JP2003511906A - Method for fair flow control in packet-switched networks - Google Patents

Abstract

(57) [Problem] To provide a fair flow control method in a packet switching network. A packet switching network includes a plurality of nodes, each node coupled to a plurality of sources for transmitting / receiving data. In addition, the packet switching network includes a data queue for storing data transmitted from a source in relation to a current queue length and a target queue length. In the fair flow control method, each node estimates the number of local bottleneck VCs using the ER and the guaranteed MCR of the corresponding source. The node assigns an ER to each source based on the difference between the current queue length and the target queue length, the derivative of the current queue length, and the estimated number of local bottleneck VCs; The ER is transmitted to each source through a feedback signal.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD OF THE INVENTION

The present invention relates to packet switched networks, and more particularly to a fair flow control method.

[0002]

[Prior art]

Asynchronous Transfer Mode (Asynchronous Transfer Mode)
ode: hereinafter referred to as ATM) network and the Internet.
Fair flow control has a significant impact on information transmission in packet switched networks. In particular, fair flow control in ATM networks is based on the available bit rate.
(Available Bit Rate: hereinafter referred to as ABR) Related to services.

The ATM includes CBR (Constant Bit Rate), VBR (Variable Bit Rate), and U.
It provides four service methods: BR (Unspecified Bit Rate) and ABR. In the ABR service, the source dynamically changes its bit rate within the available bandwidth from the network to adapt to changing network conditions when data is transmitted. The ABR service has been introduced into ATM networks to support data applications that were not efficiently supported by bandwidth guarantee services such as VBR. More specifically, S. Sathaye
, “ATM forum Traffic Management Specification, Version 4.0”, Feb. 1996
And F. Bonomi and KW Fendick, “The Rate-Based Flow Control Framework
for the Available Bit Rate ATM Service ”, IEEE Network, vol. 9, no. 2,
pp. 25-39, 1995 and R. Jain, “Congestion Control and Traffic Management
in ATM Networks: Recent Advances and Survey ”, Computer Networks and ISD
N Systems, vol.28, no.13, pp1723-1738, 1996.

Most data applications are highly irregular and difficult to predict data traffic. For successful data transmission, the network should operate according to predetermined conditions that satisfy the predefined cell loss requirements, time varying requirements, and cell delays. Due to the above characteristics, the network should change the data transmission rate according to its load condition. Therefore, the concept of elastic traffic service has been introduced which adjusts the data transmission rate based on the available bandwidth from the network. A typical example is the ABR service in the ATM network.

In principle, the ABR service does not require complex traffic characteristics and call admission control at the source and the switch, respectively. Due to the above-mentioned simplicity, the implementation and development of the ABR service is not limited to the bandwidth guarantee service (that is,
It was expected to be much easier than CBR or VBR services). However, in practice, implementing an ABR-assisted switch was much more difficult than initially anticipated.
The hardest part is the simple, scalable, and stable ABR flow algorithm, especially in explicit and decentralized network environments.
t Rate: hereinafter referred to as ER) is to design an allocation algorithm.

The ATM Forum has adopted a close-loop rate-based approach for ABR flow control. The closed-loop rate-based flow control, as its name implies, uses feedback information from the network to control the rate at which each source can transmit multiple cells to the network. The feedback information is resource management.
It is transmitted to the source by a specific control cell called cell (Resource Management: RM). The three mechanisms of the switch for writing the congestion state on the RM cell are EFCI (Explicit Forward Congestion Indication) marking,
RR (Relative Rate) marking and ER (Explicit Rate) marking,
At least one of the three mechanisms should be implemented on the switch for the rate-based flow control.

On the other hand, long and various round-trip delays included in the closed loop.
And the distributed bottleneck points of ABR VC (Virtual Circuit) make the design of high performance ER allocation algorithm difficult. A in the network when the transmission rate of the ABR source is determined by the network state information at different times.
BR cues are difficult to stabilize. In particular, the binary feedback mechanism (EFCI
Marking or RR marking, or EFCI marking and RR marking only), the ABR cue in the steady state always shows a persistent oscillation, whose amplitude is between delay and bandwidth. It increases with the product. Specifically, Hernandez-Valencia, et al., “Rate Contr
ol Algorithms for the ATM ABR Service ”, European Transactions on Teleco
mmunications, vol. 8, no. 1, pp. 7-20, 1997, and F. Bonnomi, D. Mitra, an
d JB Serry, “Adaptive Algorithms for Feedback-Based Flow Control in
High-Speed, Wide-Area ATM Networks ”, IEEE J. Select. Areas on Communica
tions, vol. 13, no. 7, pp. 1267-1283, 1995 and KK Ratmarkrishnan and J
ain, “A Binary Feedback Scheme for Congestion Avoidance in Computer Net
works with a Connectionless Network Layer ”, Proc. ACM SIGCOMM'88, pp. 3
See 03-313, 1988.

ABR queue oscillations as described above increase the likelihood of cell loss and the potential for low utilization of the link due to periodic buffer overflows and underflows. In order to achieve the asymptotic stability of the ABR queue, we overcome the shortcomings of the binary feedback mechanism by introducing an ABR flow control scheme using ER marking. However, it is still difficult to design an asymptotically stable ER allocation algorithm in a simple form. The above-mentioned problem is a problem of feedback control having a delay.

L. Benhanhaned and S. M. Meerkov formulate the problem of rate-based flow control as a problem of discrete-time feedback control with delay to achieve asymptotic stability and arbitrary closed-loop performance. We proposed an ER allocation algorithm that allows the control of ER. The contents are “Feedback Control of Congestion i
n Packet Switching Networks: The Case of Single Congested Node ”, IEEE / A
CM Trans. On Networking, vol. 1, no. 6, pp. 693-708, 1993 and “Feedback
Control of Congestion in Packet Switching Networks: The Case of Multipl
e Congested Nodes ”, International Journal of Communication Systems, vol
. 10, no. 5, pp. 227-246, 1997. The ER allocation algorithm is formed as in <Formula 1>.

[0010]

[Equation 1] Where r [k] is the ER calculated by the switch at discrete time k,
q [k] is the length of the ABR queue by class at time k, q _T is the target queue length, α _i and β _j are the controller gains, and τ _max is ABR V
Is the maximum round trip time delay of C and I is any constant greater than 0

Despite the above-mentioned theoretical grounds, the implementation complexity is so high that there is a limit to the practical use of the algorithm (1). The disadvantages and limitations of the algorithm are described by A. Kolarov and G. Ramamurty in "A Control Theo".
retic Approach to the Design of Close Loop Rate Based Flow Control for H
igh Speed ATM Networks ”, Proc., IEEE INFOCOM'97, vol. 1, pp. 293-301,
Explained in 1997. In the ER allocation algorithm, the ER term should be maintained up to the time lags τ _max , and multiple floating point multiplications are performed every discrete time slot. Should be.

At the same time, S. Chong said “Second-Order Rate-Based Flow Control with
Dynamic Queue Threshold for High-Speed Wide-Area ATM Networks ”, preprin
In 1997, a simpler control theory ER assignment algorithm was proposed. AE
Iwalid is the “Analysis of Adaptive Rate-Based Congestion Control for
High-Speed Wide-Area Networks ”, Proc. IEEE ICC'95, pp. 1948-1953, 1995
Then, we proposed a continuous-time ER allocation algorithm according to <Equation 2>.

[Equation 2] <Equation 2> provides a necessary and sufficient condition for asymptotic stabilization of a closed-loop system when the round-trip time delays of all VCs are the same. Chong extended the stability analysis of the algorithm to the general case with arbitrary round trip time delays.

On the other hand, S. Chong, R. Nagarajan, and Y. T. Wang,
“Designing Stable ABR Flow Control with Rate Feedback and Open-Loop Con
trol: First-Order Control Case ”, Performance Evaluation, vol. 34, no. 4
The simpler ER allocation algorithm proposed in [3] can easily achieve an asymptotically stable system and is expressed by <Equation 3>.

[Equation 3] [X] ⁺ = max [x, 0] means that the larger one of x and 0 should be selected.

The ER allocation algorithm (3) is described in “First-Order Rate-Based Flow Control of US Pat. No. 5,864,538 dated January 26, 1999.
with Adaptive Queue Threshold for ATM Networks. In the algorithm (2), two different stabilizing conditions were derived.
One of them is a sufficient condition for the general case with different round trip time delays, and the other one is a necessary and sufficient condition for the special case with similar round trip time delays.

The common disadvantage of said algorithms (2) and (3) compared to algorithm (1) is that of the available bandwidth for ABR traffic and the available bandwidth utilized by the remote bottleneck VC. If the controller gain and the queue length critical value are not properly selected due to the momentary recognition in the fraction, the length of the ABR queue will be insufficient because the link is not fully utilized at the equilibrium point. It can converge to 0, which is undesirable. If the transmission rate of the VC is not limited by its own PCR (Peak Cell Rate), a bottleneck situation will occur on another link, so the remote bottleneck VC
Cannot distribute the links fairly. On the contrary, when the above algorithm (1) is applied, there are no such undesirable equilibrium points.

[0016]

[Problems to be Solved by the Invention]

It is an object of the present invention to provide a method that guarantees maximum link utilization and minimum cell loss regardless of the round trip time delay of the ABR loop. Another object of the present invention is to provide a method for minimizing the size requirement of the ABR queue by ensuring asymptotic stability of the ABR queue. Still another object of the present invention is to ensure the fair distribution of available bandwidth to each ABR service user, thereby maximizing the MAX-MIN fairness based on the ATM Forum standard draft. It is to provide a way to guarantee.

Yet another object of the present invention is to provide a method for increasing responsiveness and transient control performance to changes in the communication network environment such as changes in the number of ABR service users and changes in ABR bandwidth. is there. Another object of the present invention is to provide a method for providing all functions including EFCI marking, RR marking, and ER marking specified in the ATM Forum Traffic Management Standard Draft. Yet another object of the present invention is that by the existence of an asymptotic stabilizing operating point,
It is to provide a method of achieving high availability, low cell loss, and MAX-MIN fair rate allocation.

Yet another object of the present invention is the time on multiple time scales, namely VBR and AB.
It is an object of the present invention to provide a method of increasing responsiveness to a network load change at the time of changing the cell level rate of the RVC and at the arrival and departure of the cell level of the VBR and ABR VCs. Yet another object of the present invention is to minimize the number of operations required for the calculation of the ABR service algorithm and to perform VC-specific operations including VC-specific queuing, VC-specific calculation, and VC-specific table access. The object is to provide a method of achieving the reduction of the implementation complexity and the extensibility by actually removing it.

[0019]

[Means for Solving the Problems]

The present invention for solving the above object provides a fair flow control method in a packet switching network. The packet switching network includes a plurality of nodes, and each node is connected to a plurality of sources for transmitting / receiving data. The packet switching network also includes a data queue that stores data transmitted from the source in relation to a current queue length and a target queue length. In the fair flow control method, each node estimates the number of local bottlenecks VC using the power MCR (Minimum Cell Rate) guaranteed by the ER and the corresponding source. The node determines the ER based on the difference between the current queue length and the target queue length, a derivative of the current queue length, and the estimated number of local bottlenecks VC. The ER is transmitted to each source and transmitted to each source through a feedback signal.

[0020]

DETAILED DESCRIPTION OF THE INVENTION

Hereinafter, a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings. In the following description, a detailed description of known functions and configurations incorporated herein will be omitted when it may obscure the subject matter of the present invention. FIG. 1 is a network configuration diagram for implementing the ABR service. Referring to FIG. 1, the network 2 for providing the ABR service includes a plurality of nodes 4, 6, 8 and 10. The node conceptually represents a switch, and the node is hereinafter referred to as a switch. Each switch is connected to a plurality of sources. In FIG. 1, the switch 4 is connected to a source A12 and a source C14, and the switch 8 is connected to a source B16 and a source D18. Each source transmits / receives data through a switch 4 or 8 connected to the source. Data transmitted from the source reaches a destination through a route called a VC route having a plurality of nodes. For example, the data from the source A12 reaches the destination source B16 through the VC path having the switches 4, 6, and 8.

In the ABR service, the bandwidth availability of the network is conveyed to the source through a special cell called an RM cell. In FIG. 1, the RM cell is expressed as RMc. The RM cell comprises the sources 12, 14, 16, and 18,
Alternatively, the RM cells generated in the switches 4, 6, 8, and 10 and described below are limited to the RM cells generated in the source. The RM cells generated at the source are transmitted to the destination through the switches 4, 6 and 8 via the VC path. The transmission direction of the cell as described above is forward (forwa
rd). Upon receiving the forward RM cell, the destination returns the cell to the source (backward). In FIG. 1, reference numerals 12 and 14 indicate sources, and reference numerals 16 and 18 indicate destinations. In relation to the source and destination, forward and reverse directions are defined as shown. The switches 8, 6, and 4
Writes the permissible bandwidth information to the reverse RM cell. The sauce 12
Alternatively, 14 adapts its own transmission rate according to the change of the network condition based on the received bandwidth information.

The RM cell is a CCR (Current Cell Rate), MCR, ER, NI (No) shown in FIG.
Increase) and information having a CI (Congestion Indication) field. In FIG. 2, the SR (source-receive) field provides information related to the source and the destination. The CCR field is set by the source as the current ACR (Allowed Cell Rate) when the source generates an RM cell. The MCR field indicates a minimum bandwidth allocated to each VC when the source generates an RM cell. ER is the available bandwidth written by the switch's ABR service engine to the reverse RM cells generated from the source as it passes through the switch. The former value is written to the ER field only when the calculated available bandwidth of the ABR service engine is less than the existing available bandwidth. Therefore, the minimum available bandwidth of the VC path is communicated to the source. Therefore, the source writes its own PCR (peak cell rate) in the ER field when generating an RM cell. CI and NI are fields used to control the RR. The CI informs the source that the network is too congested and the bandwidth of the source should be reduced, and the NI is used to prevent the ACR of the source from increasing.

The switch uses a switch algorithm to calculate the available bandwidth for the ABR service and writes the available bandwidth information in the reverse RM cell. What is to be obtained from the switch algorithm is
Is the bandwidth allowed for. Since the available bandwidth is transferred to the source, the source modifies the transmission rate and provides it to the ABR service as reliable.

In each switch of FIG. 1, the switch fabric 20 of FIG. 3 has an input / output port connected to an input / output port card. Each I / O port card 2
2 includes an input / output buffer management unit 30, an ABR service engine 32, and an output interface 34, as shown in FIG. The input / output buffer management unit 30 is connected to the switch fabric 20 and is in charge of input / output queuing.
The input / output buffer management unit 30 includes an ABR queue 36. According to an embodiment of the present invention, the ABR service engine 32 implements an ABR algorithm and related functions for the ABR service while an RM cell passes through the ABR service engine 32. The output interface 34 performs a user network interface function of the ATM layer.

FIG. 5 is a block diagram of the ABR service engine 32 shown in FIG. Referring to FIG. 5, the ABR service engine includes an EFCI marking unit 40, | Q
| Estimation unit 42, ER calculation unit 44, average queue calculation unit 46, and backward RM cell writing unit 48. When the EFCI congestion occurs, the EFCI marking unit 40 displays an EFCI bit indicating the EFCI congestion in the input forward data cell. The cue read signal is generated when removing a cell from the queue and the cue write signal is generated when a new cell is input to the cue. The average queue calculator 46 increments the queue instant variable by 1 when receiving the queue write signal, and decrements the queue instant variable by 1 when receiving the queue read signal, to determine a predetermined value for the ER calculation. Calculate the average queue length during the cycle. The average queue calculation unit 46 calculates the average queue length.

[Equation 4] Is output to the ER calculation unit 44. The average queue calculator 46 uses the queue read signal and the queue write signal to determine whether EFCI congestion occurs, and when the EFCI congestion occurs, the EFCI marking unit 40.
To the EFCI congestion signal EFCI_CG. According to an exemplary embodiment of the present invention, the average queue calculator 46 may also generate a signal CG (congest
ed) and VCG (very congested) indicating "very congested" are output to the reverse RM cell writing unit 48. The crowded and very crowded states are
It is determined by a predetermined minimum queue length threshold q _LT and a predetermined maximum queue length threshold q _HT for the ABR queue. The signals CG and VCG are used to control the RR according to an embodiment of the present invention.

The | Q | estimation unit 42 reads the CCR and MCR from the RM cell, and
Estimate the number of local bottleneck VCs by comparing MCR with the periodically calculated ER value r (t). The estimated value

[Equation 5] The interval W for calculating is described in detail in "(5) Discrete-time ER algorithm and | Q | estimation" below. The above

[Equation 6] Is provided to the ER calculator 44.

The ER calculator 44 periodically calculates and updates the ER, and when the backward RM cell arrives, outputs the recently updated ER value r (t) to the backward RM cell write unit 48. . The reverse RM cell writing unit 48 writes r (t) in the ER field of the reverse RM cell. More specifically, the reverse RM cell writing unit 48 receives the ER of the received RM cell, r (t) + MCR (r (t) + mi) of the RM cell,
And r (t) + mi is smaller than the ER of the RM cell, the r
Write (t) + mi in the ER field. The reverse RM cell writing unit 48
Is a signal CG received from the average queue calculator 46 for controlling RR.
And VCG write binary logic state bits in the NI and CI fields of the RM cell.

FIG. 16 is a flowchart showing the ER allocation control operation of the ABR service engine. Referring to FIGS. 5 and 16, the | Q | estimation unit 42 estimates Q in step 100, and the ER engine 44 updates ER by periodic calculation in step 102. When an RM cell arrives at the backward RM cell writing unit 46 in step 104, the ABR service engine 32 proceeds to step 106, and when no RM cell arrives, returns to step 100. In step 106, the reverse direction RM cell writing unit 46.
Reads the recently updated ER value r (t) from the ER engine 44 and the MCR value mi from the received RM cell. In step 108, the reverse RM
The cell writing unit 46 outputs the ER for VCi (i-th VC) by r (t) + mi.
The assigned value ri (t) is calculated, and in step 110, ri (t) is written in the ER field of the backward RM cell.

The purpose of the switching algorithm proposed by the present invention is to converge the ABR queue 36 to the target queue length q _T and to set the steady state queue length to the target queue length q _T. It is to maintain. Maintaining the queue length means that the input traffic of the ABR queue 36 is the same as the output traffic. The transient period occurs in the process of achieving the target queue length q _T. If the input traffic is momentarily larger than the output traffic, the ABR queue 36 will be overloaded, resulting in unexpected cell loss. Therefore, in the embodiment of the present invention, the switch algorithm is implemented in consideration of the changes in the buffer capacity and the queue length of the ABR queue 36.

Embodiments of the present invention propose a continuous-time ER allocation algorithm in order to provide a controller that can remove undesired imbalance points as simply as possible and at the same time.

[Equation 7] Where r (t) is the ER calculated by the ER engine,

[Equation 8] Is a differential value of r (t). Q is a set of local bottleneck VCs,
Q | is the number of elements (cardinality) of Q, and q (t) is the above A at time t.
The length of the BR queue 36, where A and B are controller gains that are varied by the length of the ABR queue 36 for asymptotic stability in embodiments of the invention.

The ER allocation algorithm (4) uses q (t) instead of r (t) for the damping term (first term on the right side) compared to the algorithm (2).

[Equation 9] Is a differential value of q (t) and will be described with reference to <Expression 21> and <Expression 23>.

Such a modification effectively eliminates the undesired unbalanced points in a closed loop system, so that the ABR queue 36 is used by available bandwidth and remotely bottlenecked VCs. Irrespective of some available bandwidth, it always converges to the target queue length q _T. This means that in steady state, full utilization of available bandwidth is always guaranteed. Another important feature of the proposed algorithm is the normalization of the controller gains A and B by the number of local bottleneck VCs | Q |. The normalization makes the asymptotic damping factor of the closed-loop system independent of | Q |.

The estimation of | Q | has become an important research subject in rate-based ABR flow control research. Related papers include MK Wong and F. Bonomi, “A Novel Expli
cit Rate Congestion Control Algorithm ”, it Proc. IEEE GLOBECOM'98, vol.
4, pp. 2432-2439, 1998 and L. Kalampoukas, A. Varma and KK Ramakrishn.
an, “An Efficient Rate Allocation Algorithm for ATM Networks Providing M
AX-MIN Fairness ”, Technical Report UCSC-CRL-95-29, Computer Engineering
Dept., University of California, Santa Cruz, June 1995 and A. Charny, K.
K. Ramakrishnam and A. Lauck, “Time Scale Analysis and Scalability Iss
ue for Explicit Rate Allocation in ATM Networks ”, IEEE / ACM Trans. On Ne
tworking, vol. 4, pp. 569-581, 1996 and R. Jain et al., “ERICA Switch Al
gorithm: A Complete Description ”, ATM Forum / 96-1172, 1996.

The difficulty in the | Q | estimation is that the dynamic characteristic of the | Q | estimation process is the dynamic characteristic of the ER allocation process, that is, the ER update is | Q until the closed loop system reaches a steady state. │Q│ A non-conforming design of the estimation algorithm can destabilize the closed-loop system, either because it affects the update, or because the | Q | update is coupled to the dynamic properties that affect the ER update. That is. In view of the above, in connection with the ER allocation algorithm according to the embodiment of the present invention,
We propose a stable and scalable | Q | estimation algorithm.

The control theory ER allocation algorithm (4) of the present invention achieves maximum-minimum fairness allocation according to the following principle. According to <Equation 4>, the same ER (common ER) is assigned to all VCs and shares the same link. Therefore, the minimum allocation ER of the path is transmitted to the source through the RM cell, and the source is
Data is transmitted with the minimum ER. When another VC in a different location becomes a bottleneck and the source transmits data at a rate lower than the common ER assigned by the switch, the ABR queue 36 has a time delay and the delay is attenuated. . When the ABR queue 36 is less than or equal to the target queue length q _T , the switch uses the algorithm (4) to set the common ER until the ABR queue 36 reaches the target queue length q _T. increase. Thus, each local bottleneck VC allows for fair sharing of bandwidth not used by remote bottleneck VCs.

The above algorithm (4) becomes like the following discrete time algorithm in actual application.

[Equation 10] Here, T is the duration of the discrete time slot. algorithm(
5) is a special case of algorithm (1). That is,

[Equation 11] If set to, algorithm (1) is converted to algorithm (5). algorithm(
Contrary to 1), the scaled algorithm does not allow any further control of closed-loop dynamics. However, the inventor of the present invention argues that providing arbitrary control capability is very expensive to implement and is not necessary for ABR flow control design. To support the assertion, the scaled algorithm actually considers an acceptable level of control over the closed loop performance.

In the control theory ER allocation algorithms (4) and (5), the queue length control is the main concern, and the fairness allocation is a secondary result of the queue length control. In other forms of ER algorithms (14)-(18), fairness allocation is a major concern and queue length control is secondary. In another form of the algorithm, each switch needs to keep track of available bandwidth, fair sharing of remote local bottleneck VCs on its bottleneck link, and number of local remote bottleneck VCs. Based on the above information, the switch should update the ER allocation per VC by a method that asymptotically achieves the maximum-minimum fairness allocation and the target link utilization.

(1) ER Rate Flow Control By mathematically modeling the ER allocation algorithm of the present invention, the ER allocation algorithm ensures the MCR required by each source, and at the same time the maximum-minimum fairness. To satisfy. FIG. 6 is a diagram showing a network model having nodes of interest. Figure 6
In, the plurality of VCs correspond to the corresponding plurality of sources, the VCi 54 corresponds to the source i50, and the VCj56 corresponds to the source j52. Reference numeral 36 is the ABR
Denotes the queue, q _T is the target queue length, and μ is the available bandwidth of the link shared by multiple VCs. τ _i ^f and τ _i ^b represent the forward path delay and the backward path delay in VCi 54, respectively, and the sum of τ _i ^f and τ _i ^b is VCi 54.
54 round trip time delay τ _i . τ _j ^f and τ _j ^b denote the forward path delay and the backward path delay at VCj 56, respectively, and the sum of τ _j ^f and τ _j ^b is the round trip time delay τ _j of VCj 56.

The network model is analyzed according to the following assumptions. Assumption 1. Traffic is considered a deterministic flow, and the network queuing process and the flow control mechanism are considered continuous in time. The above assumptions allow for the implementation of closed loop systems modeled by different equations.

Assumption 2. Round trip time delay τ _i of VCi 54 is the sum of forward path delay τ _i ^f and reverse path delay τ _i ^b including propagation, queuing, transmission, and processing time. . Here, it is assumed that the round trip time delay is constant. Assumption 3. The source has persistent properties until the system reaches steady state. Here, “persistent” means that the source always has sufficient data to transmit at the assigned transmission rate.

Assumption 4. There is no VC arrival or departure until the system reaches steady state. Assumption 5. The available bandwidth μ on the link is constant until the system reaches a steady state. Also assume that the size of the buffer in the link is infinite.

Α _i (t) indicates the transmission rate when the source _i 50 transmits data at the source time t, and r _i (t) is the ER of VCi calculated by the node of interest at the node time t. Indicates. Also, β _i (t) indicates the latest minimum value of ER assigned to VCi by the node along the path of VCi, excluding the ER assigned by the node of interest, and p _i (t) t) indicates PCR restriction of VCi. Neglecting linear increase and exponential decrease based on binary feedback,
The operation of the source is modeled as in <Equation 6>.

[Equation 12] Here, N is a set of all VCs including a node whose route is interested. The model is such that the source i54 is r _i (t−τ _i ^b ), which is the ER value assigned by the node of interest before the backward path delay τ _i ^b , the route other than the node of interest. It means that data is transmitted at the minimum transmission rate of the above minimum ER value b _i (t) and the above PCR value pi of VC.

The operation characteristic of the ABR queue 36 for each class of interest is obtained by <Expression 7>.

[Equation 13] The ER allocation algorithm according to the exemplary embodiment of the present invention is configured such that the queue length q (t) and a derivative value of the queue length.

[Equation 14] And an estimated value of the number of the local bottleneck VCs

[Equation 15] This is an algorithm executed in each switch based on the current network state including the following, and is obtained by <Expression 8> and <Expression 9>.

[0044]

[Equation 16] Here, A and B> 0, and mi is the MCR in the node requesting security while the VCi is alive. here,

[Equation 17] , There is a call admission control that guarantees the following conditions.

[Equation 18]

Here, r (t) is a common part of ER allocation values r _i (t) for each VC

[Formula 19] And is needed for most of the calculations of the algorithm. According to an embodiment of the present invention, the VC-specific calculation required is just to add mi to the common ER, r (t). The reason is that the algorithm reduces the computational complexity with increasing number of VCs and makes it scalable. Each RM cell of VCi 54 conveys said value mi on a round trip.

In an embodiment of the present invention, the node uses the common E in the background calculation.
It is characterized in that R is updated. The "background" calculation is to calculate the common ER periodically, regardless of the reception of RM cells. The advantage of the background calculation is that when the RM cell arrives at the corresponding node, the recently updated common ER,
That is, r (t) can be provided immediately. Through the background calculation, the node that updates the common ER value r (t) by <Equation 9> reads mi from the MCR field of the VCi 54 that passes through the RM cell, and in <Equation 8>, sets mi to the latest common ER value r ( VCi by adding to
The ER value r _i (t) assigned to 54 is calculated, and r _i (t) is written in the ER field of the RM cell.

An important feature of the ER allocation algorithm according to the embodiment of the present invention is an estimated value of the number of local bottlenecks VCs.

[Equation 20] Is the normalization of the controller gains A and B according to The normalization is an optional choice, not an absolute requirement, but a recommendation, as described below in "(4) Principal roots and asymptotic decay rate". The normalization makes the asymptotic decay rate of the closed loop system independent of the number of local VCs.

The remote bottleneck VC and the local bottleneck VC are defined in a steady state for a given network load. Remote bottleneck V on a given link
C is defined as the VC that cannot achieve fair sharing on the given link because of the transmission rate limited by the PCR or the bottleneck situation on some other links of the path. Similarly, a local bottleneck VC on a link is defined as a VC that achieves fair sharing on the given link.

To mathematically define the local bottleneck VC and the remote bottleneck VC,

[Equation 21] Leave. Here, a _is , r _is , and b _is represent a _i (t), r _i (t), and b _i (t) in the steady state, respectively. All of the above local bottleneck VCs
The set Q of is obtained by <Equation 11>.

[Equation 22] The set NQ of all remote bottlenecks VC is obtained by <Equation 12>.

[Equation 23]

In “(5) Discrete-time ER algorithm and | Q | estimation”, the | Q | estimation algorithm is

[Equation 24] Is not efficient to converge to | Q | quickly, but during the transition period,

[Equation 25] Indicates that there is a strong tendency to become larger than | Q |. Robustness as described above
tness) is necessary for the stability of the closed loop control.

(2) Steady State and Fairness Hereinafter, the steady state characteristic of the closed loop dynamic characteristic when the ER allocation algorithm according to the embodiment of the present invention is applied will be described. That is, the analysis result according to the embodiment of the present invention will be described. The closed-loop dynamic characteristic has a differential value of the system variable of 0, that is,

[Equation 26] Assuming that there is an equilibrium point that becomes, the following equations are obtained.

[Equation 27]

[Equation 28]

[0052] here,

[Equation 29] And other notations are predefined. Since q _s = q _T > 0, <Expression 17> is <
Expression 16> is meant.

[Equation 30] <Expression 17> is obtained by combining <Expression 13>, <Expression 14>, and <Expression 16> with the definitions of <Expression 11> and <Expression 12>.

[Equation 31] <Expression 17> means <Expression 18>.

[Equation 32]

As described above, the embodiment of the present invention has the following features, and as a result, <Formula 19> is obtained.

[Expression 33] , (I) the queue length is the same as the target queue length (q _s = q _T ), and (ii
) The available bandwidth on the link is fully used

[Equation 34] , (Iii) the individual MCR is guaranteed on the link, the bandwidth minus the sum of the MCRs

[Equation 35] Is the only steady-state solution (equilibrium point) that is fairly shared in terms of maximum-minimum fairness. That is,

[Equation 36]

The above-mentioned feature is that, when the ER algorithm according to the embodiment of the present invention is applied, the ABR closed-loop system has the only oscillation-free operating point that achieves MCR guarantee and maximum-minimum fairness. It means that the queue length is the same as the target value q _T , regardless of the load on the network. That is, the equilibrium point is independent of the available bandwidth for the ABR traffic and a portion of the available bandwidth used by the remote bottleneck VC, including the PCR bound VC.

A point to be noted regarding the embodiment of the present invention is that in the ER algorithm, the calculation is very simple while generating the only equilibrium point. The ER algorithm uses q (t) and

[Equation 37] Only the other measurement or monitoring result is required. Controller gain A and B
Said estimate used to normalize

[Equation 38] Is introduced to help the cue length converge to the target cue length and is independent of the equilibrium point. Benmohamed algorithm of <Equation 1>
The (Benmohamed's algorithm) also generates a unique system equilibrium point having the characteristics of the ER allocation algorithm of the present invention, but has a high implementation complexity. Conversely, <
Applying the algorithms of Equation 2> and <Equation 3>, the closed loop system has a balancing point that varies with the available bandwidth and a portion of the available bandwidth used by the remote bottleneck VC, and in the worst case Has a queue length of zero. Specifically, S. Chong, R. Nagarijan, and YT Wang, “Design Stable ABR Fl
ow Control with Rate Feedback and Open-Loop Control: First-Order Control
Case ”, Performance Evaluation vol. 34, no. 4, pp. 189-206, 1998 and S.
Chong, “Second-Order Rate-Based Flow Control with Dynamic Queue Thresho
ld for High-Speed Wide-Area ATM Networks ”, preprint 1997 and A. Elwalid,
“Analysis of Adaptive Rate-Based Congestion Control for High-Speed Wid
e-Area Networks ”, Proc. IEEE ICC'95, pp. 1948-1953, 1995.

(3) Asymptotic Stability Here, stability of the ER allocation algorithm according to the embodiment of the present invention will be described. Generally, in the case of multiple nodes, the stability property of the equilibrium point given by <equation 19> can be investigated. However, due to the complexity of the dynamic properties coupled between the nodes, the analysis may be too complicated, and thus embodiments of the present invention may have an overall stability problem in the configuration of the coupled multi-nodes. No attempt is made to solve. Only the simulation in “(6) Results of simulation” will be investigated. On the other hand, Benhamhamed and Me
erkov indicates that there is an adjoining point of the equilibrium point where the dynamic characteristics of node-to-node are separated under a specific service system (“Feedback Control of Cons.
gestion in Packet Switching Networks: The Case of Multiple Congested Nod
es ”, international Journal of Communication Systems, vol. 10, no. 5, pp
.227-246, 1997). Furthermore, the local stability condition derived from the adjacent points is F
We will show by simulation that it is also applicable to the CFS service system. According to the above results, it is assumed that the neighbor point R having the above properties exists in the embodiment of the present invention. Considering the subset of the adjacent points R, the following is satisfied.

[0057]   a)

[Formula 39] , That is, the dynamic characteristics of the remaining nodes are in a steady state. b)

[Formula 40] And

[Formula 41] , That is, the local bottleneck VC transmits data at r _i (t−r _i ^b ),
The remote bottleneck VC transmits data on mim [b _is , p _i ].

C) The saturated nonlinear values in <Equation 7> and <Equation 8> are not activated. That is,
q (t) and r (t) are positive values. d) The | Q | estimation process is in a steady state. That is,

[Equation 42] Is a constant. At the points adjacent to the equilibrium point as described above, the dynamic characteristic equations <equation 6>, <equation 7>
> And <Equation 8> can be simplified as <Equation 20>, <Equation 21>, and <Equation 22>.

[0059]

[Equation 43] <Expression 23> is obtained by combining <Expression 20> and <Expression 21>.

[Equation 44]

The error function is defined as e (t) = q (t) −q _T. By combining the differentials of <Expression 22>, <Expression 23>, and the differential of <Expression 8>, a closed-loop equation like <Expression 24> is obtained.

[Equation 45] <Expression 24> is a second-order delay differential equation. The characteristic equation (c
The haracteristic equation) is given as <Equation 25>.

[0061]

[Equation 46] <Expression 25> has an infinite number of roots. For the asymptotic stability of the closed-loop equation <Equation 24>, all roots of the characteristic equation <Equation 25> should have a negative real part.
(R. Bellman and KL Cooke, “Differential-Difference Equations”, Acad
emic Press, New York, 1963 and G. Stepan, “Retarded Dynamical Systems: S
tability and Characteristic Functions ”, Longman Scientific & Technical,
1998, see). In order to find a necessary and sufficient condition for D (s) = 0, it can be obtained from Pontryagin's criterion by assuming a discrete delay with a rational ratio (R. Bellman and KL Cooke, “Differential -Difference Equati
ons ”, Academic Press, New York, 1963 and SJ Bhatt and CS Hsu,“ St
ability Criteria for Second-Order Dynamical Systems with Time Lag ”, Jou
rnal of Applied Mechanics, pp. 113-118, 1996,). For the general case of irrational continuous or discrete delays, the Stepan's crit
erion) provides a way to configure its necessary and sufficient conditions (“Retarded Dynamical
Systems: Stability and Characteristic Functions ”, Longman Scientific &
See Technical, 1989). However, if the above conditions are used in the explicit form (explici
t form), for cases with many different round trip time delays,
Very complicated.

Alternatively, if all round trip time delays are the same, then the necessary and sufficient conditions for asymptotic stability are derived.

[Equation 47] Then, the closed loop equation <Expression 24> becomes <Expression 26>.

[Equation 48] The new equation can be normalized such that the time delay τ is one.
With t = τξ, the new variable ξ turns <Expression 26> into <Expression 27>.

[Equation 49]

[0063] here,

[Equation 50] Is. The characteristic equation of <Expression 27> is <Expression 28>.

[Equation 51] The Pontryagin criterion can be used to search for necessary and sufficient conditions in which all roots of H (z) = 0 have a negative real part.

[0064]   The theory leads to the following results.

[Equation 52] Then, the closed loop equation <equation 26> has asymptotic stability only when <equation 29>.

[Equation 53] Here, ω ₁ is the only solution of U = ω sin ω in the interval (0, π / 2). <
The controller gains A and B can be set from Equation 29>. FIG. 7 illustrates stable regions for U and V according to embodiments of the invention.

(4) Principal Roots and Asymptotic Decay Rate Now, the determination of the rate at which a stable closed-loop system approaches steady state will be described. The solution of the normalized closed loop equation <Equation 27> can be expressed by a series such as <Equation 30>.

[Equation 54] Where p _n (ξ) is a suitable polynomial

[Equation 55] Is the root of the corresponding characteristic equation <Equation 28>. As the root with the largest real part, z ^* Consider the major roots displayed in. z^*= -Α ± jβ, α> 0, β> 0,
<Expression 31> is derived from <Expression 30>.

[0066]

[Equation 56] Here, C is a constant that depends on the initial condition of <Equation 27>, and x (ξ) ≈y (ξ) is

[Equation 57] Indicates asymptotically approaching 1. Also, <Expression 32> is derived from <Expression 30>.

[Equation 58]

[0067] here,

[Equation 59] Indicates the Euclidean norm, and c is a constant that depends on the initial condition of <Equation 27>. If the original variable t (= rξ) is replaced, <Expression 32> becomes <Expression 3>
It can be rewritten as 3>.

[Equation 60]

[0068]   Here, α / τ is the asymptotic decay rate at which the original system advances to the equilibrium point.
Be careful. The inverse α / τ of the asymptotic damping factor is
Is a constant, and the time is such that a small perturbation around the equilibrium point is^-1Due to
It is the time it takes to be reduced. Similarly, α is the asymptotic decay rate and α ^-1 Is the time constant of the normalized system.   FIG. 8 shows the asymptotic decay rate α for the functions U and V. Referring to FIG.
The asymptotic attenuation rate α has a maximum value of about 0.3 at (U, V) = (0, 6, 0, 1).
Is a concave function with respect to U and V that is. At α corresponding to the boundary of the stable region
The contour lines are as shown in FIG. Once α is determined for a given (U, V) pair
Then, z = −α + jβ is replaced by the characteristic equation <Equation 28>, and the real part and the imaginary part are replaced.
Β can be easily determined by calculating several parts.

(5) Discrete-Time ER Algorithm and | Q | Estimation Up to now, a continuous-time model of the maximum-minimum flow control problem has been described. However, in practice, the feedback information is relayed through the RM cell, so the sampled form is more useful than continuous time. Further, the RM
Since the cell itself has to complete the link bandwidth along the round trip time path,
Since the feedback information is not periodic, the arrival time interval of the feedback information RM cell is variable.

The ER allocation algorithm expressed by <Equation 8> and <Equation 9> is implemented in a switch in discrete time as follows. The common ER is <equation 34>
Is updated periodically with a cycle T.

[Equation 61] here,

[Equation 62] Indicates the length of the low-pass filtered queue, that is, the average queue length. In particular, in embodiments of the invention,

[Equation 63] Since we use a periodic averaging filter to find

[Equation 64] become.

Note that the ER update equation <Expression 34> corresponds to <Expression 9> when T → ∞. In contrast to the periodic calculation r [k + 1] of the common ER value r (t), the ER allocation per VC is performed periodically when RM cells corresponding to either the forward direction or the backward direction arrive. To be done. That is, when a VCi RM cell arrives at time t,
The switch calculates <Equation 35>.

[Equation 65] Then, the result is written in the RM cell. In <Formula 35>, r (t) is a common E
The most recent value r [k + 1] of the R value r [k] is shown and updated by the periodic background calculation according to an embodiment of the present invention. The value of mi can be used from the RM cells that have arrived or from the MCR table per VC maintained in the switch and updated at the arrival or departure of the RM cells. If it is decided to take the mi value from the arrived RM cell, the MCR table per VC and access to it is not needed anymore. Therefore, the operation per VC required by the ER allocation algorithm of the present invention is only addition as in <Expression 32>.

On the other hand, many theoretical systems have been proposed for estimating or tracking the number of local bottleneck VCs. As an example of the above theoretical system, MK Wong and F.
Bonomi, “A Novel Explicit Rate Congestion Control Algorithm”, it Proc.
IEEE GLOBELCOM'98 vol. 4, pp. 2432-2439, 1998 and L. Kalampoukas, A. Var
ma and KK Ramakrishnan, “An Efficient Allocation Algorithm for ATM N
etworks Providing Max-Min Fairness ”, Technical Report UCSC-CRL-95-29, C
omputer Engineering Dept., University of California, Santa Cruz, June 19
95, A. Charny, KK Ramakrishnan and A. Lauck, “Time Scale Analysis
and Scalability Issue for Explicit Rate Allocation in ATM Networks ”, IE
EE / ACM Trans. On Networking, vol. 4, pp. 569-581, 1996 and R. Jain et al.
, “ERICA Switch Algorithm: A Complete Description”, ATM Forum / 96-1172,
There is 1996. The respective theoretical systems have different implementation complexity. 1
The basic concept of the algorithms such as Su, which estimates the number of “ON” sources shared by two links (C. F, Su, G. de Veciana and J. Warlrand, “Explicit R
ATE Flow Control for ABR Services in ATM Networks, preprint, 1997) is attractive because it does not require per-VC calculations.

In an embodiment of the present invention, the algorithm is modified to estimate the number of local bottleneck VCs without performing per VC calculations. Suppose the jth RM cell arrives at the switch at switch time t ^j . According to the ABR specification, if the j-th RM cell is a VCi RM cell, the j-th RM cell transmits the a _i (t _j −τ _i ^f ) value of the CCR field and the mi value of the MCR field. . The switch synchronously monitors the arrival of RM cells over a fixed length interval of W seconds. V for local bottleneck for the first interval
The C number is estimated by <Equation 36>.

[0074]

[Equation 66] Where 1 {·} is an indicator function, CCR (t ^j ) is the value in the CCR field, MCR (t ^j ) is the value in the MCR field, and r (t ^j ) is the common at time t ^j . It is the latest value of ER. the moment the jth RM cell arrives,
If the CCR minus MCR is greater than or equal to the closest common ER value in the switch, the VC to which the j-th RM cell belongs is calculated as the local bottleneck VC. If not, it is treated as a remote bottleneck VC. δ is a limit value for preventing an underestimation of the number of local bottlenecks VC that are particularly close to a steady state. When the system approaches steady state, the CCR of the local bottleneck VC stays close to the sum of the MCR and the common ER, so in the absence of the limit δ, due to a small perturbation in the CCR. , The remote bottleneck VC may be erroneously calculated. However, by having the limit value, the underestimation can be effectively prevented. Through the simulation, the recommendation choice δ = 0.9 is detected. Also, the expected number of VC RM cell arrivals within W seconds, that is,

[Equation 67] Note that by summing the values of the indicator functions by, the sum of said values over the W second time interval provides an accurate estimate of the number of local bottleneck VCs. Based on the estimates as above for each time interval, the iterative estimate is calculated by <Equation 37> and <Equation 38> at the end of every time interval.

[0075]

[Equation 68] In <Expression 37>, λ is an average factor, int [a] represents the lowest positive integer equal to or larger than _a , and the saturation function sat _a [b] is defined as in <Expression 39>.

[Equation 69] For all t,

[Equation 70] Note that The actual number of local bottleneck VCs can be zero for some network loads, but intentionally put a lower bound on the estimate to avoid dividing by a value close to zero.

A future problem is how to choose the interval W and the mean factor λ. As the number of VCs sharing one link increases, or as the available bandwidth decreases for a given interval W, the arrival latency of VC's RM cells increases, causing the switches to vary widely. Get started. To solve the above problem, W can be adjusted by changing the number of VCs sharing the link and the available bandwidth, but it is not easy to implement. Alternatively, one can choose a large λ close to 1, while expecting the periodic averaging operation in <Equation 37> to effectively filter the | Q | I variability. Through the simulation, for a number of VCs sharing one link in a wide area and an available bandwidth, a stable and effective | Q | estimate is λ = 0.98 regardless of the selection of W. It turns out that

The inventors of the present invention further improve the | Q | estimation algorithm for improved stability. First, it is assumed that the controller gains A and 8 are selected as in <Equation 40>.

[Equation 71]

The (U, V) pair exists within the stable region shown in FIG. The | Q | estimate is accurate enough so that for all times

[Equation 72] The selection is made so that However, in general,

[Equation 73] Therefore, U ′ and V ′, which are the actual values or effective values of U and V that manage the normalized closed-loop equation <Equation 27>, are replaced by replacing <Equation 26> with <Equation 40>. Equation 41> is given.

[0079]

[Equation 74]

[Equation 75] If is smaller than 1, the point (U ', V') exists somewhere on the straight line connecting the point (U, V) and the origin (0, 0), and therefore exists within the stable region. . Conversely,

[Equation 76] When exceeds 1, the point (U ', V') moves upward along a straight line including the origin (0, 0) and the point (U, V), and consequently moves out of the stable region. . Simply put, an overestimation of | Q | does not affect the stability of the system and is acceptable because it only changes the asymptotic damping factor, but an underestimation of | Q | It destabilizes and should be avoided. This is because the limit value δ is required for <Expression 36>. However, δ does not solve all situations, especially the new VC
However, it is not possible to solve the problem more in the situation where the participant participates in a small initial cell rate (ICR). When the new VC participates in a small initial cell rate, the VC has a sufficiently small limit value δ, and therefore CCR (t ^j ) −MCR (t ^j ) is δ · in <Equation 36>.
It is very likely that it will be smaller than r (t ^j ), so it is very likely to be calculated as the remote bottleneck VC during the transient period. In such a situation, this does not matter if the new VC is actually a remote bottleneck VC, but if the new VC is actually a local bottleneck VC, the initial underestimation of | Q | It is possible that the system will be divergent. In order to solve the above problem, in the embodiment of the present invention, the | Q | estimation algorithm is improved as follows.

As soon as a new VC arrives at time t, the estimate is intentionally as if the V
It is updated by <equation 42> and <equation 43> as if C is a local bottleneck.

[Equation 77]

By doing the above, it is possible to avoid initial underestimation of | Q |

[Equation 78] Converges to | Q | from <Expression 42> and <Expression 43>. This is a highly desirable property for system stability, as mentioned above. The performance of the | Q | estimation proposed in connection with the ER allocation algorithm according to the embodiment of the present invention is proved by “(6) Simulation result” throughout the simulation. Next, consider the case where RR marking is applied in relation to the ER marking. The ER marking is asymptotically stable, and is maximum with MCR guarantee.
Whereas the RR marking plays a major role in achieving minimum fairness,
Plays a complementary role in limiting transient overshoot of queue length, and therefore
Minimize transient cell loss.

FIG. 9 shows a proposed design of queue length critical values for the RR and ER markings. In FIG. 9, q _T is the target queue length, and q _LT is the A
The minimum queue length threshold for the BR queue and q _HT is the maximum queue length threshold for the ABR queue. q _T is associated with ER marking and q _LT and q _HT are associated with RR marking. q _LT is a critical value for setting the NI bit of the NI field, and q _HT is a critical value for setting the CI bit of the CI field. If the queue length is greater than q _HT , the switch
The CI bit of the reverse RM cell is set to 1 to indicate queue length congestion. When the queue length is between q _LT and q _HT , the switch sets the NI bit of the reverse RM cell to 1 to prevent the source from increasing bandwidth. However, by setting q _T much smaller than q _LT , the queue length stays close to q _{T due to} the control of the ER marking, unless a queue spike due to sudden changes in the network load occurs. However, the RR marking is hardly activated. The linear increasing and exponential decay modes are initiated when the queue length exceeds q _{HT due} to load changes and are activated until the ER marking regains control of the queue length.

In “(6) Results of simulation”, the RR marking effectively limits the worst queue length in the transient period, and the ER marking re-executes the actual control to obtain the queue length. It is shown through simulation that we recover from the vibration mode to the asymptotically stable mode.

(6) Simulation Results Here, the simulation results are presented to confirm the above analysis and to experiment the excellent performance of the ER allocation algorithm according to the embodiment of the present invention.
The simulation model used was developed on the NIST ATM simulator platform. The two different network topologies are peer-to-peer (p
The eer-to-peer structure and the parking lot structure are considered. <Table 1
>, The recommended values for the design variables of the switching algorithm proposed by the present invention are summarized and used in the following simulation study.

[0085]

[Table 1] In FIG. 10, first, considering the peer-to-peer configuration, 20 ABRs are considered.
The VCs have the same route, and the capacity of all the links is set to 600 Mbps and the same. In order to implement a WAN environment, the distance between each of the sources s1 to s20 and the first switch SW1 is set to 1,000 km in the path. Assuming a signal propagation speed of 2.5 × 10 5 Km / sec and ignoring the queuing time, τmax is approximately 10 msecs. The VC model used in the simulation configuration is
As shown in Table 2 all sources are assumed to be persistent. PCR
, MCR, ICR, arrival time, and time are measured in Mbps.

[0086]

[Table 2]

Changing PCR values, MCR values, ICR values, and VC arrival and departure times in order to investigate the effects of PCR restricted sources, the difference between MCR and ICR, and call activity in network performance. Be careful. For the purpose of comparison, the theoretical fairness ratio that satisfies the MCR guarantee and the maximum-minimum fairness is calculated, and the result is shown in Table 2.
Fill in>. Referring to Table 2 and FIG. 10, it is observed that the fairness of each VC changes with time due to arrival and departure of other VCs. The source s1
The switches s11 to s2 are bottlenecks due to the switch SW1.
It can be seen that 0 becomes a bottleneck in the access due to its own PCR restriction. For example, the ABR sources s1 to s4 may be 0 to 0 for maximum-minimum fairness.
1 second is 36.1 Mbps, 1-2 seconds is 33.3 Mbps, 2-3 seconds is 3
At 0 Mbps, data should be transmitted at 32.5 Mbps from 3 to ∞. The present invention creates 32 data cells between two neighboring forward RM cells. That is, NRM = 32.

11A-11D show simulation results in a peer-to-peer configuration with only ER marking and no VBR background traffic. 12A-12D show simulation results in a peer-to-peer configuration with ER marking and VBR background traffic. 11A and 11B illustrate a source rate a _i of VC PCR respectively is 150Mbps (t), and PCR is 25 Mbps VC source transmission rate a _i of (t), the. It can be seen from FIGS. 11A and 11B that the actual source transmission rate exactly matches the theoretical fairness rate of Table 2. The transmission rates of the sources s1 to s4 and s10 are the same as the common ER value r (t) calculated by the switch SW1 because the MCR is 0 Mbps, and the transmission rates of the sources s5 to s9 are the same. Since the MCR is 10 Mbps, the source s is larger than the common ER value r (t).
11 and s20 are limited by their PCR regardless of their MCR. The initial transient operating characteristic is r (0) = 0 in both the switches SW1 and SW2.
Due to the initial condition In other words, although it takes time for the common ER value to jump to the operating point, it hardly occurs in the normal operating process. FIG. 11C
Indicates the queue length in the bottleneck node SW1. The merging of the source s20 at 1 second and the source s20 at 2 seconds leads to a sudden increase in the length of the queue, and the retention of the source s20 at 3 seconds leads to a sudden sudden decrease in the length of the queue. . However, the algorithm proposed in the embodiment of the present invention quickly re-stabilizes with q _T = 250 cells. FIG. 11D shows | Q | avg (t) which is the estimated number of local bottlenecks VC in the switch SW1. By computing the smallest positive integer estimate that is greater than or equal to the estimate according to <Equation 38>, the integer estimate | Q | can be obtained as follows. Where the symbol “[x” is

[Equation 79] The symbol “y]” indicates “y>”. As shown in FIG. 11D, the integer estimated value is 18 at [0, 1] seconds, 9 at [1, 2] seconds, and 1 at [2, ∞] seconds.
8 Here, except for the initial time interval [0, 1] seconds, the integer estimates are shown in Table 2 as 9 at [0, 2] seconds and 10 at [2, ∞] seconds |
It exactly matches the true value of Q |. The difference in the time interval [0, 1] seconds is
, The estimated value | Q | avg (t) is increased by one each time a new VC arrives. 1 for 0 seconds in a given simulation scenario
8VC arrives because the estimate | Q | avg (t) initially has a value of about 18.

For VBR background traffic, the results are shown in FIGS.
Like 2D, it does not change vertically. In other words, the macroscopic (time-averaged) motion characteristics are
Identical to the case with non-VBR traffic disturbance. VB
The R background traffic has a peak rate, an on-cycle length, and an off-cycle length of 10 Mbps, 20τmax seconds, and 20τ, respectively.
Produced by the determined on / off source, which is max seconds. From FIGS. 12A and 12B, the transmission rate of the source (excluding PCR limited sources) almost completely follows the on / off pattern of the VBR background traffic. VB
The on / off behavior of R background traffic results in iterative expansion and contraction of the length of the ABR queue. However, ER according to embodiments of the present invention
The allocation algorithm quickly recovers the queue length to the target value, as shown in FIG. 12C.

Now consider the case of a VC with multiple bottleneck nodes and different round trip time delays, considering the parking lot configuration shown in FIG. A 16ABRVC with different source positions is included, the performance of the link is based on the switch SW3.
The link between SW4 and SW4 and the switch SW4 and SW5 are 300M
It is set equally at 600 Mbps, except for bps. The VC model used for the simulation setup is summarized in Table 3 and assumes that all sources are persistent.

[0091]

[Table 3]   For comparison purposes, the maximum-minimum publicity that guarantees MCR for a given scenario.
The theoretical fairness that satisfies flatness was calculated, and the results are included in Table 3. Previous
In order to make the scenario more clear, the theoretical bottleneck position of each VC is shown in <Table 3
> Included. The bottleneck position is determined by each individual fair share.
The position. 14A to 14F have only the ER marking and the VBR back
The simulation result when no ground traffic is shown is shown. FIG. 14A
14B and FIG. 14B respectively show the source transmission rate a of the VC whose PCR is 150 Mbps. _i (t) and the source transmission rate a of the VC whose PCR is 25 Mbps_i(t) is shown.
FIG. 14C shows the queue length in the switch SW3, and FIG. 14D shows the switch S.
The queue length in W4 is shown in FIG.
The estimated value of the number of Torneck VCs | Q | avg (t) is shown in FIG.
The estimated value | Q | avg (t) of the number of local bottleneck VCs in FIG. Figure 1
From 4A and FIG. 14B, the actual source transmission rate in the steady state is the own round trip time.
Regardless of delay and bottleneck position, the prescribed theoretical fairness in Table
I know that it will match. The initial transient operation characteristic is r (0) = 0 in all switches.
This is due to the initial conditions of, but this does not occur during normal operation.
is there. In a given simulation scenario, two congestion nodes SW3 and
SW4 exists. As expected, the queue length at the congested node is shown in FIG.
As shown in 4C and FIG. 14D, the target value converges to 250 cells. Where the steady state
In the state, the estimated values at the switches SW3 and SW4 remain at about 9 and 2.
I understand that. This is consistent with the data in Table 3.

Finally, a case where RR marking is applied together with ER marking will be described. The queue length critical value for the RR marking is set at q _T = 250 cells, q _LT = 750 cells, and q _HT = 1000 cells. The choice of the queue length critical value is such that, under normal operating conditions, the ER marking plays a major role in ensuring asymptotic steady state maximum-minimum flow control, while the RR marking shows the queue length transient. It is expected to play an additional role in limiting overshooting. In an embodiment of the invention, the same parking lot configuration scenario with VBR background traffic is used. The VBR background traffic has a peak rate, an ON period length, and an OFF period length of 10 Mbps, 20τmax seconds, and 2 respectively.
Generated by the determined on / off source, which is 0τ max seconds. The results are shown in FIGS. 15A to 15F. 15A-15F show simulation results for a parking lot configuration with ER marking and VBR background traffic. 15A and 15B respectively show a PC.
The source transmission rate a _i (t) of VC with R of 150 Mbps and PCR of 25 Mbps
The source transmission rate a _i (t) of the VC, which is s, is shown. FIG. 15C shows the switch SW3.
15D shows the queue length in the switch SW4, FIG. 15E shows the estimated value of the number of local bottlenecks VC in the switch SW3 | Q | avg (t), and FIG. An estimated value | Q | avg (t) of the number of local bottleneck VCs in SW4 is shown. Compared to the previous case of FIGS. 14A-14F, the transient overshoot of ABR queue length in switch SW3 was actually reduced with a maximum reduction from 6000 cells to 3200 cells. Indeed, the gain comes at the cost of temporary oscillations in queue length and source transmission, as can be seen in Figures 15A, 15B, and 15C. Also, it can be seen that the effective control of the ER allocation algorithm according to the embodiment of the present invention restores the queue length behavior characteristic from the vibration mode to the asymptotic mode.

On the other hand, in the detailed description of the present invention, the specific embodiments have been described, but it goes without saying that various modifications can be made within the scope of the present invention. Therefore,
The scope of the present invention should not be limited to the above-described embodiments, but should be defined by the claims and their equivalents.

[0094]

【The invention's effect】

As described above, the ER allocation algorithm according to the embodiment of the present invention is
By (1) ensuring maximum link utilization and minimum cell loss, regardless of ABR closed-loop round-trip time delay, and (2) ensuring asymptotic stability of the ABR queue,
(3) Maximum-minimum fairness based on the ATM Forum standard proposal is guaranteed by minimizing the size requirement of the BR queue and (3) ensuring fair sharing of available bandwidth among users, and (4) using ABR. Respond quickly to changes in the communication network environment, such as changes in the number of people and changes in ABR bandwidth, and (5) All functions including EFCI marking, RR marking, and ER marking are proposed by the ATM Forum traffic management standard proposal. Provide as specified, (6) achieve high availability, low cell loss, and maximum-minimum fair share allocation due to the presence of asymptotic stabilizing operating points, and (7) on multiple timescales. , That is, high response to changes in network load at VBR and ABR VC cell level rate changes and at VBR and ABR VC cell level arrivals and departures And to achieve transient control performance, (8) to minimize the number of operations required to compute the algorithm, (9) VC different queuing, VC
There is the advantage of achieving a reduction in implementation complexity and scalability by actually eliminating VC-specific operations, including separate calculations and VC-specific table accesses.

[Brief description of drawings]

FIG. 1 is a diagram showing a network configuration for providing an ABR service.

FIG. 2 is a diagram showing fields in an RM cell.

FIG. 3 is a block diagram showing a switch fabric for an ABR service having an input / output port connected to an input / output card.

FIG. 4 is a detailed block diagram of the input / output card of FIG.

FIG. 5 is a block diagram of an ABR service engine according to an embodiment of the present invention.

FIG. 6 shows a network model with nodes of interest.

FIG. 7 is a graph showing stable regions for U and V according to an embodiment of the present invention.

FIG. 8 is a diagram showing asymptotic attenuation rates α in functions U and V.

FIG. 9 is a diagram showing queue length critical values for ER marking and RR marking.

FIG. 10 is a diagram showing a peer-to-peer configuration.

11 shows simulation results in a peer-to-peer configuration with only ER marking and no VBR background traffic, FIG.
Shows the source transmission rate a _i (t) of a VC with a PCR of 150 Mbps, FIG. 11B shows the source transmission rate a _i (t) of a VC with a PCR of 25 Mbps, and FIG. 11C shows the queue length in the switch SW1. 11D shows | Q | avg (t) which is the estimated number of local bottlenecks VC in the switch SW1.

FIG. 12 shows simulation results in a peer-to-peer configuration with ER marking and VBR background traffic, and FIG. 12A shows PCR.
VC source transmission rate a _i (t) of but a 150Mbps, FIG. 12B, the source rate of a VC PCR is 25 Mbps a _i a (t), FIG. 12C, the switch SW
12D shows the queue length for the switch SW1, and | Q | avg (t), which is the estimated number of local bottlenecks VC in the switch SW1.

FIG. 13 is a diagram showing a configuration of a parking lot.

FIG. 14 shows simulation results for a parking lot configuration with only ER marking and no VBR background traffic, and FIG.
4A shows the source transmission rate a _i (t) of a VC whose PCR is 150 Mbps as shown in FIG.
B is the source transmission rate a _i (t) of the VC whose PCR is 25 Mbps, FIG. 14C is the queue length of the switch SW3, FIG. 14D is the queue length of the switch SW4, and FIG. FIG. 14F shows an estimated value | Q | avg (t) of the number of local bottlenecks VC in the switch SW3, and FIG. 14F shows an estimated value | Q | avg (t) of the number of local bottlenecks VC in the switch SW4.

FIG. 15 shows simulation results for a parking lot configuration with ER marking and VBR background traffic, FIG.
The source transmission rate a _i (t) of a VC whose CR is 150 Mbps is shown in FIG.
The source transmission rate a _i (t) of a VC whose R is 25 Mbps, FIG. 15C shows the queue length in the switch SW3, FIG. 15D shows the queue length in the switch SW4, and FIG. 15E shows the queue length in the switch SW3. FIG. 15F shows the estimated value | Q | avg (t) of the number of local bottleneck VCs, and FIG. 15F shows the estimated value | Q | avg (t) of the number of local bottleneck VCs in the switch SW4.

FIG. 16 is an ER in an ABR service engine according to an embodiment of the present invention.
It is a flowchart which shows allocation control operation.

[Explanation of symbols]

  2 network   4, 6, 8, 10 nodes (or switches)   12 Source A   14 Source C   16 Source B   18 Source D   20 switch fabric   22 I / O port card   30 I / O buffer management unit   32 ABR Service Engine   34 Output interface   36 ABR queue   40 EFCI marking section   42 | Q | Estimator   44 ER calculator   46 Average queue calculator   48 Reverse direction RM cell writing unit   50 Source i   52 Source j   54 VCi   56 VCj

Claims (12)

[Claims] 1. A plurality of nodes each coupled to a plurality of sources for transmitting / receiving data, an explicit rate engine, and a source and a queue in relation to a current queue length and a target queue length. A fair flow control method in a packet switched network having a data queue for storing received data, the method comprising: generating a first signal and a second signal; and responding to the first signal at one of the sources. And periodically updating the explicit rate by the explicit rate engine, and when a forward direction cell arrives at one of the sources, a local bottleneck virtual based on the updated explicit rate from the explicit rate engine. Estimating the number of lines (| Q |); and, in response to the second signal, a difference between a current queue length of the node and a target queue length, and the estimation Has been the local bottleneck the number of virtual circuits and, (| | Q)
Determining a new explicit rate based on the current derivative of the queue length; and when a reverse cell arrival is detected at one of the sources, the detected reverse cell Writing a new explicit rate in the explicit rate field, and comprising :. 2. The method according to claim 1, further comprising the step of retrieving an MCR (Minimum Cell Rate) from the reverse cell when the arrival of the reverse cell is detected at one of the sources. Method. 3. In the process of writing the new explicit rate in the explicit rate field of the reverse cell, the explicit rate of the reverse cell and the sum of the newly determined explicit rate and the MCR of the reverse cell. , And writing the new explicit rate in the explicit rate field of the backward cell if the sum is less than the explicit rate retrieved from the backward cell. The method of claim 1. 4. The number of local bottleneck VCs (Virtual Circuits) is defined by a difference between CCR and MCR (Minimum Cell Rate) of the reverse cell and an update explicit rate received from the explicit rate engine. The method of claim 1, wherein the method is estimated by comparing. 5. The method of claim 4, wherein the number of local bottleneck VCs is estimated if the difference between the CCR and MCR is greater than or equal to the updated explicit rate. 6. The method of claim 1, wherein the forward and reverse cells comprise resource management cells. 7. The method of claim 1, wherein the explicit rate updated by the explicit rate engine is determined at predetermined time intervals. 8. A process of determining whether congestion occurs in one of the sources when a data cell is detected, and a first congestion signal or a second congestion signal when the congestion is detected. The method of claim 1, further comprising the step of transmitting, the second congestion signal indicating a more congested condition. 9. The method of claim 8, further comprising controlling an RR (Relative Rate) of the reverse cell, wherein the RR is defined by the first signal and the second signal. 10. The first signal is transmitted when the difference between the current queue length of the node and the target queue length exceeds a low predetermined threshold value, and the current queue length and target of the node are transmitted. 9. The method according to claim 8, characterized in that the second signal is transmitted when the difference between the queue length and the queue length exceeds a predetermined threshold value. 11. The number of bottleneck virtual circuits (| Q |) if the difference between the CCR and the MCR detected from the reverse cell is less than the updated explicit rate transmitted from the explicit rate engine. The method of claim 1, wherein is determined. 12. The new explicit rate is estimated from (((average queue length−previous average queue length) × first gain) ÷ (estimated) from the previous explicit rate in response to the second signal. Value | Q |) + ((average queue length-target queue length) x ((second gain x second signal section) / calculated value | Q |)) 4. The method according to claim 3,
The method described.