JPH10190691A - Traffic forming device and routine - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a rate forming device for a flow-based standby routine mechanism by preparing a queuing mechanism which exclusively sorts the peak packet sending rates and the flows from each other and organizes an undecided packet of active flows included in a queue and also a scheduling mechanism for the transmission. SOLUTION: The flows are sorted into the category A flows which have the non-zero minimum packet sending rates having the contact designated relative minimum rate tolerance and the category B flows which have the traffic contracts that allow the zero packet sending rates respectively. A scheduling mechanism includes a non-work maintenance calendar queue having a relatively high priority, a work maintenance queue having a relatively low priority and other non-work maintenance queues having the undecided priorities with respect to the said two queues. Then the sending process strategy of every queue is given to the scheduling mechanism.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a packet-switched communication network, and more particularly, to creating a flow of time-multiplexed packets at queuing points in such networks or network elements to conform to specified traffic descriptors. Related to traffic shaping.

[0002]

2. Description of the Related Art Traffic Contract / Definition Most applications currently running on packet-switched communication networks can operate satisfactorily with whatever bandwidth accidentally comes from the network. This is because those applications require "elastic" bandwidth. This type of service that supports these applications is known as the "best effort" service in the Internet world or "Available Bit Rate" (ABR) in the broadband ISDN / ATM world.

However, there is an increasing demand for network services that provide bounded jitter, or bounded packet delay variation (generally referred to as cell delay variation in the context of ATM). For example, such services are needed in real-time applications such as circuit emulation and video. It is not clear whether or not the Internet will respond to such demands and how it will respond, but in the field of broadband ISDN / ATM it has responded by introducing the concept of a user-network negotiated traffic contract.

[0004] As is known, user-network ATM contracts are defined by traffic descriptors, including traffic parameters, tolerances, and quality of service requirements. Conformance definitions (conformanc) for each of the relevant traffic parameters
e definition) is specified. Thus, ATM services can use these traffic parameters and their corresponding conformance specifications to support different combinations of quality of service (QoS) targets and multiplexing schemes.

[0005] Partially overlapping ATM traffic classes have been defined by the International Telecommunication Union (ITU-T) communication standards department and the ATM Forum. In some examples, these two groups have different names for traffic classes that have essentially the same attributes, and the following name translation tables are equivalent to the currently used equivalent names. Is shown.

[0006]

[0007] An ATM service contract for a virtual circuit (VC) connection or a virtual path (VP) connection can include a plurality of parameters describing the service rate of the connection. This includes peak cell rate (PCR), sustainable cell rate (SCR), intrinsic burst tolerance (IBT), and minimum cell rate (MCR). Not all of these parameters are relevant to individual connections or individual service classes, but if they are implicit or explicitly specified elements of a service contract, they must be respected . Although VC connections are the focus of the discussion below, it will be appreciated that VP connections can be so specified. A data transfer unit for an ATM connection is usually called a "cell". But,
In this specification, the term "packet" is sometimes used to represent a unit of data transfer. This is because the term "packet" is a more general term for some of the broader concepts of the subject innovation.

[0008] The Generic Cell Rate Algorithm (GCRA) specified in ITU-T Recommendation 1.371 is well suited for testing the flow of packets or cells for conformance with traffic descriptors. In order to perform such a test, the GCRA requires an emission
It requires specifications for the interval (ie, the reciprocal of the flow rate) and the tolerance τ. In practice, this tolerance will depend on various factors such as connection, connection setup parameters, or type of service. As will be seen, GCRA can be employed as a Boolean function, and for fixed-size packet or cell flows on a connection, GCRA (output interval, tolerance) is false if the flow follows a peak rate. Yes, true if the flow follows the minimum rate. For example, if GCRA (1 / PC
If R, τ _PCR ) is false, the cell source follows PCR. Similarly, if GCRA (1 / MCR, τ _MCR ) is false, the connection or flow follows the MCR.
As will be seen later, "output interval" is the reciprocal of "cell rate".

A DBR traffic contract is appropriate for a source that establishes a connection in anticipation that a certain amount of bandwidth is continuously available for the connection while the connection is being made. Therefore, the bandwidth that the network entrusts to the DBR connection is
It is characterized by an R value. Furthermore, the flow of the cells or packets in such connection, if it _{GCRA (1 / PCR, τ PCR} ) if adapted to,
Comply with traffic contracts. On the other hand, SB
The R traffic contract is suitable for applications with known traffic characteristics that take into account the signaled selection of SCR and τ _IBT and PCR and τ _pcT . SBR or rt-SBR flow, if the flow is _{GCRA (1 / PCR, τ PCR} ) as well, GC
If conforming to RA (1 / SCR, τ _IBT ), the traffic contract is complied with.

[0010] As mentioned above, ABR traffic contracts are suitable for applications that can tolerate dynamic changes in information transfer rates as a result of using unreserved bandwidth. PCR and MCR
Is specified by the source establishing such a connection, and these parameters are subject to negotiation with the network. Thus, the available bandwidth for an ABR connection is the sum of the MCR (which may be zero) and the variable cell rate caused by sharing unreserved bandwidth between ABR connections according to a defined allocation policy ( That is, the bandwidth a source receives above its designated MCR depends not only on the negotiated PCR but also on the network policy). Feedback from the network allows the source application to dynamically adjust the rate at which cells or packets are sent for the ABR connection. The ABR flow is GCRA (1 / MCR, τ
If compatible with the _MCR), always in compliance with the traffic contract, if not conform to _{GCRA (1 / PCR, τ PCR} ), it is always non-compliance. MCR and PC
The suitability in the region between R depends on ABR feedback and is thus determined dynamically.

[0011] The UBR traffic contract is AB
Similar to the R contract, but the UBR contract is not adaptable to the MCR specification and has no dynamic conformance definition. Therefore, if the UBR flow is G
_{CRA (1 / PCR, τ PCR} ) if adapted to, in compliance with traffic contract.

B. Traffic shaping ITU-T Recommendation 1.371 addresses the possibility of reshaping traffic at network elements in order to adapt traffic to traffic descriptors with respect to:

"Traffic shaping is used to improve network efficiency while meeting QoS targets, or to ensure adaptation at subsequent interfaces.
A mechanism that changes the traffic characteristics of the cell stream on the CC or VPC to achieve the required changes in these traffic characteristics. Traffic shaping must maintain the integrity of the cell sequence on the ATM connection. Shaping alters the characteristics of the cell flow, thereby increasing the average transfer delay of the cell.

Examples of traffic shaping are peak cell rate reduction, burst length limitation, CDV reduction by appropriate cell time intervals, and queue service schemes.

Whether or not to perform traffic shaping and where to perform it depends on the choice of the network operator. For example, a network operator may select a suitable UPC /
You can choose to perform traffic shaping with the NPC function.

Performing traffic shaping on separate or collective cell flows is an option for the operator.

Therefore, traffic shaping can be performed for any ATM connection.

Options available to the network operator / service provider include:

A. No shaping-Determine the size or dimension of the network so as to allow any flow of conforming cells at the network entrance while ensuring compatibility at the network exit without using any shaping functions.

B. Shape and operate the network such that any flow of conforming cells at the ingress is carried by the network or network segment while meeting the QoS targets, and shape the output of traffic to meet the conformance test at the egress. Do.

Shape traffic at the entrance of the network or network segment, satisfy QoS targets and subsequent conformance tests at the exit of the network or network segment, and allocate resources according to the traffic characteristics obtained by the shaping.

Also, use traffic shaping in the customer's equipment or at the source to ensure that cells generated by the source or at the UNI comply with the negotiated traffic contract for the ATC used. (See Section 5.5). "IT
UT Recommendation 1.371, Chapter 6.2.5.

C. Scheduling / Existing Tools and Techniques for Real-Time and Non-Real-Time Connections As is known, bandwidth is "fair" split between applications employing "best effort" Internet services or ABR ATM services If not, various undesirable phenomena will occur.

At present, most ATM switches use FIFOs.
Queuing has been implemented. However, FIFO queuing behaves abnormally when used for ABR traffic. Also, FIFOs cannot protect a user who is acting correctly from a user who is acting incorrectly (does not provide isolation). Because of these drawbacks, non-FIFO queuing mechanisms such as weighted fair queuing or schemes that approximate fair queuing such as round robin are often proposed.

Services of the type that require inelastic bandwidth sometimes require that data be transmitted over the network with constrained jitter (ie, limited cell or packet delay variation). Request. As is known, weighted fair queuing can be used to provide constrained jitter for real-time streams. In addition, tested delay limits can be provided for systems that use the closely related mechanisms of virtual clocks and self-timed fair queuing.

Therefore, the elasticity (best effort / A
It is known that both BR) and inelastic (or real-time) services benefit from using fair queuing and related algorithms.

1. Weighted fair queuing and virtual clocks Fair queuing and related algorithms (eg, frame-based fair queuing, defective round robin, etc.) may be packets or other data transfer units (eg, ATM cells may be packets in the discussion herein). ). In the case of ATM, these sequences are recognized by either VCI or VPI, and suitable Internet protocols for this recognition are based on <IP address, protocol, port> triples (IPv4) or flow identifiers (IPv6). . In both self-timed weighted fair queuing and virtual clocks, packets are ordered (sorted) by time stamp (round-robin and similar schemes approximate packet ordering by time stamp). This timestamp represents the virtual end time (or similarly the virtual start time of the packet), gets the start time, and multiplies the packet length by a weight that represents the bandwidth allocation for a particular packet sequence. Calculated by adding the offset obtained by

More specifically, in the case of a virtual clock, the virtual end time is calculated as follows.

VT (f, 0) = 0 VT (f, j + 1) = max ｛arrival (f, j + 1), VT (f, j)｝ + length (f, j + 1) / rate (f) (1) where VT (f, j) is the packet j of the flow (virtual line) f
Is the virtual end time associated with, and the arrival (f, j) is the flow f
Is the arrival time of packet j, and the length (f, j) is
The length of packet j in f.

The self-timed weighted fair queuing specifies a virtual end time according to the following equation.

VT (f, 0) = 0 VT (f, j + 1) = max ｛system virtual time, VT (f, j)｝ + length (f, j + 1) * weight (f) (2 However, the system virtual time is a virtual time relating to the packet provided (output) at the time when the packet (f, j + 1) arrives.

In the case of ATM, since the cell has a fixed size (for example, 53 bytes in length), the packet length is constant. Therefore, the rightmost terms of Equations (1) and (2) are
It becomes a constant for each flow. For virtual clocks, the equation can be simplified as follows:

VT (f, j + 1) = max {arrival (f, j + 1), VT (f, j)} + constant (f) (3)

For self-timed weighted fair queuing, the equation can be simplified as follows:

VT (f, j + 1) = max ｛system virtual time, VT (f, j)｝ + constant (f) (4)

That is, the ATM queuing point that performs the virtual clock or the self-timed weighted fair queuing performs the following steps.

1) Current virtual time for VC (a)
And the maximum value between the cell arrival time i) and any one of the system virtual times ii) (b) is calculated. 2) In the result of the above step 1, for each VC indicating the occupancy of the bandwidth VC 3) Provide (send) the cells in the order in which the values of the virtual timestamps specified by steps 1 and 2 increase.

2. Priority To give one traffic class higher priority than the other means that if the higher priority traffic class has cells ready for transmission, these cells will always be It is transmitted in priority to cells of a traffic class with a lower priority.

The priority mechanism is preemptive.
e) can be non-preemptive. This term comes from the operating system literature. The non-preemptive priority mechanism gives priority to objects (processes in the operating system world, VCs in the ATM world) at scheduling time, and the object gives this priority until it is provided. maintain. On the other hand, a preemptive priority mechanism can change the priority of an object while waiting for the object to be provided. For example, in a preemptive system,
It can be said that "schedule this VC with priority 3", but if it is not provided within 200 microseconds, its priority is raised to 2.

3. Work Conservation and Non-Work Conservation Queuing The term "work conserving" is sometimes used to describe any queuing system in which work is neither created nor destroyed. Consistent with this term, a switch that always transmits a cell to an output link when a standby cell is provided is a "work maintenance switch". Switches employing pure FIFO, weighted fair queuing or virtual clock scheduling algorithms are all work-maintaining. In contrast, a non-working switch can choose not to send a cell even if the cell is waiting for a transfer. As will be seen, a way to do this is to program the switch to wait until the current time is equal to or greater than the time stamp associated with that particular cell before transmitting it.

The maintenance switch attempts to make full use of the transmission link, but does not necessarily eliminate bursts,
Do not block. In contrast, non-work-maintaining switches can strategically delay cells to reshape traffic to meet more stringent conformance tests (ie, GCRA with smaller τ). In addition, non-work-maintaining switches, where only a specified amount of buffering is allocated to a given connection, can be provided with a policing function (by discarding or tagging cells that overflow the allocated buffer space). In the ITU, this is called UPC / NPC). An example of a non-work keeping queuing system is a stalled virtual clock that adapts a virtual clock algorithm where virtual time is not allowed to be too fast (real or stuck in work) than real time.

4. Calendar Queue A calendar queue is a chronological list of actions, each action being taken from the queue and executed when the actual time is equal to or greater than the time associated with the action. A calendar cue with a limited time interval can be represented as a linear array known as a "time-wheel" or "time-line". The time wheel assigns events for the bucket relative to the pointer,
The bucket index is calculated using the wheel size as an arithmetic modulo. These data structures are well known in the literature as queuing mechanisms.
In the time wheel, the absolute time is represented as an offset from the current time (actual time), and each element in the array is
A bucket that contains one or more actions (typically in the form of a linked list) to be performed at a specified time on the bucket to which they belong. Any of such time wheel buckets can be empty, i.e., have no events associated with them.

For each time wheel, there are two important times, _earliest and t _latest , which correspond to the head and tail pointers for the active entry in the array, and the _earliest is the next entry to provide ( (Eg, packet or cell) time, and t _latest is the time associated with the latest (farthest in time) bucket that contains the scheduled event. The difference between _earliest and t _{late st} cannot exceed one time wheel length b minus one. This is due to the fact that time is maintaining modulo b, and that there is no offset (packet length multiplied by either rate or weight in virtual clock or weighted fair queuing, respectively) greater than b-1. Is ensured by ensuring that For an ATM link operating at OC-3 speed (149.76 Mbps-SONET payload rate), there are about 353208 cells per second on the link. Thus, a 64 Kbps (voice phone speed) flow (approximately 174 cells per second if AAL type 1 is used) is the slowest connection to be supported and the fastest supported for the slowest speed. Speed ratio is 20
29, rounded up to 2 ¹¹ . This ratio is the maximum offset that is added while calculating the actual time. Thus, a timewheel of length 2030 (2048 when considering rounding to the power of two) encodes the actual time associated with the line at speeds between 64 Kbps and the speed of the full OC-3 link. Is enough to make

The length of the time wheel array can be reduced by allowing the array elements to include more than one time offset. For example, if the above time wheel was reduced from 2048 elements to 256 elements, each bucket would have eight time offsets mapped into it. Actions within a single bucket spanning multiple offsets may be performed out of order, but between buckets the actions will be in order. This can reduce the amount of memory that needs to be allocated to such a time wheel at the expense of reducing the accuracy of the ordering of the actions in the calendar queue.

D. Traffic shaping for time multiplexed flows on multiple output channels Any traffic shaping necessary to adapt the time multiplexed packet or cell flows to their traffic contracts is the flow for the different output channels. It is preferably performed after the switching or routing operations required to separate the two from each other are completed. Thereby, the throughput efficiency of the multiplexer can be optimized.

However, previous output standby ATM switches generally employed FIFO (first in, first out) output buffers. These buffers cannot participate in controlling the reshaping of any of the flows passing through them. Instead, the per-VC time mutiplexed flows output from these buffers are essentially a time multiplex of the input flows loaded into them. Of course, these output flows are delayed in time with respect to the input flows due to the inherent latency of the buffer. Also, if a scheduling conflict occurs between the data transfer restrictions of different flows, the cell delay variation (CDV) of one or more of these output flows
Will increase. This is because these interferences cause so-called "transmission collisions".

As will be seen, the increase in CVD is particularly troublesome for traffic with relatively tight tolerances, such as DBR traffic. Thus, if each hop between the source and the destination includes a simple FIFO output queue of the above type, this CDV-sensitive traffic must be passed to ensure conformance within its specified tolerance. It will be necessary to limit the number of hops that can be made.

Thus, there is a need for more efficient and more effective traffic shaping mechanisms and processes for ATM switches or other routers that send traffic from multiple inputs to multiple outputs for the delivery of time multiplexed outputs. Is clear.

[0049]

SUMMARY OF THE INVENTION The present invention provides for rate shaping in a per-flow wait routing mechanism for available bit rate services.

[0050]

SUMMARY OF THE INVENTION An aspect of the present invention includes a contract in a packet-switched communication system, wherein a time multiplexed flow is adapted to substantially match the individual network traffic contract of each flow. A traffic shaping device for continuously transmitting packets, wherein the contract specifies a respective peak packet transmission rate and an associated peak rate tolerance of some of the flows and has mutually exclusive categories, ie, each has a mutually exclusive category. Classifying each of said flows into one of a Category A flow having a non-zero minimum packet delivery rate with an associated minimum rate tolerance specified in the contract and a Category B flow having a traffic contract allowing a zero packet delivery rate; Said The traffic shaping device includes a queuing mechanism for organizing pending packets of an active flow including the number of flows in each queue according to the oldest pending packet at the head of the queue, wherein the queuing mechanism And a scheduling mechanism for scheduling the packets of the active flow for transmission, the scheduling mechanism comprising a relatively high priority non-work maintenance calendar queue and at least one relatively low priority calendar queue. A work maintenance queue and another non-work maintenance queue whose priority is not determined with respect to the high-priority queue and the low-priority queue, wherein the scheduling mechanism comprises: (i) the high-priority queue; In the non-work maintenance calendar queue, A contract-specified minimum transmission rate for the first flow when packets of the first flow are being transmitted at a rate that exceeds a tolerance for a specified minimum transmission rate for one given flow. To send at a rate exceeding
(ii) in the other non-work maintenance calendar queue, when a packet of the first flow is transmitted at a rate exceeding a tolerance of a designated peak transmission rate of the first flow, the first flow (Iii) in the work maintenance queue, when the packets of the first flow are being transmitted according to the traffic contract of the first flow, so as to transmit at substantially the same rate as the specified peak transmission rate for Schedules any given Category A packet for delivery according to the selected continuous service strategy, the scheduling mechanism comprising: (iv) in the other non-work keeping calendar queue the second given flow The rate at which packets exceed the tolerance for the specified peak delivery rate for the second flow. (V) in the work maintenance queue, the packet of the second flow is transmitted at the same rate as the specified peak transmission rate for the second flow when the second flow is transmitted. Schedule any given Category B packet to be sent according to the selected strategy as it is being sent according to the two-flow traffic contract.

A second aspect of the present invention, in a packet switched communication system, includes a contract and continuously sends out packets of a time multiplexed flow to substantially match individual network traffic contracts of each flow. A traffic shaping device, wherein the contract entrusts each of the allowed peak packet transmission rates to some of the flows,
Classifying each of the flows into one of a mutually exclusive category, ie, a Category A flow to which a respective non-zero allowable minimum packet transmission rate is committed and a Category B flow to receive a traffic contract allowing a zero packet transmission rate. And said traffic shaping device includes a queuing mechanism for organizing packets of active flows including said several flows in respective queues according to the oldest pending packet at the head of the queue order; A scheduling mechanism coupled to the scheduling mechanism for scheduling the packets of the active flow for transmission, the scheduling mechanism including a relatively high priority non-work maintenance calendar queue, and at least one relatively superior calendar queue. Low-priority work maintenance queues, other non-work maintenance queues whose priorities have not been determined relative to the high-priority queues and low-priority queues, and the high-priority queues and low-priority queues A control mechanism that selectively couples the scheduled packets in the high priority queue and the low priority queue in combination with the monitoring transmission from
The scheduling mechanism comprises: (i) scheduling packets for each flow in the high priority calendar queue to be sent at a rate that exceeds a minimum rate commitment for each flow; and (ii) a predetermined work maintenance queue. Preparing packets of any category A flows by placing them in the work queue to be sent according to the outgoing strategy, thus creating a race condition for the packets of these category A flows, the control mechanism comprising: When arriving at one of the high-priority queue and the maintenance queue, this packet is taken out of the other packets of the queue by removing the packet of the delayed category A flow. Resolve the race condition and The forwarding mechanism sends all Category A and Category B flow packets that are sent at a rate that exceeds the allowable peak sending rate entrusted to each flow at a rate approximately the same as the peak sending rate entrusted to each flow. Scheduling in the other non-work-keeping calendar queue, the scheduling mechanism arbitrates packets of any category B flow that are being sent at a rate compatible with contracting for the respective flow. Queued for delivery according to the outgoing strategy.

An aspect of claim 3 of the present invention is to link fixed bit length packets of a plurality of time multiplexed flows to a link of known bandwidth so as to substantially conform to the individual traffic contract for each flow. In a continuously transmitting traffic shaping process, at least some of the flows are sent to respective available bit rate traffic contracts specifying respective peak rates and minimum packet transmission rates greater than zero for the some flows. A shaping routine, comprising the step of organizing packets of active flows including the number of flows in each queue according to the oldest pending packet at the head of the queue order; Bit rate traffic co Dynamically determining the available bandwidth on the link as a percentage of the known bandwidth for flows having tracts, and dynamically determining the number of active flows having available bit rate traffic contracts. Including steps
For each head of a queue packet of each flow receiving an available bit rate traffic contract, a first theoretical delivery time based on a contract specified minimum delivery rate for the flow, and A second theoretical delivery time based on an active available bit rate per-flow allocation of bandwidth available on the link, and a third theoretical delivery time based on a contract-specified peak delivery rate for said flow; And discarding the earliest of the second and third theoretical sending times, and maintaining the non-operation so as to send at the earliest of the remaining theoretical sending times. Scheduling the packets in a queue.

[0053]

DETAILED DESCRIPTION OF THE INVENTION Exemplary Environment Referring to the drawings, and in particular to FIG. 1, the input and output ports of the ATM switch 21 are typically coupled to one or more physical layers via a Utopia2 interface,
It is connected to the switch control processor module 22 via a second suitable interface. As a result, the switch 21 exchanges data and control cells with any connected physical layer, and exchanges control cells with the control processor module 22. Consistent with this standard practice, communication channels are one-way, so two-way communication requires a set of channels.

The switch 21 includes a switching fabric 24, a fabric control module 25, and a reservation ring 26 for switching data and control cells from an input queue to a per-VC output queue. Cells in these queues are stored in data memory 27.
Stored in the data path inside, these input and output queues are managed by the queue control module 28.
Typically, the data memory 27 has a maximum of about 12000
This is the capacity that can store cells. Data and control cell flow connection records are stored in the control path in control memory 29 along with certain types of control cells captured by rate based engine / traffic multiplexer 31 for routing to control processor module 22. The control RAM 29 has a maximum of about 82
Preferably, a connection record of 00 and a cell record of 64K can be stored. The interaction between the control processor module 22 and the switch 21 is beyond the scope of the present invention and will not be described herein. But A
If you are familiar with the design of TM switches, the control processor is mainly responsible for establishing and terminating connections, OAM
(Operation and Maintenance) function.

The data path of the switch 21 is synchronously clocked at a predetermined rate, for example, 40 MHz.
However, in line with traditional synchronous pipeline design practices, the phase of this clock signal must be set so that the data has the appropriate time to settle before being transferred from one pipeline stage to the next. Delayed by different amounts (by means not shown) at different points along the data path.

According to standard practice, the source that wants to communicate with the destination is the switch 21
Initiates negotiations with an existing ATM network by sending a SETUP message to the network. This message identifies the destination and specifies, explicitly or implicitly, all of the relevant traffic parameters for the requested connection. If the network is ready to commit to the traffic contract defined by these traffic parameters (or a variant of the parameters that the source wants to accept), the network sends a SETUP message to the destination. Send to Nation. Then, if the destination is ready to receive message traffic from the source according to the conditions of the traffic contract, the destination returns a CONNECT message to the source.
This CONNECT message confirms that a connection has been established on the specified virtual circuit (VC) in the specified virtual path (VP) for the cell flow conforming to the traffic contract. ITU
-T Recommendation 0.2391 and ATM Forum UNI
See the 4.0 specification. By supplying these signaling protocols without generating them, a "permanent" virtual connection can be established.

After the connection is established, the data cell starts to flow. As shown in FIG. 2, the shape of the cell changes as it passes through switch 21 for the actions performed by switch 21. Cells may be duplicated for multicasting, but the discussion below is limited to unicast operation to avoid unnecessary complications.

As shown in FIG. 2, each upstream cell 41 received by the switch 21 has a header including a VP index and a VC index. These indices are combined to define a unique address for one hop of the connection. A connection can be made up of multiple hops, so the VP and VC indexes for the next hoop are written in the header of the cell as it passes through the switch 21, as shown at 42 in FIG.

The switch 21 calculates the V of the upstream cell (41 in FIG. 2) to calculate the address in the control RAM 29 where the connection record for the related flow exists.
Adopt P and VC indexes. Typically, this connection record is a bit vector for recognizing the output port from which the flow exits the switch 21 (ie, the destination at the switch level), on a granular priority scale. And a line index (“Circuit Index”) that uniquely identifies a flow in the switch 21.
x "). These connection parameters are written in the cell header, as shown at 43 in FIG. The cell is then written to data RAM memory 29, and a pointer to the cell is linked to an appropriate one of the plurality of FIFO input queues. The selection of the queue is performed based on the priority of the related flow.

The relative priority of the heads of the cue cells in the input queue is examined during each cell time, and the head of the cue cell with the highest priority is determined for the arbitration session during the next arbitration session. Selected for. In addition, the priority of the lower priority head of the cue cell (ie, any unselected head of the cue cell) is increased by a fixed amount (by means not shown), thereby increasing the arbitration of the next arbitration session. To increase the likelihood that a cell will be selected. Thus, higher priority input queues have higher throughput per unit time than lower priority queues, but lower priority queues are limited in delay because the priority of the queue cell heads increases over time. You.

Each arbitration cycle requires one cell time of switch 21, so the routing information of the cell selected for arbitration is:
The payload of the cell or cells that have won the arbitration is provided to the reservation ring 26 one cell time before being released to the switching fabric 24. That is, as indicated at 44 in FIG. 2, the cells received by the reservation ring 26 and the switch fabric 24 comprise the header of the cell for the next arbitration cycle (ie, the “current cell”), The body or payload of the cell that has been successfully arbitrated in the previous arbitration cycle (ie, the “previous cell”) follows. Therefore,
When the cell bodies arrive at the fabric 24, the fabric is controlled by the fabric controller 2 to send these cells to their respective output port destinations.
5 is already configured.

In the illustrated embodiment, cells are broken into 4-bit wide nibbles for arbitration and routing. After this (except for the "idle cells" that would be provided to test the switching process), the cells are reassembled and (a) time-scheduled transfer to the appropriate output port of switch 21 And / or (b) queued in the data path, control path, or both, for transfer to the control processor module 22. Switch 21
The time-scheduled transfer of cells to the output ports of a is central to the present invention and is further described below. On the other hand, cell disassembly and reassembly, test process, and RM (Resource Management) of control processor 22
And interaction with OAM (Operation and Maintenance) cells is an incidental topic that does not need to be considered in depth.

Referring to FIG. 3, it can be seen that switch 21 is fanning out on the output side, ie, the transmission side, of switching fabric 24. Thus, although only one output channel of switch 21 is shown, it will be appreciated that this channel is representative of the other channels.

As shown, fill cell module 5 to accept cell bodies from switching fabric 24 and their associated line indexes
1 is preferably provided. The "effective cell time" at the output of the switching fabric 24 is determined by the ratio of the normal cell time to the "k" acceleration factor. Thus, for example, if the normal cell time is 113
In the case of clock cycles / cell, the effective cell time at the output of the switching fabric 24 is 56.5 cycles / cell if k = 2.

When a valid cell is received, fill cell module 51 typically outputs a cell from linked, numbered free list 52 for a data structure for writing the cell to data memory 27. Use the configuration. For this reason, it is preferable that the fill cell module 51 includes a footch state machine 53 for reading a cell structure from the top of the free list 52 on demand. Thereby, the fill cell module 5
1 can insert a line index for the cell and a pointer to the location of the cell in the data memory 27 into an "arrival" message sent to inform the cell flow control unit 55 of the arrival of the cell. From the connection record in the control memory 29, the flow control unit 55
Or the flow can be ascertained. Then, the cell flow control unit 55 can check the traffic shaping state of the flow. An OAM / RM recognizer 57 is provided to allow the flow control unit 55 to recognize these control cells and determine whether they are queued in the data path, control path, or both. Is advantageous.

The memory pointer for the cell of the traffic contract compliant flow is transmitted by the cell flow control unit 55 to the queue control unit 58 as “a”.
Queued per VC in response to the "ddCell" message. Each addCell message identifies the cell to which it belongs and the line index for the associated flow or VC. The addCell message also indicates whether the cell should be queued on the data path, control path, or both. Once the cells are properly queued, the newly queued cells are queued to inform the flow control unit 55 that consideration should be taken during future rate shaping operations for the VC to which it belongs. The control unit 58 returns an "added" message to the flow control unit 55.

Advantageously, the queue control unit 58 monitors the per-VC queue length for the depth control limits set for each queue. Thus, the control unit 58 can start the congruence control operation for the ABR flow when the queue for each VC becomes excessively long. The control unit 58
Can activate the appropriate policing function (not shown) to identify flows that are exceeding the traffic contract and to delete or record cells for such non-conforming flows.

The admission controller 61 monitors the "added" message returned from the queue control unit 58 so that the scheduler 62 transfers the queue cell head for the non-empty per-VC queue at the scheduled time. Put it on the calendar queue 63. The scheduler 62 determines the incorporation of the heads of the queue cells onto the calendar queue 63 and the respective virtual end times VT for which the scheduler 62 calculates.
In order to perform according to (f, j + 1) (or “virtual start time”), a virtual clock for each VC is employed. C. above. See 1.

The calendar queue 63 keeps track of the "real time" and "current time" of the system to prevent scheduled cells from being released for transmission prior to the scheduled time. That is, the scheduler 6
2 and the calendar queue 63 provide a stalled virtual clock, so that cells scheduled for transmission are released for transmission only when the real time of the system reaches the respective transmission time. As shown, connections having cells released for transmission by the calendar queue 63 are linked to a linked list of connections having cells ready for transmission on the transmission list 65.

The calendar queue 63 informs the flow control unit 55 whether to release cells for transmission on any given connection. The flow control unit 55 then requests a reference (ie, the new head of the queue cell) for the next cell (if any) on the per-VC queue for the given connection, and This reference is used by scheduler 62 for scheduling.
Notify to This ensures that the admission control 61 is effectively involved in closed loop communication with the calendar queue 63, and that the head of the queue cell that has granted scheduling is allowed to exclude all other cells in the queue per VC. To ensure.
Thus, the calendar queue 63 contains one or more time-limited time wheels, or “timelines” 66.
Can be implemented. The time span of these time wheels must be at least as large as the period of the lowest frequency flow designed to support the system to prevent ambiguity caused by time wrap, and Is 2
Times so that the relative times can be compared using two's complement arithmetic.

B. Specified peak flow rate by shaping the flow of fixed bit length data transfer limit Referring to FIG. 4, the stalled virtual clock transmission control is a time multiplexed flow of fixed bit length data transfer units such as ATM cells from the output waiting routing mechanism. To
It can be seen that it is very suitable for shaping to a flow rate of a specified peak data unit, such as PCR for DBR / CBR ATM service. As described above, a data transfer unit of a flow sent to a predetermined output port is put into a queue for each flow after being sent.
The admission controller 61 recognizes that the data transfer unit at the head of these queues is incorporated into the timeline calendar queue 63 by the scheduler 62 (except for all other transfer units). Then, the scheduler 62 performs a virtual clock operation for each flow with respect to the heads of these queue transfer units, and releases them from the calendar queue 63 according to their respective theoretical end times VT (f, j + 1) or their respective theoretical start times. To schedule.

The actual time is advantageously increased on the timeline 63 at such a rate that the shaped time multiplexed output traffic can essentially fill the bandwidth of the output link 71. As will be recalled, the maximum number of resolvable timeslots to which scheduler 62 can map the queue member head of each flow is based on the ratio of the highest to the lowest allowable frequency of these flows. Therefore, the actual time is increased by a logical multiple of the cell rate for each bucket.

The data transfer units present in the time slot representing a time before or equal to the current actual time, which is a reference for the timeline 63, are eligible to be selected for transmission, and therefore, as described above. To the transfer list 65. However, the data transfer units present in the time slots associated with the later time slots of the time line 63 will wait until the actual time of the system has advanced enough to reach these time slots. To avoid ambiguity due to rollover, the timeline 63 shows that all references to previously scheduled data transfer limits are replaced by any reference to later scheduled transfer units in anticipation of the next scan. It is removed from each time slot before it is inserted.

The above structure is adapted to DBR / CBR
Although effectively shaping ATM flows into PCRs specified by their traffic contracts, they do not help in adapting the cell delay variation (CDV) of these flows to the τ _PCR parameters of their traffic contracts.

C. Multiple Priority Levels to Minimize Relative CDV According to the present invention, different frequencies of AT
Give priority to data transfer units supplied to multiplexing points such as the output port of the M switch, and give higher priority to data transfer units of high-frequency flows than data transfer units of low-frequency flows that may collide Is also given a higher transmission priority. As shown in FIG. 3, the transmission priority is obtained by arranging data transfer units for which scheduling is permitted by the stalled virtual clock scheduling mechanism 63 or the like on the basis of the frequency of the flow to which these data transfer units belong. Implemented by leading to one or other of a plurality of timelines 66a-66e or output FIFO queues. For example, A
In the case of TM switches, it is recommended to implement with five different frequency dependent / service type dependent output priorities. This priority includes (1) the highest priority for cells from flows that have a negotiated output rate of at least 1/16 of the maximum rate of the output link; (2)
A second priority for cells from flows having a negotiated output rate in the range of 1/16 to 1/256 of the output link rate, and (3) one of the output link rates
A third priority for cells from flows with negotiated output rates in the range of / 256 to 1/4096 is included. The two lower priorities are for ABR connections with a non-zero negotiated MCR rate,
URB connection and ABR with zero MCR rate
It is preferable to set each for the connection.

As will be seen, the present invention effectively reduces the CDV of high frequency flows without substantially increasing the CDV of low frequency flows. In general terms, the acceptable CDV varies depending on the negotiated rate of the flow. For example, a CVD of 100 cell times is very large for one output interval per 10 expected cells, but if the negotiated output interval is only one for 2029 cells. Does not generally matter.

When employing a calendar queuing mechanism to schedule the transmission of data transfer units or cells for flows of different frequencies, high frequency, high priority flows are scheduled to achieve an acceptable low CDV, And while low frequency / low priority flows need to be decomposed to a single cell time accuracy, they can be coarsely decomposed with, for example, 16 cell time accuracy. This means that the number of time slots on the calendar queue 63 can be reduced. This allows the scheduling of the heads of the queue cells of low frequency flows at the expense of some accuracy which is generally not needed.
3 can be reduced in the amount of memory required to implement the third method.

The frequency-based prioritization technique according to the present invention, which is provided to eliminate the problem of transmission collisions at multiplex points between flows of nominally different constant frequencies, is disclosed in the accompanying US Pat. It can be employed in many different applications to reduce the relative jitter of a flow, such as an application having the task of maintaining a per-flow output queue to supply cells or other data transfer units.

D. Source to destination A
Traffic shaping of MCR and PCR parameters for ABR service in a network with a BR control loop As will be recalled, the standard ABR traffic contract is an explicit or implicitly specified PCR and M
Expected CR parameter, MCR may be zero.
An ABR flow that conforms to GCRA (1 / MCR, τ _MCR ) is valid and eligible to be offered, but the network has not granted it the allowed bandwidth. In contrast, GCRA (1 / PCR,
Flows that violate τ _PCR ) are incompatible. Thus, an ABR connection or flow sent to the output of a network or network element, such as by the ATM switch 21, may be: (1) ABR flows with a non-zero MCR guarantee of MCR:
Accepting the appropriate output bandwidth to effectively satisfy these guarantees, and (2) shaping to ensure that none of the ABRs violate their PCR commitment at such output It is advantageous. In networks with multiple multiplexing points, it is useful to perform this shaping at each of the multiplexing points.

FIG. 4 illustrates one technique for shaping ABR flows or VC connections at the output of a network or network element to conform to their traffic contracts. As shown,
Each flow monitors the rate at which it requires an egress service, and first at 85 the flow receives a GCRA (1 / MC
R, τ _MCR ) is determined, and at 86, the flow not conforming to GCRA (1 / MCR, τ _MCR ) conforms to GCRA (1 / PCR, τ _PCR ). Judge. These series of tests 85 and 86
Is reflected back to the scheduler 62 and its associated steering logic 87 so that appropriate rate shaping adjustments can be made to the rate at which the data transfer units (eg, cells) of each flow are provided at the output. To As will be seen, this required state is such that all ABRV flows have GCRA (1 / MC
R, τ _MCR ) (or barely fit) and false / match to GCRA (1 / PCR, τ _PCR ). This is because these algorithms test for the minimum and maximum allowable flow rate, respectively.

More specifically, GCRA (1 / MC
R, τ _MCR ) will not have received the required output bandwidth to meet their MCR guarantee. Thus, when such a flow is identified at 85, the scheduler 62 and its steering logic 87 will provide a relatively high priority stall virtual clock calendar for output at a speed somewhat above its guaranteed MCR. Set up on queue 88 to schedule subsequently received references to the data transfer unit (eg, cell) of the flow. For example, these subsequently received cell or packet references are incorporated on a calendar queue 88 for output at a rate determined by multiplying the MCR for the flow by a predetermined acceleration factor. This allows the network to give these additional, possibly "lagging" flows the additional output bandwidth needed to meet their MCR guarantee.

GCRA (1 / MCR, τ _MCR ) non-conforming A
BR flow, high _{GCRA (1 / PCR, τ PCR} ) than can be low or acceptable to the extent that allowed each further tested in 86 to determine whether the requesting output service in compliance or non-compliance rate Is done. The results of this test are captured by other control signals and the scheduler 62 and its steering logic 87
Are fed back to these GCRAs (1 / MC
(R, τ _MCR ) Set up scheduler 62 and its steering logic 87 to queue subsequently received references to non-conforming flows onto lower priority queue 89 in a speed sensitive manner. That is, 8
Subsequent received references for flows found to be GCRA (1 / PCR, τ _PCR ) compliant at 6 are appended to the transmission list for ABR flows with MCR guarantee. Because it is provided in round robin order. On the other hand, at 86, the flow is GCRA (1 / PCR, τ
_PCR ) If found to be non-conforming, then a subsequently received reference for that flow will follow the PCR for that flow and place the scheduler 6 on the non-work-maintained stalled virtual clock calendar queue 89.
2 thereby imposing a PCR limit on the flow.

More generally, it will be apparent that the embodiment of FIG. 4 implements the following ABR traffic shaping algorithm: if GCRA (1 / MCR,
τ _MCR ) if enqueued, queue to higher priority at MCR * acceleration factor rate; otherwise, if GCRA
(1 / _PCR, τ PCR) if compatible, queued low work conserving priority, otherwise, based on the PCR output interval, put in a non-work conserving lower priority queue.

However, it is clear that alternatives based on similar or different principles can be used.

For example, as shown in FIG. 5, one attractive alternative is to use GCRA at 85 and 86 (FIG. 4).
(1 / MCR, τ _MCR ) compliant and GCRA (1 / PC
R, τ _PCR ) The data transfer unit of the flow found to be conforming, ie, the reference to the cell, is (1)
In the MCR scheduling interval, it is to be placed in both the calendar queue 88 with high priority and the (2) work maintenance area of the queue with low priority. This creates a "race" situation. Because each time one instance of such a doubly queued reference is removed for service in any of the queues, the delinker 91 removes another instance of that reference from the other queue. Because. This means that a tighter limit on the tolerance τ will be maintained. The traffic shaping algorithm in this variant is as follows: when enqueuing, enqueue at higher priority with speed MCR, if GCRA (1 / PCR,
τ _PCR ) if acceptable, put into low priority work-maintaining queue; otherwise, based on PCR output interval, put into low priority non-work-maintaining queue and leave queue If leaving the queue with the higher priority, remove the link from the queue with the lower priority; otherwise, remove the link from the queue with the lower priority.

FIG. 6 shows yet another embodiment for shaping an ABR connection or flow. In this embodiment, MC has a guarantee that MCR is not zero.
The output bandwidth provided for each of the R non-conforming ABR flows is based on the average spacing between ABR cells on the output link (ie, the forwarding side of the line interface) and the current length of the ABR "ready queue". Adjusted dynamically.
The current length Len of this ABR “Lady Queue” is 93
Is determined by measuring the length of the work maintenance queue on the ABR transmission lists 65e and 65f. on the other hand,
The average spacing S between ABR cells on the calendar queue is ABR after grant for any non-ABR flow
By dividing the bandwidth available for the flow, it can be determined as the ratio of the link. The bandwidth required by these non-ABR flows is determined at 94 by calculating the overall bandwidth required to provide an entry for the non-ABR flows to transmission list 65.

For example, if at 94 it is determined that 50% of the output link bandwidth is available for the ABR flow, then naturally the average spacing S between ABR cells or data transfer units on the current output link is considered. Becomes 2 cell times. With this information and the calculation length Len of the ABR "ready queue", the scheduler 62 schedules the arriving references for these flows on the non-work keeping calendar queue 89 at each of the output intervals _{Tt described} below. Services of non-conforming ABR flows can be optimized.

T _t = Min ( _St , Max (P _t , R)) (3) where R = predicted round robin ABR service time (R = current time + S * Len), P _t = next to flow t the earliest service contract compliant with Dinner Departure (P _t = the current time _{+ 1 / PCR), S t} = flow t of the slowest service contract compliant with Dinner Departure (S _t = current time + 1 / MCR for the next cell for the cell ).

[Brief description of the drawings]

FIG. 1 is a simplified block diagram of an ATM switch in which it is advantageous to use the present invention.

FIG. 2 illustrates various forms that ATM cells preferably take while traversing the switch shown in FIG. 1;

FIG. 3 is a more detailed block diagram of an exemplary channel at the output side of the chip shown in FIG.

FIG. 4 is a schematic diagram illustrating an implementation of a stalled virtual clock calendar queue of the present invention.

FIG. 5 is a schematic diagram illustrating an implementation of another stalled virtual clock calendar queue of the present invention.

FIG. 6 is a schematic diagram illustrating an implementation of yet another stalled virtual clock calendar queue of the present invention.

[Explanation of symbols]

 58 Queue Control Unit 62 Scheduler 63 Calendar Queue 66 Timeline 87 Steering Logic 88 High Priority Calendar Queue 89 Low Priority Calendar Queue

 ──────────────────────────────────────────────────続 き Continuation of front page (72) Inventor Landis See. Rogers United States 03848 Kingston, Hampshire Rock Limon Road 61 (72) Inventor Christopher Jay. Coupler United States 02129 Charlestown, Massachusetts 13th Street 106 apartments 230

Claims (3)

[Claims] 1. In a packet switched communication system, a traffic shaping device that includes a contract and continuously sends out packets of a time multiplexed flow to substantially match individual network traffic contracts of each flow, The contract specifies a respective peak packet transmission rate and an associated peak rate tolerance for some of the flows, and a mutually exclusive category, i.e., non-zero minimum packet transmissions, each having an associated minimum rate tolerance specified by the contract. A category A flow having a rate;
Classifying each of the flows into one of the category B flows having a traffic contract that allows a zero packet transmission rate, the traffic shaping device determines each of the flows according to the oldest pending packet at the head of the queue order; A queuing mechanism for organizing pending packets of the active flow including the number of flows in a queue; a scheduling mechanism coupled to the queuing mechanism for scheduling the packets of the active flow for transmission; The scheduling mechanism may include a relatively high priority non-work maintenance calendar queue, at least one relatively low priority work maintenance queue, and the high priority queue and the low priority queue. Priority And a non-work-maintaining queue of decisions, said scheduling mechanism comprising: (i) in said high-priority non-work-maintaining calendar queue, a specified minimum delivery rate for a first given flow; (Ii) transmitting the packet of the first flow at a rate exceeding a contract-specified minimum transmission rate for the first flow when the packet of the first flow is transmitted at a rate exceeding the tolerance. In a maintenance calendar queue, when a packet of the first flow is being sent at a rate that exceeds a tolerance of a specified peak sending rate of the first flow, a designated peak transmission for the first flow is performed. (Iii) in the work maintenance queue, the packets of the first flow are transmitted so as to be transmitted at substantially the same rate as the first flow. Scheduling any given Category A packet to be sent according to the selected continuous service strategy when being sent according to the traffic contract, the scheduling mechanism comprising: (iv) at the other non-work keeping calendar queue , A specified peak for the second flow when packets of a second given flow are being sent at a rate that exceeds a tolerance relative to a specified peak delivery rate for the second flow. (V) To send at almost the same rate as the sending rate,
Scheduling any given category B packets to be sent according to the selected strategy in the work maintenance queue when the packets of the second flow are being sent according to the traffic contract of the second flow; Traffic shaping device. 2. A packet shaping communication system, comprising: a traffic shaping device that includes a contract and continuously sends out packets of a time multiplexed flow to substantially conform to individual network traffic contracts of each flow, The contract entrusts each of the flows with a respective allowable peak packet transmission rate and a mutually exclusive category, ie, a category A flow to which each non-zero allowable minimum packet transmission rate is entrusted, and a zero packet transmission rate. Classifying each of the flows into one of the category B flows that receive a traffic contract that allows the traffic shaping device to receive a traffic contract according to the oldest pending packet at the head of the queue. Go above A queuing mechanism for organizing packets of the active flow including the flows, and a scheduling mechanism coupled to the queuing mechanism for scheduling the packets of the active flow for transmission; A high-priority non-work maintenance calendar queue, at least one relatively low-priority work maintenance queue, and a priority not yet determined for the high-priority queue and the low-priority queue. Not to selectively schedule packets scheduled in the high-priority queue and the low-priority queue in combination with monitoring and sending from the high-priority queue and the low-priority queue. Including a control mechanism; The juling mechanism includes: (i) scheduling packets for each flow in the high priority calendar queue to be sent at a rate exceeding the minimum rate commitment for each flow; and (ii) defining a predetermined work maintenance queue. Preparing packets of any category A flows by placing them in the work queue to be sent according to the outgoing strategy, thus creating a race condition for the packets of these category A flows, the control mechanism comprising: When arriving at one of the high-priority queue and the maintenance queue, this packet is taken out of the other packets of the queue by removing the packet of the delayed category A flow. Resolve the race condition and The Nism sends out all Category A and Category B flow packets that are sent at a rate that exceeds the allowable peak sending rate entrusted to each flow at approximately the same rate as the peak sending rate entrusted to each flow. Scheduling in the other non-work maintenance calendar queue such that the scheduling mechanism removes from the work maintenance queue any Category B flow packets that are being sent at a rate that is compatible with the contract delegation for each flow. A traffic shaping device, wherein the traffic shaping device queues the work maintenance queue for delivery according to a delivery strategy. 3. A traffic shaping process for continuously sending fixed bit length packets of a plurality of time multiplexed flows over a link of known bandwidth to substantially conform to an individual traffic contract for each flow. A routine for shaping at least some of said flows into respective available bit rate traffic contracts specifying respective peak rates and minimum packet transmission rates greater than zero for said some flows, The routine includes the steps of organizing the packets of the active flow including the number of flows in each queue according to the oldest pending packet at the head of the queue, the flow having an available bit rate traffic contract. Dynamically determining the number of active flows with available bit rate traffic contracts, comprising dynamically determining the bandwidth available on the link as a percentage of the known bandwidth for For each head of a queue packet of each flow that receives a bit rate traffic contract, a first theoretical delivery time based on a contract specified minimum delivery rate for the flow and the link for an available bit rate flow A second theoretical transmission time based on an active available bit rate per-flow allocation of available bandwidth and a third theoretical transmission time based on a contract-specified peak transmission rate for said flow. Deciding, the second Discarding the earliest of the third theoretical transmission times, and scheduling the packet in the non-work-maintaining queue to transmit at the earliest of the remaining theoretical transmission times. , Routine.