow shall be taken to mean that the host generating the flow is required to shape the traffic in terms of packet separation time as if the host's output link speed were constrained to the specified peak-bandwidth.
• the means to control the peak-rate sum may be either appropriate dimensioning, or the application of CAC rules. However, control of the peak-rate sum is a necessary but not always sufficient condition to provide bounded-delay operation. Depending on the particular configuration of the element, it may also be necessary to control the number and packet-size of bounded-delay flows.
• Equations 1 A and 1 B below represent a multi-packet delay bound for an element, where: link ate is the operating speed of the output link being considered, mux atio is the ratio of output port link-speed to input-port link-speed, N ⁇ p is the number of input ports carrying Bounded-delay (BD) class flows and which are driving the output links, N violat ows is the maximum number of bounded-delay flows, / max is maximum size of bounded-delay class packets, and / be the maximum size of best-effort packets: for mux_ratio ⁇ N ip
• Equation 1 A and Equation 1 B can be used as the basis for CAC rules whereby new flows can be tested as to whether they can be admitted without violating a specified delay bound.
• Equation 1 A and Equation 1 B represent the worst-case distribution of flows across input ports.
• Equations 1 A and 1 B are sufficiently general to apply to a range of situations. However, at certain particular cases of interest they reduce to a simplified form, and for the purposes of illustration, two examples are given. In each example the term / de /link_rate has been assumed small enough to be neglected.
• Equation 1 A Equation 1 A reduces to:
• the delay across this element is bounded automatically, in that a delay bound can be set even with no knowledge of the actual number of flows.
• a peak bandwidth CAC test may be expressed by Equation 4:
• P k flow is the peak-bandwidth of the new flow requiring admission to the Bounded Delay (BD) class
• P k accum is the sum of the peak bandwidths of all currently-accepted BD flows
• P k aMoc is the total bandwidth allocated for BD traffic. This describes the test made against new flows in checking that the sum of the peak-bandwidths of the new flow and currently-admitted flows does not exceed the allocated bandwidth.
• P k a noir oc may be chosen to be less than the link-rate in order to guarantee a certain level of bandwidth for the best-effort class.
• Equation (2) In the case described by Equation (2), and given known restrictions on N ip and / max the test described by Equation (4) may be the only necessary condition for CAC.
• 0C in addition to applying the peak- bandwidth test, it is also necessary for the CAC process to restrict N flows to some value, N f , ows a ,
• P k min a minimum peak-bandwidth
• each flow could also specify a maximum packet size, / ' max representing that for the ith flow.
• the delay bound may be approximated by a more general form of Equat ⁇ on(3)
• CAC also checks that there is sufficient spare buffer space available and applies the CAC rules described by Equation (7) .
• This type of CAC approach of implicitly limiting the number of flows is still applicable for single packet flows, and for networks using the P k m ⁇ n concept described earlier, Equation (7) is directly applicable for both single and multi-packet flows.
• the independent dual test CAC approach could enable more elaborate CAC methods to be considered than is possible with- the P k m ⁇ n approach.
• One example of such a method is where the CAC processor, which maintains running totals of the available buffer space and bandwidth, allocates either all, or some proportion of the spare bandwidth to each new single packet flow. In the case where a proportion is allocated, the amount may, for example, be made proportional to the ratio of packet-size to available buffer space. Using either of these approaches does not affect the form of any of the previous CAC tests, it simply results in particular instances having slightly different definitions for the CAC parameters and values.
• An additional benefit of this approach is that it also enables simple FIFO scheduling to be used, thus avoiding the complexity and significant processing overhead associated with schemes such as WFQ. Imposing a maximum packet size is particularly important when using FIFO scheduling because it also determines whether or not an acceptable level of bandwidth sharing occurs across all flows.
• the packet identifier (5, 1 6, 28, 38) having identified the packet as such may cause additional processing to take place to associate a peak rate bandwidth requirement with the bounded delay flow.
• This additional processing might, for example, involve the calculation of a bandwidth requirement on the basis of a delay requirement proposed by the user
• Further processing in the processing element (6, 1 7, 29, 39) may then determine whether the first buffer (4, 1 8, 30, 40) and associated output port (1 1 , 23, 32, 42) are able to meet the CAC conditions and as a result of this may decide as to whether or not the new flow may be admitted to the first buffer (4, 1 8, 30,
• One approach to dealing with this possibility is to admit flow packets refused by the CAC rules to the second buffer, to be dealt with on a best effort basis At the same time an appropriate 'connection refused' message may be relayed to the source Subsequent packets from the same flow may continue to carry 'connection request' signalling information to ensure the continuing possibility that the flow may eventually be admitted to the first buffer.
• worst case delay occurs at a network element when the buffer contains a backlogged packet from each admitted flow, so the worst case bound is equal to the sum of the largest packet size for each admitted flow, divided by the output link speed
• ⁇ ase a particular critical value
• T re i eas ⁇ could be made independent of packet size by using /' ⁇ multi _pkt m ; ⁇ ⁇ release r j k -" mr (Eq. 10)
• T release can be calculated as part of the CAC process which could then start a release counter, which on reaching a value equal to T re , ease , could decrease the value of / max accum by an amount equal to l ' mgx , and the value of P k acc ⁇ m by an amount equal to P k f!ow .
• P k _ together ow term is replaced by the appropriate bandwidth term. For instance, when using the P J , paragraphric, concept, as discussed earlier, all resource release computations using P f
• peak-rate CAC is used because it gives a deterministic bound on worst case buffer size, and hence it bounds worst case link delays to values determined by both the maximum packet size and the maximum number of simultaneous users. Consequently, packet loss can be prevented by using the worst case sized buffer, and link delays can be reduced as necessary simply by imposing an appropriate limit on maximum packet size.
• the use of buffer size less than the worst case will introduce a statistical probability of buffer overflow and consequent packet loss
• CAC rules may be defined to enforce the provision of a combination of zero packet loss, together with an absolute or "worst-case” delay bound, which will represent the greatest delay that could ever be seen under any conditions (other than some failure).
• any such transient fluctuations in peak-rate may be sufficiently limited so as not to breach the delay bound. Indeed, rather the reverse, and in many instances the conditions at particular nodes may be such that the worst-case bound is in statistical terms, overly pessimistic.
• CAC CAC rules can in some instances be relaxed somewhat, allowing more flows to be admitted, and that CAC might be applied in at least three ways, according to the location of a node:
• admission control is based on a combination of a "peak-rate sum test", and a "worst-case delay test".
• link-speed is such that the servicing time of an individual packet is sufficiently small compared with the required delay bound, statistical considerations may lead to the conclusion that only the "peak-rate sum test" need be applied.
• admission control may well be appropriate for admission control to be based on a combination of a "peak-rate sum test” and a “statistical delay bound test” .
• the latter is based on admitting flows up until the probability of exceeding a specified delay exceeds a certain (small) value.
• Futhermore in relation to the peak-bandwidth sum test, it may be considered practical to apply a degree of statistical multiplexing, particularly for nodes which are intended to carry large numbers of flows. This entails allowing the sum of the peak-rates for all admitted sessions to exceed the allocated peak-rate sum for the bounded-delay class traffic, based on the knowledge or likelihood that hosts will for some of the time, generate traffic at a level below their specified peak-rates.
• the best-effort mode may adapt from the given predetermined shares to use any spare bounded-delay mode bandwidth. Effectively, this means that the bounded-delay mode will have a guaranteed maximum amount of the total bandwidth and that the best effort mode will have a guaranteed minimum amount, which on average will be larger as determined by the bounded-delay modes utilisation statistics.
• buffer utilisation as will now be explained.
• any unused buffer space could be released as necessary for use by the best-effort mode.
• One implementation of this feature could be to use particular levels of buffer fill as metrics for reclaiming / releasing buffer space. Hence it ought to be possible to dimension the total buffer space such that a reasonable level of buffer utilisation is achieved but without having to accept a certain probability of packet loss.
• Another way to view this is that rather than the bounded-delay mode having a finite probability of packet loss, as is usually the case, it becomes instead the probability of the best-effort mode giving up some buffer space.
• an ingress node ( 1 ) based on the Figure 1 (a) node architecture and an egress node (25) based on the Figure 1 (c) architecture allow connection to physically separate respective connection - oriented and connectionless networks.
• Figure 2(b) shows a similar network example to that of Figure 2(a) but utilising a number of nodes ( 1 3) as shown in the Figure Kb) node architecture to effect connection between the Figure 1 (a) node architecture ingress node (1 ) and the Figure 1 (c) node architecture egress node (25) .
• Figure 2(c) depicts a network formed only from elements according to the Figure 1 (d) node architecture (35) .
• CAC might only be performed at a single node or subset of nodes rather than at every node. It might be practicable, for example, to apply CAC at only the ingress or egress nodes.
• An important feature of the invention is the ability to remove the need for delay negotiation when setting up a bounded delay mode flow. This is achieved by reducing network delays to the point where they become a negligible part of the overall end to end delay. As discussed several factors impact delay at an element but fundamentally they are determined by the speed of the output port link and the buffer fill which in turn depends on the maximum packet size and the CAC rules (i.e. the maximum allowed number of simultaneous users and the level of statistical multiplexing). Purely schematic examples of variations in relative delay as a function of these parameters are as indicated in Figures 3(a) to 3(c) .
• Figure 3(a) shows a schematic representation of the component parts of the overall end to end delay for a particular network configuration containing a number of network elements having a range of link speeds supporting only connection mode flows with, by way of example, a maximum packet size of 1 Kbyte and using a statistical multiplexing ratio of 2.5. It is clear for this particular example that the overall end to end delay is dominated by the network link delays. However, comparing Figure 3(a) with Figure 3(b) shows how these delays reduce when CAC prevents statistical multiplexing (i.e. instantaneous bandwidth requirements never exceed the amount allocated) and comparing yet further Figure 3(c) with Figure 3(b) shows the additional reductions achieved with a five-fold reduction in packet size. It is immediately clear from these comparisons that at some point the overall end to end delay makes a transition from being dominated by link delays to being dominated by host delays.
• a further factor which influences link delay is the proportion of the total bandwidth allocated for connection mode operation.
• the delay reductions achieved by decreasing the packet size are also representative of those that would result from reducing the bandwidth allocated to the bounded delay class from 1 00% to 25 % of the total output link bandwidth.
• RSVP Resource Reservation Set-Up Protocol
• the RSVP Path and Resv signalling messages would operate as normal, using Tspec and Rspec fields to identify the BD class flow and the peak rate requirements, P kate ow , for this flow, these signalling messages being passed hop by hop towards the receiver and then hop by hop back to the sender.
• each hop back to the sender would compare the sum of the requested bandwidth, P k
• RSVP reservations for IP version 4 are identified by source address/destination address/port triples which act as the flow classifier.
• IP version 6 has the notion of a single field called the flow-id.
• Other classifiers relating to groups of senders or receivers (address prefixes), traffic type (port number/transport protocol) or other identifiers such as 'tags' are well known and could be used to provide an aggregated classifier.
• Multiple flows can be easily aggregated in the CAC and signalling processes of our invention by simply summing up the total peak bandwidth of the aggregated flows and storing this as an aggregated classifier, and also sending this aggregated total peak bandwidth, P ⁇ ⁇
• the mode identification process may be accomplished in many ways and in this case could be part of the general 'out of band' RSVP signalling and classification process. If CAC fails then the classifier state is stored but no resources are allocated. This classifier state is used to divert incoming BD packets from this flow into the BE buffer at this node to protect the other accepted BD flows in this node. Downstream nodes in which this flow passed CAC can still put this packet back into the BD class. RSVP messaging could ensure that the customer is informed of an incomplete reservation or other protocols or diagnostic tools could provide this feedback. The customer could at that point decide to cease the flow due to the BD class request failing, which would result in the classifier state in all nodes timing out.
• Policing could either be done at all nodes, or only at the ingress and egress nodes of this delay-bounded network when sufficient capacity has been provided for the traffic matrix and the allocated amount of BD class capacity.
• the policing would act to ensure that the BD class flows kept within the declared maximum packet length and peak rate, and also to implement functions such as packet counting, reservation policy and billing.
• Routing changes would result in the packets for this flow using different nodes and the refresh mechanism would be used to invoke CAC in the new nodes whilst the soft state would ensure that the old nodes returned the previously reserved bandwidth.
• the flow packets (45) are defined to have a header portion with a number of fields and a data carrying payload portion (47) .
• the first two header fields (48, 49) carry the source and destination addresses, these being used to route packets across the network in the usual manner. Other well known fields may also be included.
• the mode identifier field (50) is used to distinguish between best effort and bounded delay mode packets.
• the packet identifier field (51 ) is used to indicate packets of particular functionality.
• the additional processing information field or fields (52) is used to indicate to network elements information such as the effective peak data rate that the host has selected for its bounded delay mode flow. All of these header fields (46) may be specified by the application.
• the payload portion of the packet (47) contains the information to be transported across the network to the destination host.
• the mode identifier may carry out the mode identification by means of reading the Mode Identifier field (50) in the header of the received packets (45) .
• the leading packet of a new flow may be used to initiate the Connection Admission Control (CAC) procedure.
• CAC Connection Admission Control
• the packet identifier having identified the leading packet of a new flow from the Packet Identifier field (51 ) may cause the proposed Bandwidth requirement to read from a leading packet header Bandwidth field (52), whereupon the peak rate CAC decision may be made. Other such parameters could be acquired in a similar fashion.
• bounded delay class flows may be established without a separate signalling phase, using 'signalling on the fly' .
• a particular packet identifier could also be defined for the purpose of sending intermediate packets within a flow for additional processing thus enabling applications to dynamically adjust their end-to-end delay requirements as necessary.
• All packets could include the mode identifier for the BD class. However, the mode identifier could be taken from the first 'signalling packet' and stored
• the mode could be statically configured for specific classifiers and in particular for aggregated classifiers to automatically provide BD class for VPN configurations and for specific traffic types such as Internet telephony.
• the first packet in the flow includes the associated peak rate bandwidth for the flow, P k Mow and any other parameters needed for CAC.
• a new flow could be admitted in every node on the end to end path using the same CAC rules as described above. If the flow fails CAC then failure classifier state is installed as before, but now a new 'CAC Failed' message is returned to the sender. This could be a new ICMP code or some other means.
• a 'CAC failed' flag is set in the first packet header so that downstream nodes and the receiver are also informed of the failure and downstream nodes do not also send back CAC failure messages which could otherwise swamp the sender.
• the remaining packets in the flow which match the classifier and have the BD mode set, are admitted into the BD buffer, but use the BE buffer in nodes which failed the flow during CAC. Refreshing of the signalling packet, indicating the P k flow value and other CAC parameters, is used to cope with routing changes, time-out of classifier state and later admittance to resources on initial failure during
• the preceding example could be implemented by specifying mode and packet identifier codes for use in the TOS header field.
• the required bandwidth could be either pre-configured or placed in the IP v 4 options field.
• the mode and packet identifier codes could be placed in the flow ID field and the bandwidth requirement could be placed in the hop-by-hop extension (options) header.
• appropriate ICMP messages could be generated, for example flow admission refused etc. It is very likely that some form of traffic processing capability will be required within both the originating and receiving host terminals depending on both type of application being used and the number of simultaneous flows being generated. Therefore, it is expected that in general hosts might also make use of the dual buffering nature of the network elements.
• An example of an arrangement of such a host architecture is as indicated in Figure 5.
• the host may, for example, be implemented using well known Unix work station or PC technology equipped with a suitable network interface.
• Application A produces one or more flows which are to be sent in the bounded delay class.
• a bounded delay buffer 55
• buffer allocation and packet size controls 56, 57
• the scheduling element 58
• the scheduling element is controlled by an appropriate scheduling algorithm (59) .
• a best-effort buffer (60) is provided purely to accommodate the fact that the scheduling algorithm(s) (59) will have been designed to give highest priority to packets stored in the bounded delay buffer (55) .
• Application B (61 ) could also have similar control over buffer allocation (not shown) and packet size (62) as shown for Application A (54) .
• Traffic shaping is an important aspect for flows that are to be guaranteed a delay bound, and primarily shaping enables the inter-arrival times between packets generated by an application to be modified prior to being input into the network. Hence, the burstiness of a packet flow can be controlled by shaping, the outcome of this being that a lower peak bandwidth is needed from the network than may otherwise have been the case. Effectively, traffic shaping is simply a trade-off between network bandwidth requirement and host delay, with the latter decreasing as the former increases. To date, traffic shaping has in general been achieved using either token bucket or leaky bucket shapers, with combinations of both being used in some instances.
• token bucket shaping cannot prevent packet bursts from being input into the network, it does bound the size of the bursts. Therefore, token bucket shaping does not eliminate the possibility of the flow rate instantaneously exceeding the specified peak-rate when the instantaneous value is calculated over any time period that is measured from the beginning of any packet sequence (which could be a single packet) and an adjacent one, as illustrated in Figure 7 Based on this definition, the instantaneous peak-rate for a flow is given by
• a good example of a scheduling algorithm that satisfies this is one whereby the decision to input a packet into the network is based on the size of the previous packet and the time at which it was sent. Therefore, the scheduling element does not admit a new packet into the network until the time lapse since admitting the previous packet, ⁇ 'j , satisfies the relationship
• / packe represents packet size in bits and represents the specified peak rate for the flow in bits/second.
• Superscript n represents the packet to be sent and superscript n - 1 the previously sent packet.
• the only information required by the network is the flow's required peak-rate, and this can be provided by 'on the fly' signalling Consequently, the application could now continually monitor the host performance during the life-time of a flow and adjust as necessary buffer allocation, packet size, packet discard or peak-rate as often as required and without any need to involve the network.
• bounded-delay class For example, for the bounded-delay class, if the approach is adopted whereby new flows signal their maximum packet sizes as part of the CAC process, separate classes could be defined for two or more specific ranges of packet size. By separately controlling allocations of bandwidth or buffer resource for each of these, some control is obtained over the distribution of packet-sizes for admitted flows, which may lead to an improvement in the bandwidth-utilisation of the overall bounded-delay class. One potential benefit of this might be in improving the probability of admission for flows with small packet sizes. A further possibility is that sub-classes could be established to provide different values of bounded-delay. This could be achieved by separating packets into multiple virtual queues within the main bounded-delay queue.
• RED Random - early detection
• IETF IntServ IETF IntServ
• CBQ class-based queuing

Priority Applications (6)

Applications Claiming Priority (4)

Publications (1)

Family

ID=26147594

Family Applications (1)

Country Status (7)

Cited By (15)

Families Citing this family (72)

Citations (2)

Family Cites Families (5)

• 1998
  • 1998-09-09 EP EP98942868A patent/EP1013049B1/en not_active Expired - Lifetime
  • 1998-09-09 AU AU90843/98A patent/AU9084398A/en not_active Abandoned
  • 1998-09-09 WO PCT/GB1998/002727 patent/WO1999013624A1/en active IP Right Grant
  • 1998-09-09 JP JP2000511292A patent/JP4201978B2/ja not_active Expired - Lifetime
  • 1998-09-09 CA CA002302218A patent/CA2302218C/en not_active Expired - Lifetime
  • 1998-09-09 US US09/180,102 patent/US6538989B1/en not_active Expired - Lifetime
  • 1998-09-09 DE DE69818846T patent/DE69818846T2/de not_active Expired - Lifetime

Patent Citations (2)

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events