JP2003249953A - Packet forwarding control system, method for the same, program for the same, recording medium, and communication equipment - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a traffic congestion control system which automatically rationalizes control parameter values by a learning function following the change of traffic characteristics. <P>SOLUTION: A traffic controller which controls packet abandonment in order to maintain an average queue length of FIFO queues of a transmission buffer within a certain range and monitors the average queue length in order to control the congestion of the transmission buffer discriminates the transmission buffer control operation is adapted to current traffic characteristics or not on the basis of a preset adaptability discrimination condition. In the case that the transmission buffer control is not adapted, not only the extent of increase/decrease of the queue length is calculated by fuzzy inference in order to suppress quick increase or decrease of the queue length but also parameters of rules used in the fuzzy inference are corrected in a control parameter corrector by the learning function of a neural network so as to make the calculation proper when a maximum packet abandonment rate is adjusted to increase or decrease the proportion of abandonment of packets in the buffer. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a packet transfer control technique and, more particularly, to a technique suitable for efficiently controlling network congestion in response to dynamic fluctuations in traffic characteristics. .

[0002]

2. Description of the Related Art As conventional techniques relating to congestion control and packet transfer control techniques for networks such as the Internet that perform packet transfer, for example, "S. Floyd, V. Jac.
obson: Random early detection gateways for congest
ion avoidance.IEEE/ACM Transaction on Networking
vol.1, no.4, Aug-93 ”RED (Random EarlyD
etection). The congestion control technique and packet transfer control technique using this RED will be described below.

FIG. 14 is a block diagram showing a configuration example of a conventional system for performing congestion control and packet transfer control using RED, and FIG. 15 is an average queue length and packet discard rate referred to in the system of FIG. 16 is an explanatory diagram showing an example of the relationship with FIG. 16, and FIG. 16 is an explanatory diagram showing an outline of the operation of the RED queue manager in the system of FIG.

In FIG. 14, a gateway (Gatewa
When a packet from the source (Forwarder) arrives at the output port of (y) (Packet Incoming), the RED queue manager (RED-Queue Manager) determines the weighted average queue length q from the queue length Q of the transmission buffer (FIFO-Queue). Is calculated, and the average queue length q is compared with the predetermined "minimum threshold value min _{th of} average queue length" and "maximum threshold value max _th of average queue length" to perform the following operation in each case.

First, in the case of "q <min _th ", the arrived packet is stored in the queue of the transmission buffer (FIFO-Queue), and in the case of "min _th ≤q≤max _th ", the packet discard probability Pa is calculated. Then, the arriving packet is discarded with this discard probability Pa, and if "max _th <q", all the arriving packets are discarded.

That is, as shown in FIG. 16, the RED queue manager first calculates the average queue length q for each packet arrival. For example, when the queue is not empty, it is based on the formula of “q = (1−ω _q ) q + ω _q Q” (ω _q : a weighting coefficient of the low-pass filter for estimating q), and when the queue is empty, “q = The average queue length q is calculated based on the formula of (1-ω _q ) ^m q ”(m: elapsed time from when the queue was last emptied until now).

As a result of this calculation, "min _th ≤ q ≤ max _th "
In the case of, first, "Pb = P _max (q-min _th ) / (max _th
-Min _th ) "(P _max : maximum value of Pb)," Pb "is obtained, and then" Pb "is used to calculate" Pa = Pb / (1-
count × Pb) ”(count: the number of packets accommodated in the queue since the last discard), the packet discard rate Pa is determined, and the arriving packets are discarded at this packet discard rate Pa. From the above equation, “P _max ” (maximum value of Pb) is also the maximum value of Pa.

If "max _th <q", all arriving packets are discarded. Note that the packet discard rate Pa and the average queue length q obtained by calculation in FIG.
The relationship is as shown in.

By this packet transfer control operation, the average queue length q of the transmission buffer (FIFO-Queue) can be kept within a certain range, and buffer congestion can be avoided in advance.

However, as described above, in order to efficiently operate the queue control using the RED, it is necessary to set the control parameter to an appropriate value, but the tune-up technique has been theoretically established. Not difficult

Even when the tune-up succeeds and the average queue length q remains almost constant, the number of active flows increases or decreases on the same link and traffic is rapidly increased. If the characteristics fluctuate, the average queue length q cannot be maintained within the packet random discard section (min _{th to} max _th ) within a certain range, and desired throughput and connection delay cannot be obtained.

As a conventional technique for dealing with such a problem, for example, "S. Floyd, G. Gummadi, and S. Shenke
r. Adaptive RED: An Algorithm for Increasing the
Robustness of RED. Technical Report, to appear, 20
As described in "01", a technique has been proposed in which the maximum packet discard rate P _max is not kept constant, but the maximum packet discard rate P _max is made variable (operated) in accordance with a rapid increase / decrease in q. There is.

The operation policy of the maximum packet discard rate P _max is "if q <min _th , decrease P _max (= increase queue length, increase packet transfer amount)", "min _th ≤ If q ≦ max _th , P _max is kept constant, and if max _th <q, P _max is increased (= the queue length is reduced and the packet transfer amount is reduced) ”.

When the maximum packet discard rate P _max is reduced, as is apparent from the above equation, the slope of the graph shown in FIG. 15 becomes smaller, and when the maximum packet discard rate P _max is increased, the graph shown in FIG. The gradient of the graph shown in (1) becomes large, and the packet discard probability Pa for each average queue length q changes.

However, this technique does not solve the problem of how to determine the operation amount ΔP _max of the maximum packet discard rate P _max with respect to the change amount of the average queue length q in various traffic situations.

[0016]

The problem to be solved is that the conventional technique cannot appropriately adjust the maximum packet discard rate P _max in response to the fluctuation of traffic characteristics.

An object of the present invention is to solve these problems of the prior art, to enable appropriate congestion control even when a sudden change in traffic characteristics occurs, and to improve the reliability of the network.

[0018]

In order to achieve the above object, the present invention, for example, in a gateway on the Internet, sets an average queue length q of a FIFO queue of a transmission buffer within a certain range regardless of fluctuations in traffic characteristics. To control congestion of the transmission buffer by controlling packet discard in order to maintain
The average queue length q of the FO queue is monitored, and based on the monitoring result of the average queue length q, the control operation for the transmission buffer is adapted to the current traffic characteristic by referring to the pre-stored traffic characteristic adaptability determination condition. If it is determined to be non-adaptive, the maximum packet discard rate P _max is adjusted to increase or decrease the rate of discarding packets in the buffer, and the average queue length q is rapidly increased. Or suppress a sharp decrease. In order to suppress this, the control rule for the maximum packet discard rate P _max is described as a fuzzy relationship, and the manipulated variable ΔP _max is calculated by fuzzy inference. Further, the manipulated variable ΔP _max is corrected by the learning function of the neural network.

[0019]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing a configuration example of a packet transfer control system according to the present invention, FIG. 2 is a block diagram showing a configuration example of a traffic control device in FIG. 1, and FIG. FIG. 4 is a block diagram showing a configuration example of a control parameter correction device, FIG. 4 is an explanatory diagram showing an example of control conformity determination conditions stored in the conformity control condition storage unit in FIG. 2, and FIG. 5 is used in the traffic control device in FIG. FIG. 6 is an explanatory diagram showing a first configuration example of a fuzzy control rule, FIG. 6 is an explanatory diagram showing a second configuration example of a fuzzy control rule used in the traffic control device in FIG. 1, and FIG. 7 is a control in FIG. FIG. 8 is an explanatory diagram showing an example of a convergence condition stored in the parameter correction condition storage unit, and FIG.
10 is an explanatory view showing an example of a control rule generation condition stored in the control parameter correction condition storage unit in FIG. 9, FIG. 9 is an explanatory view showing a configuration / operation example of a neural network of the control parameter correction device in FIG. 1, and FIG. Explanatory diagram showing a specific operation example of the neural network in FIG.
11 is an explanatory diagram showing an example of the processing operation of the control parameter correction device in FIG. 1, FIG. 12 is an explanatory diagram showing an example of the processing operation of the maximum estimation error region selection section in FIG. 3, and FIG.
4 is an explanatory diagram showing an example of a processing operation of a control rule generation unit in FIG.

In FIG. 1, the packet transfer control system of this example is provided in a gateway (described as “Gateway” in the figure) 1 as a communication device for transferring packets on the Internet or the like. , CP
It has a computer configuration including a U (Central Processing Unit), a main memory, a display device, an input device, an external storage device, and the like, and a CD-ROM via an optical disk drive device and the like.
After the programs and data recorded in the storage medium such as the above are installed in the external storage device, the functions of each processing unit according to the present invention are executed by reading the external storage device into the main memory and processing them by the CPU.

That is, in the gateway 1, a transmission buffer (“FIFO-QUEU” in FIG.
2) and the average queue length in this transmission buffer 2, and according to the packet discard rate Pa preset corresponding to this average queue length, the sender (Forward
RED queue manager (“RED-Queue Ma” in the figure) that controls the discard of packets (Packet Incoming) received from r).
3) is provided to control the transfer amount of packets.

Further, in the gateway 1 of this example, the packet discard control by the RED queue manager 3 is controlled to control the average queue length in the transmission buffer 2 to prevent congestion in the transmission buffer 2. Adaptive adjustment type traffic congestion system 4
Is provided.

Further, the adaptive adjustment type traffic congestion system 4 includes a fuzzy inference unit ("Fuzzy Control
ller) 41a and a traffic control device 41 and a neural network (“Neural Network” in the figure).
(described as “rk”) 42a.

The adaptive adjustment type traffic congestion system 4 and the RED queue manager 3 constitute the packet transfer control system according to the present invention, and the average of the FIFO queues of the transmission buffer 2 is averaged regardless of the fluctuation of the traffic characteristics on the network. In order to keep the queue length q within a certain range, packet discard is controlled and congestion of the transmission buffer 3 is controlled.

That is, the RED queue manager 3 monitors the average queue length q of the FIFO queue of the transmission buffer 2, and based on the monitoring result of this average queue length q,
In the adaptive adjustment type traffic congestion system 4, it is judged whether or not the control operation for the transmission buffer 3 is adapted to the current traffic characteristics by referring to the pre-stored traffic characteristic adaptability judgment condition, and it is judged as non-adaptive. In this case, the packet discard rate (here, the maximum packet discard rate P _max ) is adjusted to increase or decrease the rate of discarding the packets in the transmission buffer 3, and the average queue length q is rapidly increased or abruptly increased. Suppress the decrease.

In order to suppress this, in the traffic control device 41, the control rule of the maximum packet discard rate P _max is described as a fuzzy relationship, and the operation amount ΔP _max is calculated by the fuzzy inference unit 41a. Further, this manipulated variable ΔP _{max is set} to the control parameter correction device 4
In the second neural network 42a, the learning function is used for correction.

Specifically, the RED queue manager 3
Compares the calculated average queue length q with a lower limit value α and an upper limit value β of a preset target range, and if the average queue length q is smaller than the lower limit value α, the packet discard rate (maximum packet discard rate P _max ) is calculated. If the average queue length is smaller than the upper limit value β, the packet discard rate (maximum packet discard rate P _max ) is increased, but the packet discard rate by this RED queue manager 3 (maximum packet discard rate P _max )
The increase / decrease amount ΔP _{max of} is calculated by fuzzy inference in the fuzzy inference unit 41a of the traffic control device 41.

Further, the control parameters of the rule used in the fuzzy inference in the fuzzy inference unit 41a of the traffic control device 41 are converted into the control parameter correction device 42.
In the neural network 42a, the weight is calculated by the learning function as the weight of the neural network.

A traffic control device 41 having a fuzzy inference section 41a and a neural network 4
The control parameter correction device 42 including 2a has a configuration as shown in FIG. 2 and FIG. 3, respectively. Hereinafter, referring to FIG. 2 and FIG. 3, the traffic control device 41 and the control parameter correction device 42 will be described. Details will be described.

In the traffic control device 41 of FIG. 2, 411 is an adaptive control condition input / output unit, 412 is an adaptive control condition storage unit, 413 is a control parameter input / output unit, 414.
Is a monitoring data receiving unit, 415 is a control suitability determining unit, 416
Is an input value calculation unit, 417 is a control analysis unit, 418 is an output value calculation unit, and 419 is a control value transmission unit.

With this configuration, the traffic control device 41 monitors the average queue length q of the FIFO queues of the transmission buffer 2, and the average queue length q is a predetermined threshold value ma.
When x _th is exceeded, the traffic characteristics fluctuate and the average traffic volume increases, so that it is determined that the control operation is incompatible with the current traffic characteristics, and the maximum packet discard rate P _{max is set} to the manipulated variable (ΔP _max ) To increase the number of dropped packets.

On the contrary, the average queue length q is the threshold value min _th.
If it is less than the above, the traffic characteristic fluctuates and the average traffic volume decreases, so that it is determined that the control operation is incompatible with the current traffic characteristic, and the maximum packet discard rate P _max is equal to the operation amount (ΔP _max ). Reduce it to reduce the number of discarded packets. However, when the average queue length q remains between the two threshold values min _{th to} max _th , it is determined that the control operation is suitable for the current traffic characteristics, and the maximum packet discard rate P _max remains constant. To do.

In the control parameter correction device 42 of FIG. 3, 421 is a control parameter input unit, 422 is a control parameter correction condition input / output unit, 423 is a control parameter correction condition storage unit, 424 is a learning data input / output unit, and 425 is learning. A data storage unit, 426 is an output estimated value calculation unit, 427 is a convergence condition determination unit, 428 is a control rule generation condition determination unit,
429 is a control parameter update processing unit, 4210 is a maximum estimation error region selection unit, 4211 is a control rule generation unit, 42
Reference numeral 12 is a control parameter output unit.

With such a configuration, the control parameter correction device 42 has the target value r (mi
error e between n _th ≤r ≤max _th ) and the actual average queue length q
(= R−q) and error change Δe (= e _t −e _t−1 ).
Is input, and the fuzzy inference unit 4 of the traffic control device 41 is used by using the neural network 42a which outputs the operation amount ΔP _max of the maximum packet discard rate P _max.
In 1a, the control parameter of the rule used when calculating the manipulated variable (ΔP _max ) is corrected.

That is, when any input / output data is given as the learning data, the error between the output estimated from the input data and the actual output is fed back to update the connection weighting parameter value on the neural network 42a. To go.

The updating process of the connection weighting parameter value on the neural network 42a is sequentially executed for each learning data, and is continued until the output error becomes equal to or less than a certain fixed value (δ ₁ ). Thus, we determined the average appropriate operation amount according to the amount of change in queue length q (the decrease amount [Delta] P _max for the maximum packet loss ratio P _max) in the transmission buffer 2.

Hereinafter, such a traffic control device 4 will be described.
1 and the operation of each processing unit of the control parameter correction device 42 will be described.

First, with reference to FIGS. 4 to 8 together with FIGS. 2 and 3, based on a setting instruction operation from the system terminal 43 in FIG.
(5) will be described.

(1) The setting of the control suitability determination condition (see reference numerals (a) and (b) in FIG. 2) will be described.

By the setting instruction operation from the system terminal 43 in FIG. 2, the “control conformity determination condition” is input to the traffic control device 41 (a), and the conformity control condition storage unit is input via the adaptive control condition input / output unit 411. It is stored in 412 and set (b).

The "control conformity determination condition" means that the control conformity determination unit 415 uses the RED sequence manager 3
Shows the condition for judging whether the control operation of (1) can meet the current traffic characteristics and keep the average queue length of the FIFO queue of the transmission buffer 2 in a balanced state. Show.

"Control conformity determination condition 41" in the example shown in FIG.
Then, the average queue length q of the FIFO queue of the transmission buffer 2 is set to a preset target section α to β (min _th ≦
If α ≦ β ≦ max _th ) (α ≦ q ≦ β), it is determined that the control operation of the RED sequence manager 3 is “adapted” to the current traffic characteristics, and the preset target section α˜ If it is outside β (q <α or β <
q), the control operation of the RED sequence manager 3 is set to determine that it is “non-conforming” to the current traffic characteristics.

(2) Next, the setting of the target value r of the average queue length q of the FIFO queue of the transmission buffer 2 (see symbols (c) to (e) in FIG. 2) will be described.

Based on the setting instruction operation from the system terminal 43, the target value r is input to the traffic control device 41 (c), stored in the adaptive control condition storage unit 412 via the adaptive control condition input / output unit 411, and set. (D). At the same time, the target value r is also set in the output inference value calculation unit 426 of the control parameter correction device 42 via the control parameter input / output unit 413 and the control parameter input unit 421 (e).

Here, the "target value r" is an index when controlling so that the average queue length q of the FIFO queue of the transmission buffer is maintained at a constant length, and in the input calculation unit 416,
The reference value when calculating the input value to the control analysis unit 417 is shown.

(3) Next, the setting of the control rule (see symbols (f) to (h) in FIG. 2) will be described.

Based on the setting instruction operation from the system terminal 43, the control rule is input to the traffic control device 41 (f) and stored in the adaptive control condition storage unit 412 via the adaptive control condition input / output unit 411. It is set (g). In addition, the initial value of the control parameter to be parameter-adjusted by the learning function of the neural network 42a is output through the control parameter input / output unit 413 and the control parameter input unit 421, and the output inference value calculation unit 426 of the control parameter correction device 42 is executed. Is set to (h).

Here, the "control rule" is a fuzzy inference unit 41a of the traffic control device 41 when the control analysis unit 417 determines that the control conformity determination result is non-conformance.
In, we present a control rule for determining the maximum packet loss rate according to changes in traffic characteristics.

In the following, this control rule will be fuzzy.
y) Average queue length q and eye
An error e (= r−q) from the standard value r and a change Δe (=
e_t-E_t-1(T: time variable)) input (x₁,
x_Two), And the maximum packet discard rate P _maxManipulated variable ΔP_max
Is output (y).

Further, the membership functions (A _j1 , A _k2 ) in the antecedent part for each input have three types (j
= 1 to 3, k = 1 to 3)), and the consequent part is a constant in order to use the simplified fuzzy inference. Figure 5 shows the contents of the fuzzy control rules that can be described by these.
And illustrated in FIG.

That is, the fuzzy shown in FIG.
The control rule 51 is in the form of “if to then”, the input is “x ₁ (= e), x ₂ (= Δe)”, and the output is “y (= Δ”).
P _max ) ", the i-th fuzzy control rule (R
ⁱ ) is the antecedent part of “if x ₁ is A _j1 , x ₂ is A _{k 2} ” and “th x
en y = ω _bi ”consequent part (i = 1 ... 9, j = 1 ... 3
k = 1 ... 3).

The membership functions (A _j1 , A _k2 ) in the antecedent part have sigmoid function f as an internal function, and each have three types as initial values, and the consequent of the fuzzy control rule 51. The section is usually "y = c
₁ x ₁ + c ₂ x ₂ + ω _b ”(c ₁ , c ₂ , ω _b : constant), but since the simplified fuzzy inference is used in this example,
“Y = ω _b ” is used as “c ₁ = c ₂ = 0”. Note that ω _b in the consequent part, the slope ω _g in the membership function, and the center ω _c are control parameters of the fuzzy control rule 51.

As shown concretely in FIG. 6, in the fuzzy control rules 51 and 61 of this example, each input "x ₁ (=
e), x ₂ (= Δe) ”,
Output (rule) is obtained.

For example, if the input “x ₁ ” is “A ₁₁ ”, then “q>
r ”, if the input“ x ₂ ”is“ A ₁₂ ”and“ Δe> 0 ”, the rule (y = manipulation amount ΔP _max ) is significantly increased” is output, and the input “x ₁ ” is “A ₁₁ , "Q>r",
If the input “x ₂ ” is “A ₃₂ ” and “Δe <0”, the rule (y = increase the manipulated variable ΔP _max ) is increased, and the input “x ₁ ” is “A ₃₁ ”. in "q <r", the input "x _2" is "Δe in" a ₁₂ "> has a content if 0", the rule decreases (y = operation amount [Delta] P _max) significantly "is output .

(4) Next, the setting of the convergence condition (see symbols (i) and (j) in FIG. 3) will be described.

Based on the setting instruction operation from the system terminal 43, the convergence condition in the neural network 42a is input to the control parameter correction device 42 (i), and the control parameter correction condition input / output unit 422 is used to store the control parameter correction condition. It is stored in the unit 423 and set (j).

Here, the "convergence condition" indicates a condition for ending the control parameter update processing routing by the neural network 42a based on the learning data, and is exemplified in FIG.

Under the convergence condition 71 shown in FIG. 7, the output y * estimated from the input data “x ₁ ” and “x ₂ ” and the actual output y
^{"(Y-y *) 2/} 2 " the error E with a, if the output error E becomes a preset threshold ([delta] ₁₎ below ( ^{"E = (y-y *)} 2/2 ≦ δ ₁ ″), the control parameter updating process by this learning is terminated.

(5) Next, the setting of the control rule generation condition (see the reference signs (k) and (l) in FIG. 3) will be described.

Based on the setting instruction operation from the system terminal 43, the control rule generation condition is input to the control parameter correction device 42 of FIG. 3 (k), and the control parameter correction condition is input via the control parameter correction condition input / output unit 422. It is stored in the storage unit 423 and set (l).

Here, the "control rule generating condition" means a trigger condition for automatically generating a control rule when the estimation error does not decrease even if the control parameter is adjusted by the neural network 42a based on the learning data. FIG. 8 illustrates the contents of the control rule generation condition.

In the control rule generating condition 81 shown in FIG.
Control parameters are learned by learning with the neural network 42a.
The estimation error E does not decrease even if the meter is adjusted.
In other words, that is, the variation ΔE (= E of the estimation error E_p−
E_p-1... p: learning count) is a preset threshold value
(Δ_Two) When it becomes below (“ΔE = E_p-E_p-1
≤ δ _Two))), Content to automatically generate new control rules
Has become.

Similar to the setting / registration of each condition in (1) to (5) above, the learning data used in the neural network 42a is input to the control parameter correcting device 42 based on the setting instruction operation from the system terminal 43. ((M) in FIG. 3), and is stored and set in the learning data storage unit 425 via the learning data input / output unit 424 ((n) in FIG. 3).

The correction operation of the control parameter based on the learning data set in this way will be described below (see symbols (o) to (x) in FIG. 3).

The control parameter correction device 42 in FIG.
Output estimation value calculation unit 426 of learning data storage unit 425
The learning data stored in is read for each set (o), and x ₁ (= e) and x ₂ (= Δ of the data are read.
The estimated output value y * (= ΔP _max ) is calculated using e) as an input (p).

This maximum packet discard rate P_maxManipulated variable Δ
P_maxCalculation of the estimated value of
It is performed on the basis of gorhythm. That is, as shown in FIG.
Antecedent parts (A) to (D) of the neural network configuration
And the consequent parts (E) and (F), as shown in FIG.
Output (manipulation amount y = ΔP)
_maxPerforms processing related to the inference value of.

For example, first, in the antecedent section (A), x
_From the bias unit that inputs ₁ (= e) and x ₂ (= Δe) and outputs x ₁ (= e) and x ₂ (= Δe) and a constant value “1” in the antecedent part (B). based on the weighting to be added with respect to the output _{(ω c11} ~ω cj2), to calculate the input value ^{O B} of the sigmoid function ^{f (x) (O B =} ω g (x j + ω c), or, O ^B = -ω
_g (x _j + ω _c )).

The example shown in FIG. 10, that is, "Rule2: if
x ₁ is A ₂₁ , x ₂ is A ₂₂ theny = ω _b2 ”,
"I ₂ = ω _g21 (x _j + ω _c21 )", "-i ₂ = ω
_g21 (x _j + ω _c21 ) ”is calculated.

Next, in the antecedent section (C), sigmoid
Output value O of function f (x)^CAnd calculate (O^C= 1 ÷ (1+
exp [-ω_g(X_j+ Ω_c)]) Or O^C= 1
÷ (1 + exp [ω_g(X_j+ Ω_c)])),further,
In the antecedent part (D), the membership function A (x)
Output value O^DCalculate (O^D= F (x) or O ^D
= F (x) -f (-x)).

In the example shown in FIG. 10, "f (i ₂ ) ... (x
₁ ≤ ω _c ) "and" f (i ₂ ) -f (-i ₂ ) ... (ω _c <
x ₁ ) ”is calculated.

Then, in the consequent part (E), fuzzy
Goodness of fit of the antecedent part of the control rule (of each membership function
Calculate the grade product) μ (μ_i= A_i1(X₁) A_j2
(X _Two)), Suitable for the antecedent part obtained from all units in layer (E)
The value M standardized by the sum of the degrees is output (M_i= Μ_i÷ Σ
μ_k), The consequent part (F), the standard value of the conformance of the antecedent part
(M) and consequent constant (coupling weight) ω_bCompute the product of all
The sum of units is set as the output inference value y * (y * = ΣM ·
ω_b).

In the example shown in FIG. 10, the consequent part (E)
, "Μ_Two= A₂₁(X₁) × A_{Two Two}(X_Two) ”
The total of the antecedent conformity obtained in all units in the (E) layer.
Value normalized by sum "M_Two= Μ_Two÷ Σμ_iOutput M,
In the consequent part (F), with the standard value (M) of the conformity of the antecedent part
Consequent part constant (coupling weight) ω_bCalculate the product of_Two= M _Two
・ Ω_b, The sum of all units “output inference value y * = y₁+
y_Two+ ... + y₉Is obtained.

The output estimated value calculation unit 426 operates in this way.
Notify the convergence condition determination unit 427 of the estimated value y * calculated by
(Q). The convergence condition determination unit 427 starts learning.
Before reading from the control parameter correction condition storage unit 423
Based on the set convergence condition (r), the estimated value notified
The error E between y * and the output y of the learning data is a constant threshold value δ. ₁
Check if it is below (s).

If the result of this determination is that the convergence condition has not been cleared, the convergence condition determination unit 427 notifies the estimation error E to the control rule generation condition determination unit 428 (t). In the control rule generation condition determination unit 428, based on the control rule generation condition read from the control parameter correction condition storage unit 423 before starting the learning process (u), the notified estimation error E _p (p: learning frequency) And the previous estimated error E _p-1
It is checked whether or not the difference ΔE of (1) becomes equal to or less than a certain threshold value δ ₂ (v). When p = 1, the estimation error is stored and the determination process is not performed.

As a result of this judgment, when ΔE> δ ₂ , the control rule generation condition judging unit 428 judges that the control parameter correction process by learning is necessary with the current control rule, and the estimation error E _The control parameter update processing unit 429 is notified of _p (w).

In the control parameter update processing unit 429, error backpropagation (“BackPropagatio” is performed as a learning algorithm.
n ”) method, the estimated error E _p notified is used to control the slope ω _g and the center ω _c of the sigmoid function which is the internal function of the membership function of the antecedent part which is the control parameter, and the constant ω of the antecedent part. _The value is fed back to _b to update the parameter value (x). The processing operation for correcting such control parameters is performed as shown in FIG.

That is, FIG. 11A and FIG.
As shown in (b), error backpropagation (“BackPropagatio
n ”) method, input / output data d _p = (x _1p ,
_{x 2p, y p) (P} = 1 ... N: When the number of data) is given (step 1101), the input _{data, x 1p,}
The output y _p estimated from x _2p and (step 110
2) By feeding back the error from the actual output y, the values of the parameter ω _g and the center ω _{c of} the fuzzy controller model and the constant ω _b of the consequent part are updated (step 11).
03-1105).

Then, the data is sequentially updated and output.
Error E_pIs a preset threshold (δ ₁) When
If (step 1103), this learning procedure is ended.

The control parameter update processing unit 429 stores the parameter update value corrected in this way in the control parameter correction condition storage unit 423 (y).

Then, the output estimated value calculation unit 426 sets the updated control parameter from the control parameter correction condition storage unit 423, and then reads the next learning data from the learning data storage unit 425 to obtain a new output estimated value. calculate.

As described above, the control parameter correction process by learning is repeated until the estimation error satisfies the convergence condition (δ ₁ ).

However, when the estimation error E does not decrease even if the parameter adjustment is performed by such learning, that is, step 1 in the flow shown in FIG. 11B.
In the processing of 104, the change amount ΔE of the estimation error E = E _p −E
_{When p-1} becomes a certain threshold value δ ₂ or less, the control rule generation condition determination unit 428 in FIG. 3 determines that the control parameter correction process by learning is insufficient with the current control rule. Then, a new control rule is automatically generated.

The processing operation for automatically generating a new control rule in such a case will be described below (see reference numerals (z) to (h ² ) in FIG. 3).

The control rule generation condition determination unit 428 instructs the maximum estimation error selection unit 4210 to select the control rule generation region in order to automatically generate a new control rule (z).

Upon receiving this instruction, the maximum estimation error selection unit 42
The operation of selecting the control rule generation area by 10 is as follows.

In the maximum estimation error selection unit 4210, x is calculated by the membership function of the antecedent part of the fuzzy control rule.
From among all the _{regions 1} -x ₂ plane is divided, selects the largest area of juvenile constant error than the other regions as a region for generating a new control rule (b ^2). This operation will be specifically described with reference to the example of FIG.

In FIG. 12, the input variable (x₁, X_Two)
Center of two adjacent membership functions ω_c
The region divided by_jAnd For example input x₁
Of A ₁₁Center of_c11And A₂₁Center of_c21By
Area S divided by₁Is the fuzzy control rule shown in FIG.
, That is, 1, 3, 6, in FIG.
It is a rectangle surrounded by 4.

In addition to this, the region divided by the membership functions of x ₁ and x ₂ is S ₂ = (4, 6, 9,
7), S ₃ (1, 2, 8, 7), S ₄ = (2, 3, 9,
8).

Next, for each region S _j , the average estimation error D _{s j} of all learning data contained therein is calculated, and the region S _max having the maximum value among the average estimation errors D _sj is controlled. The control rule is selected as an area to be newly set in order to improve the accuracy of

The area 2 thus selected is constructed.
The maximum estimation error area selection unit 4210 notifies the control rule generation unit 4211 of the center ω _c of one membership function (c ² ). Control rule generation unit 421 that received the notification
1 generates a new control rule as follows.

That is, the control rule generation unit 4211 newly adds a membership function having a new center at the midpoint between the centers of the two membership functions forming the region selected by the maximum estimation error region selection unit 4210. Generate (d ² ).

FIG. 13 shows an example of the generation of the membership function when the area S ₁ of FIG. 12 is selected, and FIG.
The three "black circle points (12, 11, 10)" in 3) indicate newly generated rules. That is, when a new membership function A ₄₁ (x ₁ ) is generated in the area S ₁ , new rules 10, 11, 12 are generated.

As shown in FIG. 13B, when the new membership function A ₄₁ (x ₁ ) of the antecedent part is generated (see FIG. 3).
Code ^(e 2)) in the center omega _c41 is membership function _A 11 _{(x 1)} of the center omega _c11 and membership functions _A 21 _{(x 1)} center omega _{c 21} midpoint of the (omega _c41 =
(Ω _c11 + ω _c21 ) / 2), and the slope ω _g41 is the membership function A ₁₁ (x ₁ ) and the membership function A _21.
Each intersection with (x ₁ ) is set to 0.5.

Further, the generation of the real number of the consequent part (symbol (f ² , g ² ) in FIG. 3) is promoted by the pre-generation rule for the center of the new membership function so that the estimation error does not increase or decrease before and after the generation. _{Let the} output be a new real value of the _consequent part (ω _b1k = y _old (ω _{c4 1} + ω _ck2 ) ... (k =
0, 1, 2)).

In this way, the constant of the consequent part of the newly generated rule does not increase the estimation error before and after the generation of the rule, so that the central value of the membership function of the antecedent part of the new rule is updated. The new rule R ^1k is output as “R ^1k : if x
₁ is A ₄₁ , x ₂ is A _k2 then y = ω _{b1 k} ”.

The control rule generator 4211 produces the raw data in this way.
New rule ("R^1k: If x ₁ is A₄₁, x_Two i
s A_k2 then y = ω_b1k)) Control parameter correction
Stored in the case storage unit 423 (h^Two).

Hereinafter, reference numerals (i ² ),
As shown in (j ² ), the control parameter adjustment completion process by learning is performed.

That is, the convergence condition determination unit 427 determines that the estimated error between the estimated value y * calculated by the output estimated value calculation unit 426 and the output of the learning data is equal to or less than a certain threshold value δ ₁ .
When it is determined that the convergence condition is satisfied, the control parameter correction process by learning is terminated, and the output inference value calculation unit 42 is executed.
6 is notified of the end of updating (i ² ).

The output inference value calculation unit 426 that has received the notification
The control rule output unit 421 sets the control rule at the end point.
2 and the control parameter input / output unit 413, and stores it in the adaptive control condition storage unit 412 of the traffic control device 41 (j ² ).

Next, the operation of monitoring the average queue length and the operation of determining the control suitability by the traffic control device 41 having the configuration shown in FIG. 2 (refer to symbols (k ² ) to (m ² ) in FIG. 2),
And adaptive control operation (reference numeral (n ² ) to FIG.
(See (w ² )).

The traffic control device 41 receives the average queue length q of the FIFO queue of the transmission buffer 2 calculated by the RED queue manager 3 every time a packet arrives via the monitoring data receiving unit 414, and judges the control suitability. The unit 415 monitors (l ² ) and determines based on the (k ² ) control conformity determination condition read from the conformity control condition storage unit 412 whether the active queue control operation by the current RED is adapted to the traffic characteristic. Do (m ² ).

When the result of this determination is "non-conforming", the conformity determining unit 415 notifies the input value calculating unit 416 of the non-conformance occurrence event (o ² ), and at the same time sets the average queue length to the input value. And hand it over to calculate

Upon receiving the notification, the input value calculation unit 416 calculates the error e and the error change amount Δe based on the received average queue length q and the (n ² ) target value r read from the adaptive control condition storage unit 412. (P ² ), and inputs these to the next control analysis unit 417 (r ² ).

In the control analysis unit 417, based on the (q ² ) control rule read from the adaptive control condition storage unit 412,
The operation amount ΔP _max of the maximum packet discard rate P _max to be set next is determined according to the input error e and the error change Δe (s ² ). After the determination, the control analysis unit 417 inputs the determined operation amount ΔP _max to the output value calculation unit 418 (t
² ).

In the output value calculation unit 418, the input operation amount ΔP _max and the maximum packet discard rate P _max before the control suitability determination are made.
Then, the maximum packet discard rate P _max adapted to the current traffic characteristics is calculated (u ² ). Then, the calculated maximum packet discard rate P _max is reset in the RED queue manager 3 via the control value transmission unit 419 (v ² ).

The control suitability determination unit 415 repeats suitability adjustment of the maximum packet discard rate P _max for the current traffic characteristics until the average queue length satisfies the suitability determination condition.

When the control conformity determination unit 415 determines that the average queue length satisfies the conformity determination condition, it is determined that the active queue control operation by the RED queue manager 3 is “conforming” to the current traffic characteristics, and the conformance adjustment is performed. The control analysis unit 417 is notified of the completion report (w ² ).

As described above with reference to FIGS. 1 to 13, in this example, the average queue length q of the transmission buffer 2 in which the packets to be transferred are accumulated is calculated and preset in correspondence with this average queue length q. The received packet is discarded according to the determined packet discard rate, and when the packet transfer amount is controlled, the obtained average queue length q is compared with the preset lower limit value α and upper limit value β of the target range. , If the average queue length q is smaller than the lower limit value α, the packet drop rate is reduced, and if the average queue length q is larger than the upper limit value β, the packet drop rate is increased. Along with the calculation, the control parameter of the rule used in this fuzzy inference is obtained as the weight of the neural network by the learning function.

More specifically, in the gateway 1 on the Internet, packet discard is controlled in order to keep the average queue length q of the FIFO queue of the transmission buffer 2 within a certain range, regardless of fluctuations in traffic characteristics, and the transmission buffer is controlled. FIF of the transmission buffer 2 when controlling the congestion of
In the traffic control device 41 that monitors the average queue length q of the O queue, the traffic characteristic adaptability determination condition preset by the system terminal 43 is stored in the adaptability control condition storage unit 412, and based on this adaptability determination condition. , It is determined whether the control operation for the transmission buffer 2 is adapted to the current traffic characteristics, and the result of the determination is
When it is determined to be non-adaptive, the maximum packet discard rate P _max is adjusted to increase or decrease the rate of discarding the packets in the transmission buffer 2 in order to suppress a rapid increase or a rapid decrease in the queue length. , Maximum packet discard rate P
A rule for controlling _max is described as a fuzzy relation, and the operation amount Δ of the maximum packet discard rate P _max is calculated by fuzzy inference.
Calculate P _max . In addition, the control parameter correction device 42
In the above, the control parameter correction condition storage unit 4 stores the control parameter correction conditions preset by the system terminal 43.
Stored in 23, when the judgment result of the non-adaptive, in order to appropriately adjust the maximum packet loss rate P _{ma x} according to the amount of change in the average queue length q, the appropriate operation amount [Delta] P _max of the neural network training Modify according to the function.

As described above, in this example, for the adaptive adjustment of the operation amount ΔP _max of the maximum packet discard rate P _max in the RED queue manager 3, the control rule is fuzzy modeled to perform the fuzzy inference, and the fuzzy inference is performed. By adjusting the control parameter of the rule used in the above with the learning function as the weight of the neural network, the maximum packet discard in the RED queue manager 3 according to the fluctuation of the traffic characteristics can be performed without using the theoretical analysis model. The adaptive adjustment of the operation amount ΔP _max of the rate P _max can be automatically and stably performed, and the average queue length of the FIFO queue of the transmission buffer is set in the gateway 1 or the like on the Internet regardless of the fluctuation of traffic characteristics. You can control buffer congestion by keeping it within a certain range

The present invention is not limited to the examples described with reference to FIGS. 1 to 13, and various modifications can be made without departing from the spirit of the invention. For example, in this example, the configuration and operation in the gateway 1 on the Internet have been described, but the present invention can be applied to other communication devices on the network that transfer packets.

In this example, the RED queue manager
Maximum packet discard rate P in -3 _maxManipulated variable ΔP_max
The maximum packet discard rate P_max
Manipulated variable ΔP_maxIntermediate packet discard due to adaptive adjustment
The operation amount of the rate Pa is also adaptively adjusted.

In this example, the simplified fuzzy inference (direct method) is used as the fuzzy inference, but the indirect fuzzy inference may be used.

Further, the computer configuration of the gateway 1 in this example may be a configuration without a keyboard or an optical disk drive. Further, in this example, an optical disc is used as a recording medium, but an FD (Flexible Disc) is used.
k) or the like may be used as the recording medium. Further, regarding the installation of the program, the program may be downloaded and installed via the network via the communication device.

[0116]

According to the present invention, the control amount of the packet discard rate in the RED queue manager according to the fluctuation of the traffic characteristics is obtained by fuzzy inference, and the parameter of the rule used in this fuzzy inference is determined by the neural network. Since the learning function is used as the weight of the transmission buffer, the queue length in the transmission buffer can be automatically and stably maintained within a certain range even if the traffic characteristics fluctuate, and the congestion of the transmission buffer can be efficiently controlled. It is possible to improve the reliability of the network.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of a packet transfer control system according to the present invention.

FIG. 2 is a block diagram showing a configuration example of a traffic control device in FIG.

3 is a block diagram showing a configuration example of a control parameter correction device in FIG.

FIG. 4 is an explanatory diagram showing an example of control conformity determination conditions stored in a conformity control condition storage unit in FIG.

5 is an explanatory diagram showing a first configuration example of a fuzzy control rule used in the traffic control device in FIG. 1. FIG.

FIG. 6 is an explanatory diagram showing a second configuration example of a fuzzy control rule used in the traffic control device in FIG.

7 is an explanatory diagram showing an example of a convergence condition stored in a control parameter correction condition storage unit in FIG.

FIG. 8 is an explanatory diagram showing an example of control rule generation conditions stored in a control parameter correction condition storage unit in FIG.

9 is an explanatory diagram showing a configuration / operation example of a neural network of the control parameter correction device in FIG.

10 is an explanatory diagram showing a specific operation example of the neural network in FIG.

11 is an explanatory diagram showing an example of the processing operation of the control parameter correction device in FIG.

12 is an explanatory diagram showing an example of the processing operation of a maximum estimation error region selection unit in FIG.

FIG. 13 is an explanatory diagram showing an example of the processing operation of the control rule generation unit in FIG.

FIG. 14 is a block diagram showing a configuration example of a system for performing congestion control and packet transfer control using conventional RED.

15 is an explanatory diagram showing an example of the relationship between the average queue length and the packet discard rate referred to in the system of FIG.

16 is an explanatory diagram showing an outline of operation of a RED queue manager in the system of FIG.

[Explanation of symbols]

1: Gateway (“Gateway: communication device”), 2: Transmission buffer (“FIFO-QUEUE”), 3: RED queue manager (“RED-Queue Manager”), 4: Adaptive adjustment type traffic congestion control system, 41: Traffic control device, 41a: Fuzzy inference unit (“Fuzzy Controlle
r "), 42: control parameter correction device, 42a: neural network (" Neural Network "), 43: system terminal, 50: control conformity determination condition, 51, 61: fuzzy control rule (" Fuzzy control Rule "), 71 : Convergence condition, 81: control rule generation condition, 411: adaptive control condition input / output unit, 412: adaptive control condition storage unit, 413: control parameter input / output unit, 414: monitoring data receiving unit, 4
15: control conformity determination unit, 416: input value calculation unit, 41
7: control analysis unit, 418: output value calculation unit, 419: control value transmission unit, 421: control parameter input unit, 422: control parameter correction condition input / output unit, 423: control parameter correction condition storage unit, 424: learning data Input / output unit, 42
5: learning data storage unit, 426: output estimated value calculation unit, 4
27: convergence condition determination unit, 428: control rule generation condition determination unit, 429: control parameter update processing unit, 4210:
Maximum estimation error area selection unit, 4211: control rule generation unit, 4212: control parameter output unit.

Claims (11)

[Claims] 1. An average queue length of a transmission buffer in which packets to be transferred are accumulated is calculated, a received packet discard control is performed according to a preset packet discard rate corresponding to the average queue length, and a packet transfer amount. And a means for determining whether or not the packet discard rate is adapted to the current traffic characteristics by comparing the obtained average queue length with a preset target value. If the determination result is non-adaptive, means for fuzzy inference to obtain the increase / decrease amount of the packet discard rate so that it is determined to be adapted to the current traffic characteristics by comparison with the target value, and the fuzzy The control parameter of the rule used in the inference is used as the weight of the neural network, and the learning function is used. And a means for changing the packet discard rate. 2. An average queue length of a transmission buffer accumulating packets to be transferred is calculated, and a received packet discard control is performed according to a preset packet discard rate corresponding to the average queue length to transfer the packet. A packet transfer control system for controlling the packet transfer control system, which compares the calculated average queue length with a lower limit value and an upper limit value of a preset target range, and reduces the packet discard rate if the average queue length is smaller than the lower limit value. First means for increasing the packet discard rate if the average queue length is larger than the upper limit value, and second means for calculating the increase / decrease amount of the packet discard rate by the first means by fuzzy inference. When,
A packet transfer control system comprising: a third means for obtaining a control parameter of a rule used in fuzzy inference by the second means as a weight of a neural network by a learning function. 3. An average queue length of a transmission buffer in which packets to be transferred are accumulated is obtained, and a received packet discard control is performed according to a preset packet discard rate corresponding to the average queue length to transfer the amount of packets. A packet transfer control system for controlling a packet transfer control system, which compares the calculated average queue length with a lower limit value and an upper limit value of a preset target range, and reduces the packet discard rate if the average queue length is smaller than the lower limit value. If the average queue length is larger than the upper limit, the first means for increasing the packet discard rate, and the target value and the average queue length from the target value preset for the average queue length Error and the change amount of the error, and the fuzzy inference using the obtained target value and the error and the change amount as input, the packet by the first means. Second to calculate the amount of increase or decrease of disposal rate
And the control parameter of the rule used in the fuzzy inference by the second means, the neural network as a weight in the neural network that inputs the target value and the error and the variation and outputs the increase and decrease of the packet discard rate. A packet transfer control system comprising: a third means for obtaining by network learning. 4. An average queue length of a transmission buffer in which packets to be transferred are accumulated is calculated, a received packet discard control is performed according to a packet discard rate preset corresponding to the average queue length, and a packet transfer amount. A packet transfer control system for controlling the above, comparing the obtained average queue length with a lower limit value and an upper limit value of a preset target range, and if the average queue length is smaller than the lower limit value, the maximum drop rate of the packet And a first means for increasing the maximum drop rate of the packet if the average queue length is larger than the upper limit value, and the target value and the The error from the average queue length and the change amount of the error are obtained, and by the fuzzy inference using the target value and the error and the change amount as input,
The second means for calculating the increase / decrease amount of the maximum discard rate of the packet by the first means, and the control parameter of the rule used for the fuzzy reasoning by the second means, A packet transfer control system, comprising: a third means for obtaining a weight in a neural network, which has an input and the increase / decrease amount as an output, by learning the neural network. 5. A step of obtaining an average queue length of a transmission buffer in which packets to be transferred are accumulated, and performing a discard control of a received packet according to a preset packet discard rate corresponding to the average queue length, A packet transfer control method for a system for controlling a packet transfer amount, comprising:
By comparing the obtained average queue length with a preset target value, a procedure of determining whether or not the packet discard rate is adapted to the current traffic characteristics, and a determination result of non-adaptation, A procedure for determining the increase / decrease amount of the packet discard rate by fuzzy inference so that it is determined that it is adapted to the current traffic characteristics by comparison with the target value, and the control parameters of the rules used in the fuzzy inference are stored in the neural network. The procedure to obtain the weight by the learning function,
And a procedure for changing the packet discard rate by the increment / decrement amount obtained by the fuzzy inference. 6. A procedure for determining an average queue length of a transmission buffer that stores packets to be transferred, and performing control of discarding received packets according to a preset packet discard rate corresponding to the average queue length, A method of controlling packet transfer of a system for controlling a packet transfer amount, comprising: comparing a calculated average queue length with a lower limit value and an upper limit value of a preset target range; If it is smaller, the packet drop rate is reduced, if the average queue length is larger than the upper limit, the packet drop rate is increased, and the packet drop rate increase / decrease is calculated by fuzzy inference.
A packet transfer control method comprising: a control function of a rule used in the fuzzy inference as a weight of a neural network and a learning function. 7. A procedure for determining an average queue length of a transmission buffer in which packets to be transferred are accumulated, and performing a discard control of received packets according to a preset packet discard rate corresponding to the average queue length, A method of controlling packet transfer of a system for controlling a packet transfer amount, comprising: comparing a calculated average queue length with a lower limit value and an upper limit value of a preset target range; If it is smaller, the packet drop rate is decreased, if the average queue length is larger than the upper limit value, the packet drop rate is increased, and the target value is preset from the target value set for the average queue length. The error between the value and the average queue length, and the procedure for obtaining the variation of the error, and the fuzzy inference using the obtained target value and the error and the variation as input. A procedure for calculating the increase / decrease amount of the packet discard rate and a control parameter of the rule used for the fuzzy inference, the target value, the error and the change amount are input, and the neural network that outputs the increase / decrease amount of the packet discard rate is output. And a procedure for obtaining the weight as a weight by learning the neural network. 8. A procedure for obtaining an average queue length of a transmission buffer in which packets to be transferred are accumulated, and performing control of discarding received packets according to a preset packet discard rate corresponding to the average queue length, A method of controlling packet transfer of a system for controlling a packet transfer amount, comprising: comparing a calculated average queue length with a lower limit value and an upper limit value of a preset target range; If it is smaller, the procedure of decreasing the maximum drop rate of the packet, if the average queue length is larger than the upper limit value, the procedure of increasing the maximum drop rate of the packet, and a preset target value for the average queue length. From the target value and the average queue length, and a procedure for obtaining a change amount of the error, and a fuzzy using the target value and the error and the change amount as inputs. A procedure for calculating the increase / decrease amount of the maximum discard rate of the packet by inference, and a control parameter of the rule used for the fuzzy inference, the target value, the error, and the change amount are input, and the increase / decrease amount is output as a neural network. And a procedure for obtaining the weight as a weight by learning the neural network. 9. The computer according to any one of claims 5 to 8.
A program for executing each procedure in the packet transfer control method according to any one of 1. 10. A computer-readable recording medium in which a program for causing a computer to execute each procedure in the packet transfer control method according to claim 5 is recorded. 11. A computer which executes each procedure in the packet transfer control method according to claim 5, and which is based on the processing of the computer,
A communication device for controlling the amount of packets to be transferred.