CN100435542C - Overload detection method for communication transaction processing system - Google Patents

Abstract

The present invention provides an overload detection method for a communication transaction processing system, which is in view of the rule that high system load certainly causes the queue time of background messages to be lengthened and the response time to be prolonged according to the change relationship of system resource occupancy rate and response time. The method has the steps that two measurable physical quantities, a transaction average successful call rate lambda and a corresponding system response time T, are measured, and then resource occupancy rate R is calculated according to correlation formulas and mechanisms designed by the present invention so as to judge whether a system is overload or not. The method does not depend on various concrete environments or realizing conditions in a system, but only automatically carries out overload detection according to the size of transaction load in the system. The method has the advantages of simple method, accurate judgment and reliable work. The method is combined with a patent application of a method for controlling overload of intelligent networks based on service control points in multi-service environments of the applicant, and can nicely realize the overload detection and the overload control of network systems.

Description

A kind of overload detection method that is used for communication transaction processing system

Technical field

The present invention relates to a kind of overload detection method that is used for communication transaction processing system, belong to the overload detection technical field in the communication system.

Background technology

It is a very important technical problem that the communications transaction system is carried out overload detection, because it is the prerequisite of prevention communications transaction system overload, enforcement flow control.At present, take following several method commonly used usually about overload detection:

(1) adopt the system of message queue whether formation is overflowed as testing conditions, it waits pending message queue with regard to extending according to being that processing speed is slack-off when system loading increases, when queue length is increased to certain value, just thinks overload;

(2) judge whether link load is too high, promptly when link load reaches the warning value of its design capacity, just think overload;

(3) judging whether the occupancy of CPU is too high, because CPU usage has been reacted the busy not busy degree of system to a certain extent, as the overload detection condition, is effective with the occupancy of CPU in suitable scope;

(4) system to the response speed of message as the overload detection condition.

Though above-mentioned first three methods can both detect, judge system and whether transship that certain limitation is all arranged within the specific limits.Wherein whether formation overflows relevantly with concrete hardware platform in the method (1), and the queue length of overload is not easy to determine.Method (2) can not really reflect the system overload situation sometimes, such as in situation such as the configuration of link capacity and treatment system is disproportionate.Method (3) is to judge method the most intuitively whether transshipped in system, but, for the complicated distributed system, accurately obtain the occupancy of CPU and be not easy, particularly the measurement of the occupancy of CPU can only be in the system that manages business of reality, usually with take to transship the locus of equipment of control and separate, will cause transmitting the hysteresis of CPU usage information like this, the information that the control point knows CPU usage of promptly transshipping lags behind in time certainly, therefore, even take to transship control measure, also inevitable hysteresis in time makes the control effect to have a greatly reduced quality.And, in case the situation that the occupancy rate information of CPU is lost occurs in transmittance process, will make the serious situation of overload control complete failure.

Method (4) is though the easier realization of first three methods relatively, because the overload that causes of reason whatsoever, the most direct form of expression is that the response speed of message is slack-off, promptly judges by the variation of detect-message response time normally a kind of more effective detection method whether transshipped in system.Yet, up to now, adopt the response time to detect the method for transshipping, all will be according to the base condition of the response time value of measuring in advance or being provided with as overload detection; And this method only is applicable to fixed system and fixed service, promptly to each system that is applied to existing network before practical application, all to carry out measurement, to judge the detection reference of overload based on the average response time of set business.Otherwise, in case environmental change (comprising the environmental change of system hardware and software and/or the variation of service environment) all will cause the variation of overload detection Rule of judgment or benchmark.

Therefore, how to find the overload detection point of system effectively, and the various variations that can conform, also do not find a good solution so far, become the problem that those skilled in the art paid close attention to.

Summary of the invention

In view of this, the purpose of this invention is to provide a kind of overload detection method that is used for communication transaction processing system, this method can not rely on the various concrete environment or the realization condition of system, just carries out overload detection automatically according to the size of the traffic load in system situation; And computational methods are simple, accuracy of judgement, reliable operation.

In order to achieve the above object, a kind of overload detection method that is used for communication transaction processing system is characterized in that: comprise the steps:

(1) when system's nonoverload, is provided with and has following relational expression between system response time T and the professional average call arrival rate λ: T=a+b λ; In the formula, parameter a is the transmission time of every message in system; Parameter b is the proportionality coefficient of every consumption per-unit system resource to the message response time;

(2) lower at system loading, when being underload, measuring the data of many professional average call arrival rate λ of group and corresponding system response time T thereof, and adopt least square method to calculate above-mentioned two parameter a, b;

(3) in the time period of setting, measure the professional average call arrival rate λ of this time period, utilize the relational expression T=a+b λ of step (1) again, calculate the predicted value T of the system response time T of this time period _Prediction

(4) in the same time period of setting, measure the actual numerical value of the system response time T of this time period, i.e. system's actual response time T _Measure

(5) according to the actual response time T of system _MeasurePredicted value T with system response time T _PredictionBetween numerical values recited, whether decision-making system " overload detection point " occur:

If T _Measure＞T _Prediction+ ω σ, then system overload; Otherwise, system's nonoverload; ω is the overload protection factor in the formula, prevents the erroneous judgement that the response time fluctuation causes, and gets empirical value: 1～5; σ is a residual standard deviation.

This method further comprises the steps:

(6) described two parameter a, b are regularly dynamically adjusted,, calculate the numerical value of two parameter a, b, with the variation of adaptive system and service environment promptly every setting-up time execution in step (2).

In the described step (2), the concrete steps of calculating a, two parameters of b according to least square method are:

(21) when system loading is low, the professional average call arrival rate of measuring N group λ _iAnd corresponding system response time T _iData, i is the sequence number of every group of data, N is a positive integer, the maximum sequence number of the data set that expression is measured;

(22) according to every group of professional average call arrival rate λ _iWith system response time T _iData and following formula, calculate following three numerical value respectively:

$S_{\mathrm{\λ\λ}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i^2-\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)}^2;$

$S_{\mathrm{TT}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i^2-\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i\right)}^2;$

$S_{\mathrm{\λT}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{T}_i-\frac{1}{N}\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i\right);$

(23) according to following two formula, calculate two coefficient a and b respectively, and calculate residual standard deviation σ:

$b=\frac{S_{\mathrm{\λT}}}{S_{\mathrm{\λ\λ}}};$

$a=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i-\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)b;$

The formula that calculates residual standard deviation σ is: $\mathrm{\σ}=\sqrt{\frac{1}{N-2}[S_{\mathrm{TT}}-{\mathrm{bS}}_{\mathrm{\λT}}]}.$

The present invention is a kind of overload detection method (claiming linear regression method again) that is used for communication transaction processing system, the advantage of this method is: various concrete environment, hardware platform or other realization condition that can not rely on system, as long as measure the several number certificate and carry out some calculating, just can carry out overload detection automatically according to the size of the traffic load in system situation; And computational methods are simple, accuracy of judgement, reliable operation.The patent application of this method and applicant the control method of the overload of intelligent network of service control point " under the multiservice environment based on " combines, and can realize the overload detection and the overload control of intelligent net system well.And the inventive method has versatility, can be applied to variety of network systems, is not limited to the overload detection of intelligent net system.

Description of drawings

Fig. 1 is the average response time in the communication transaction processing system and the variation relation figure of CPU usage (when being different loads).

Fig. 2 is the operating procedure flow chart of overload detection method of the present invention.

Fig. 3 uses Service Control Point of intelligent network (SCP) system was revised and be applied to overload detection method of the present invention to the self-adapting window control algolithm embodiment test curve figure.

Embodiment

For making the purpose, technical solutions and advantages of the present invention clearer, the present invention is described in further detail below in conjunction with accompanying drawing.

Now introduce the working mechanism of the inventive method: in transacter, system is the reflection of system's spare time, busy degree to the response time speed of message, therefore study the Changing Pattern of response time with system loading (being often referred to CPU usage), just can be well according to the response time situation of change of understanding system loading directly perceived.The great deal of experiment data analysis result shows, under the same load situation, system approximately distributed according to negative exponent to the response time of message.Under the different loads situation, system to the variation relation of the occupancy of response time of message and CPU as shown in Figure 1.

Referring to Fig. 1, the figure illustrates the Changing Pattern of average response time when different loads: during system load light (occupancy of CPU is lower), the response time of system is steady substantially, the approximate linear rule that is; Only when system loading is very heavy, promptly the occupancy of CPU is greater than after 60%, and system response time just can be index law to be increased.

Studies show that the response time T of system and the relation between the resources occupation rate R can be expressed as:

T＝c+de ^βR (1)

In the formula, c is a constant, the disposal ability of reflection system; D is a constant, the complexity of reflection message; β is an intensity factor, and is relevant with professional complexity usually.C=42 among Fig. 1, d=0.36, β=0.075.

Though cpu resource occupancy that Fig. 1 showed and the Changing Pattern of system response time just based on the test result of a platform specific, have general significance, are equally applicable to other platform.Although when different professional of different platform operation, each coefficient of above-mentioned formula (1) may be different, but the cpu resource occupancy is consistent with the rule that the variation of response time is followed: any system is in case load is very high, must cause the formation lengthening of backstage message queueing, the response time will be index law and strengthen.

Generally, in the transaction system, response time T is a scalable physical quantity, and resources occupation rate R then is the physical quantity that can not accurately measure, particularly in distributed system, and the more difficult measurement of the numerical value of R.In aforementioned formula (1), if can confirm the value of three parameter c, d, β, the T numerical value by measuring just can be obtained resources occupation rate R according to formula (1) at an easy rate, thereby judge whether system transships again.Because adopt the numeric ratio of above-mentioned three parameter c of the accurate measurement of conventional method, d, β difficult, must find a kind of convenience, effective method to calculate the numerical value of c, d, β automatically, to obtain the functional relation of T and R.This also is the starting point of development of the present invention.

Because resources occupation rate R is the physical quantity that almost can not accurately measure, must seeks a kind of not only measurement easily, but also can reflect that the physical quantity of occupation condition substitutes it.By analysis, find that this physical quantity exists.The occupancy of system resource depends primarily on professional rate of arriving calls λ, and rate of arriving calls λ is a physical quantity that ratio is easier to measure.If therefore can grasp the relation between λ and the R, just can understand R indirectly by measuring λ.For same business, say that from the statistical significance each calls out the system resource that consumes is certain, rate of arriving calls is high more, and the system resource that is consumed is just many more; Suppose that system resource is abundant, λ and R should be linear relationships.Experimental data shows: this linear relationship is set up when system's nonoverload, that is: R=γ λ (2)

In the formula, γ is a proportionality coefficient, and is relevant with concrete business.For the ease of analyzing, formula (1) is made Taylor expansion near the origin of coordinates 0, get its linear segment, obtain following formula:

T＝c+d+dβR (3)

Formula (2) formula substitution formula (3), obtain following formula again:

T＝a+bλ (4)

In the formula, a=c+d, b=d * β * γ, parameter a represent the time that every group of message is consumed when transmitting in system, are included in the queuing time of message queue, the passing time between disparate modules etc., and its numerical values recited is relevant with the complexity of message.Parameter b is represented every consumption per-unit system resource to the proportionality coefficient of message response time, and message is complicated more usually, and consume system resources is many more, reflects that the delay of response time is big more.

Based on above-mentioned analysis, overload detection method of the present invention comprises following operating procedure (referring to Fig. 2):

(1) when system's nonoverload, is provided with and has following relational expression between system response time T and the professional average call arrival rate λ: T=a+b λ; In the formula, parameter a represents that every group of message transmits the time that is consumed in system, be included in the queuing time of message queue, and the passing time between disparate modules etc., its numerical values recited is relevant with the complexity of message.Parameter b is represented every consumption per-unit system resource to the proportionality coefficient of message response time, and message is complicated more usually, and consume system resources is many more, and then the delay of response time is big more.

(2) when system loading lower (promptly underload), gather measuring N group system response time T and corresponding service average call arrival rate λ data thereof, and according to two parameter a, b in the least square method calculating above-mentioned relation formula; Concrete steps are:

(21) when system loading is low, the professional average call arrival rate of measuring N group λ _iAnd corresponding system response time T _iData, i is the sequence number of every group of data, N is a positive integer, the maximum sequence number of the data set that expression is measured;

(22) according to every group of professional average call arrival rate λ _iWith system response time T _iData and following formula, calculate following three numerical value respectively:

$S_{\mathrm{\λ\λ}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i^2-\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)}^2;$

$S_{\mathrm{TT}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i^2-\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i\right)}^2;$

$S_{\mathrm{\λT}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{T}_i-\frac{1}{N}\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i\right);$

(23) according to following two formula, calculate two coefficient a and b respectively, and calculate residual standard deviation σ:

$b=\frac{S_{\mathrm{\λT}}}{S_{\mathrm{\λ\λ}}};$

$a=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i-\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)b;$

The formula that calculates residual standard deviation σ is: $\mathrm{\σ}=\sqrt{\frac{1}{N-2}[S_{\mathrm{TT}}-{\mathrm{bS}}_{\mathrm{\λT}}]}.$

(3) in the time period of setting, measure the professional average call arrival rate λ of this time period, utilize formula T=a+b λ to calculate the predicted value T of the system response time T of this time period again _Prediction

(4) in the section of setting at the same time, measure the actual value of the system response time T of this time period, i.e. system's actual response time T _Measure

(5) according to the actual response time T of system _MeasurePredicted value T with system response time T _PredictionBetween numerical values recited, whether decision-making system " overload detection point " occur:

If T _Measure＞T _Prediction+ ω σ, then system overload; Otherwise, system's nonoverload; ω is the overload protection factor in the formula, prevents the erroneous judgement that the response time fluctuation causes, and gets empirical value usually: 1～5; σ is a residual standard deviation.

(6) two parameter a, b are regularly dynamically adjusted, promptly carry out above-mentioned steps (2), calculate the numerical value of two parameter a, b, with the variation of adaptive system and service environment every setting-up time.

In order to verify the validity of the inventive method, this method is applied to the Chinese patent application control method of the overload of intelligent network of service control point " under the multiservice environment based on " (application number: 200510064628.9) test enforcement.Introduce the test situation of this embodiment now:

According to the Chinese patent application control method of the overload of intelligent network of service control point " under the multiservice environment based on " (application number: the content 200510064628.9) can obtain following formula:

$T=\frac{W}{\mathrm{\λ}}---\left(6\right)$

In the formula, T is the average response time of system, and λ is for entering system's average call arrival rate, and W is average occupied window number.Formula (6) substitution formula (4), obtain:

$T=\frac{1}{2}(a+\sqrt{a^2+4\mathrm{bW}})---\left(7\right)$

According to the content of the patent of the control method of the overload of intelligent network of service control point " under the multiservice environment based on ", can predict the average occupied window number W of next measurement point according to λ _Prediction, utilize the predicted value T that formula (7) just can solving system response time T again _Prediction, again this T _PredictionNumerical value and the actual response time T of system _MeasureCompare, if satisfy formula T _Measure＞T _Prediction+ ω σ can determine the overload detection point.

Referring to Fig. 3, this figure utilizes the revised self-adapting window control algolithm of linear regression method of the present invention to be applied to the embodiment test result curve chart of Service Control Point of intelligent network (SCP) system, its ordinate is the average call arrival rate, and abscissa is a system operation time.Hardware platform is a HP ALPHA DS20 minicomputer, in 50 fens clock times, increases rate of arriving calls to 450 calling/second gradually; protection factor ω=5; overload detection method according to the present invention judges, in case when overload conditions takes place, start overload control immediately.The data of embodiment are: the best window number when starting overload control is 11.As can be seen from the figure: under overload detection method control of the present invention, the maximum average call arrival rate that the SCP system receives was 196 calling/seconds, and higher rate of arriving calls is rejected.

Before overload detection point is determined, be that the average call arrival rate is during less than 196 calling/seconds, when each use linear regression method of the present invention is sought overload detection point, the linear relationship of average response time and resources occupation rate is better in this is interval, correspond to the non-overload region of Fig. 1, be judged as and do not reach the overload detection condition, do not limit calling out.But when the average call arrival rate reached for 196 calling/seconds, response time and average call arrival rate are non-linear relation, correspond to the overload region of Fig. 1, can determine the overload detection point according to the formula in the abovementioned steps (5), for subsequent voice calls, then start flow control according to the overload detection condition.In the process of seeking overload detection point, system adopts the inventive method to finish automatically according to the system loading situation and detects judgement.Experimental result shows: reached the effect of overload detection and overload control, realized goal of the invention.

Claims (3)

1, a kind of overload detection method that is used for communication transaction processing system is characterized in that: comprise the steps:

(1) when system's nonoverload, is provided with and has following relational expression between system response time T and the professional average call arrival rate λ: T=a+b λ; In the formula, parameter a is the transmission time of every message in system; Parameter b is the proportionality coefficient of every consumption per-unit system resource to the message response time;

(2) lower at system loading, when being underload, measuring the data of many professional average call arrival rate λ of group and corresponding system response time T thereof, and adopt least square method to calculate above-mentioned two parameter a, b;

(3) in the time period of setting, measure the professional average call arrival rate λ of this time period, utilize the relational expression T=a+b λ of step (1) again, calculate the predicted value T of the system response time T of this time period _Prediction

(4) in the same time period of setting, measure the actual numerical value of the system response time T of this time period, i.e. system's actual response time T _Measure

(5) according to the actual response time T of system _MeasurePredicted value T with system response time T _PredictionBetween numerical values recited, whether decision-making system " overload detection point " occur:

If T _Measure＞T _Prediction+ ω σ, then system overload; Otherwise, system's nonoverload; ω is the overload protection factor in the formula, prevents the erroneous judgement that the response time fluctuation causes, and gets empirical value: 1～5; σ is a residual standard deviation.

2, the overload detection method that is used for communication transaction processing system according to claim 1, it is characterized in that: this method further comprises the steps:

(6) described two parameter a, b are regularly dynamically adjusted,, calculate the numerical value of two parameter a, b, with the variation of adaptive system and service environment promptly every setting-up time execution in step (2).

3, the overload detection method that is used for communication transaction processing system according to claim 1 is characterized in that: in the described step (2), the concrete steps of calculating a, two parameters of b according to least square method are:

(21) when system loading is low, the professional average call arrival rate of measuring N group λ _iAnd corresponding system response time T _iData, i is the sequence number of every group of data, N is a positive integer, the maximum sequence number of the data set that expression is measured;

(22) according to every group of professional average call arrival rate λ _iWith system response time T _iData and following formula, calculate following three numerical value respectively:

$S_{\mathrm{\λ\λ}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i^2-\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)}^2;$

$S_{\mathrm{TT}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i^2-\frac{1}{N}{\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i\right)}^2;$

$S_{\mathrm{\λT}}=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i{T}_i-\frac{1}{N}\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i\right);$

(23) according to following two formula, calculate two coefficient a and b respectively, and calculate residual standard deviation σ:

$b=\frac{S_{\mathrm{\λT}}}{S_{\mathrm{\λ\λ}}};$

$a=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{T}_i-\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\λ}}_i\right)b;$

The formula that calculates residual standard deviation σ is: $\mathrm{\σ}=\sqrt{\frac{1}{N-2}[S_{\mathrm{TT}}-{\mathrm{bS}}_{\mathrm{\λT}}]}.$

Images