CN101834901A - Network coordinate system input delay pre-treatment method based on t detection model - Google Patents

Abstract

The invention discloses a network coordinate system input delay pre-treatment method based on t detection model. The method is characterized in that: each network node records nearest H direct measuring delay values between the network node and partial neighbour nodes; under the t detection model, according to the delay observed value of the delay queue, confidence interval of observed value at the next moment of the node is estimated, so as to detect and inhibit abnormal delay observed value and obtain delay result smoothly output by the node. The algorithm is based on probability theory t detection model, history record delay sample information between nodes is utilized, and the abnormal delay observed value of the node at the next moment is detected and inhibited, so as to obtain delay result smoothly output by the node, the delay result is used for carrying out network distance semi measure space embedding, thus ensuring the accuracy of the node to establish network coordinate system delay system predication and convergent cycle thereof.

Description

Network coordinate system input delay pre-treatment method based on the t detection model

Technical field

The present invention relates to Internet technical field, be specifically related to network coordinate system input delay pre-treatment method based on the t detection model.

Background technology

In recent years, along with other growth rate of IP network scaled index level, network configuration isomerism and complexity degree increase, and cause the knowability variation of network internal performance.While is continuing to bring out along with new network application again, no matter the continuous rising that the user requires network service quality is that the network service improves the merchant, or the user, wish that all energy is timely, the firsthand information of exact grasp reflection network current performance, at utmost optimize network application.In the practical IP network, can reflect that the parameter of network operation performance and behavior is a lot, for example: bandwidth, time delay, throughput etc., and time delay is one of its key parameter between node, can directly reflect the performance condition of current network.We often are referred to as " network distance " (Network Distance) to time delay between node.As DHT (Dynamic Hash Table) structure, Overlay route, multicast tree structure etc. in P2P (Peer-to-Peer) network, they can utilize time delay information that its performance is optimized and improve.How by a kind of mode of measuring efficiently, obtaining the time delay between network node, is the hot issue of research now.

In order to obtain network distance, the simplest and direct mode is exactly in internodal initiation Ping probe data packet.Yet this mode is measured number of times, has exponential other quantitative relation with network size, brings very big measurement expense to network thus.For example: in network, need to measure O (N with N platform main frame ²) inferior, low, the poor expandability of its efficient.Another kind is to adopt the indirectly measurement mode, and this only needs the direct measurement result of limited number of time between part of nodes just can all internodal distances predict that its complexity is reduced to O (N) to other; And node can be stored, calculate network delay independently of each other and operation such as processing with method of geometry, makes things convenient for network application.

The latency prediction method of coordinate Network Based is the good indirectly measurement method of a class application prospect that proposes recently.These class methods are to utilize the limited number of time between node and part neighbor node directly to measure time delay information, embed theory based on measure space, network host is mapped as point in the Virtual Space, and, just can utilize in the Virtual Space measure distance between 2 to predict time delay between respective host thus for it distributes corresponding virtual coordinates.The network coordinate method can be with less measurement expense prediction time delay, in the real network environment, they are when the building network coordinate system, usually the limited number of time that will obtain earlier in a period of time between institute's reference section neighbor node is directly measured time delay information, respectively the time delay value that mediates on the numerical values recited in the time delay formation is extracted, form static sparse time delay matrix as the input delay of setting up network coordinate system.In this case, because what make up that coordinate system imports is static sparse time delay matrix,, can not reflect congested and change in topology in the network dynamically though time delay value is stable between node.In reality scene, no matter be network congestion, Network Load Balance or network topology change or the like, can cause the time delay value instability between network node.If these input delays that participate in the network coordinate system structure are not done any processing, then can't guarantee to set up the accuracy and the convergent cycle thereof of network coordinate system latency prediction with it.

Summary of the invention

Problem to be solved by this invention is: how a kind of network coordinate system input delay pre-treatment method based on the t detection model is provided, this method can overcome existing defective in the prior art, guaranteed under the network context environment of complexity, network distance half measure space embeds and theoretical its input delay value stabilization to be changed and can in time reflect the requirement of current network conditions, has ensured the accuracy of network distance indirectly measurement.

Technical problem proposed by the invention is to solve like this: a kind of network coordinate system input delay pre-treatment method based on the t detection model is provided, it is characterized in that, each network node writes down between itself and part neighbor node recently H and directly measures time delay value, under the t testing model, time delay observed value according to this time delay formation, estimate the confidential interval of this internodal next moment time delay observed value, to detect and to suppress unusual time delay observed value, obtain its level and smooth output time delay result, concrete steps are as follows:

1. variable-definition:

A, Sample are the time delay values that limited number of time is directly measured between node AB, A and B are two nodes in the network coordinate system, wherein A is a local node, carry out the renewal process of network coordinate, B is the reference neighbor node of A, comprises whole possible time delay values between node AB in this time delay formation, be called overall, this is to carry out the measure distance space to embed required time delay set, with the form of one dimensional numerical, as the input delay data that will carry out smoothing processing;

B, SA={Sa ₁, Sa ₂... Sa _HBe simple random sampling from overall Sample, by the individual Sa that directly measures time delay between node AB recently for H time ₁, Sa ₂... Sa _HForm its individual Sa ₁, Sa ₂... Sa _HBe the observed result from time delay information among the overall Sample, sample size is that historical record time delay window size is H, H 〉=3,, and, sample SA will along with among the overall Sample up-to-date obtain the time delay individuality directly measured arrival and upgrade;

C, AVER are simple random sampling SA ({ Sa among the overall Sample ₁, Sa ₂... Sa _H) sample average, with the maximal possibility estimation of this sample average as population mean;

D, MAXV are the confidential interval upper bounds of population mean, with simple random sampling SA ({ Sa ₁, Sa ₂... Sa _H) sample average and sample variance as its independent variable:

$\mathrm{MAXV}=\stackrel{\‾}{X}+\frac{S}{\sqrt{n}}{t}_{\mathrm{\α}}(n-1);$

E, Be the sample average of simple random sampling SA, S is the sample standard deviation of simple random sampling SA, and N is the sample size of SA, and what use is the t method of inspection, and 1-α is called confidence level;

F, RTT _IDBe the output result, as the input that network distance half measure space embeds, be used to set up network coordinate system with this time delay value;

2. processing procedure:

A, for the time delay individuality of up-to-date direct measurement, judge the overall Sample that this time delay individuality belongs to according to " neighbor node ID value " in its format fields, " the original time delay value between node " that extract in the individual format fields of this time delay treats the time delay value of smoothing processing as this;

B, according to simple random sampling the SA ({ Sa in the history window between this node ₁, Sa ₂... Sa _H) information, calculate sample average

With sample standard deviation S, thereby further calculate the confidential interval upper bound MAXV of population mean, wherein

Be all the sample average of simple random sampling SA with AVER,

Be used to calculate MAXV, and AVER is as the maximal possibility estimation of population mean;

If c newly in individual format fields " original time delay value between node " greater than MAXV, it is unusual to think that this time delay observed value exists, make smoothing processing output RTT as a result _IDEqual the maximum likelihood estimator AVER of population mean; Otherwise, make RTT _IDEqual " original time delay value between node " in the individual format fields of this new time delay;

D, with " original time delay value between node " in the individual format fields of this time delay, more time delay individual information in the history window of new samples SA guarantees that sample size remains H;

E, with smoothing processing output RTT as a result _ID,, under network coordinate system core algorithm (as Vivaldi), upgrade the coordinate figure of this node as the input delay value of network distance half measure space;

F, the new individual time delay observed value of wait if there is new individual time delay observed value to arrive, jump to step a; Otherwise, continue to wait for.

Useful result of the present invention is: utilize the direct measurement delay data of having learnt between part of nodes, can leach the random delay contamination accident in this link, produce the output time delay value of steady change, embed the virtual coordinates system that sets up thereby carry out the network distance measure space.Guaranteed that under the network context environment of complexity network distance half measure space embeds and theoretical its input delay value stabilization to be changed and can in time reflect the requirement of current network conditions, has ensured the accuracy of network distance indirectly measurement.

Description of drawings

The individual format sample of Fig. 1 time delay;

Fig. 2 algorithm function schematic diagram;

Fig. 3 algorithm function structure chart;

Fig. 4 RS-TDM algorithm workflow diagram.

Embodiment

Below in conjunction with accompanying drawing the present invention is further described:

Time delay information how to utilize node and part to select limited number of time between node directly to measure acquisition is efficiently constructed a stable network coordinate system that satisfies certain measure space definition, and this is extremely important to the latency prediction accuracy that improves the indirectly measurement method.Yet in real network, internodal time delay information suffers the influence of network congestion, Network Load Balance and network topology change etc. easily.If they are not handled, will cause the network coordinate system instability that makes up in view of the above, and can't obtain latency prediction result accurately.For this reason, the present invention proposes a kind of time delay Preprocessing Algorithm, by to participating in the preliminary treatment of the input delay that network coordinate system makes up, improves the stability of network coordinate system.This algorithm utilizes the time delay information of historical record between node, next time delay value of directly measuring constantly between this node is carried out smoothing processing, to suppress its random fluctuation; And then the delay data that smoothing processing is crossed makes up metastable network coordinate system as input.

In order to improve the stability of network coordinate system, improve the network distance forecasting accuracy, the present invention proposes preprocess method at network coordinate system input delay, promptly based on the delay data smoothing processing algorithm of t detection model (RTT Smoothing Algorithm based on t Detection Model, RS-TDM).In the RS-TDM algorithm, each network node writes down between itself and part neighbor node recently H (H 〉=3) and directly measures time delay value, under the t testing model, time delay observed value according to this time delay formation, estimate the confidential interval of this internodal next moment time delay observed value, to detect and to suppress unusual time delay observed value, obtain its level and smooth output time delay result.In the RS-TDM algorithm, we are called totally (Sample) with a node to whole possible direct measurement time delay values, and each the possible time delay observed value during this is overall is called individuality.Directly measure time delay with nearest H time and be designated as Sa ₁, Sa ₂... Sa _H, and with Sa ₁, Sa ₂... Sa _HBe called a sample from overall Sample, H is a sample size.Can calculate sample average and sample variance by this sample, and calculate the confidential interval of population mean based on the t testing model.Because being the nothing of population mean, sample average estimates partially, can be with the confidential interval of the population mean of calculating, check next credibility of sample average constantly, if next moment sample time delay observed value is in the confidential interval of the population mean of being calculated, then use this time delay sample observation, as the input delay value of setting up network coordinate system; If this time delay sample observation is not in the confidential interval of the population mean of being calculated, the RS-TDM algorithm adopts the maximum likelihood estimator of the sample average of previous calculations as population mean, as the input delay value of network distance half measure space.

This algorithm is based on probability theory t detection model, utilize historical record time delay sample information between node, detection also suppresses next constantly unusual time delay observed value between this node, be used for carrying out the embedding of network distance half measure space smoothly to export the time delay result, guarantee to set up the accuracy and the convergent cycle thereof of network coordinate system latency prediction with it.

Delay data smoothing processing algorithm (RS-TDM) based on the t detection model: in the RS-TDM algorithm, each network node writes down between itself and part neighbor node recently H (H 〉=3) and directly measures time delay value, under the t testing model, time delay observed value according to this time delay formation, estimate the confidential interval of this internodal next moment time delay observed value, to detect and to suppress unusual time delay observed value, obtain its level and smooth output time delay result.In the RS-TDM algorithm, we are called totally (Sample) with a node to whole possible direct measurement time delay values, and each the possible time delay observed value during this is overall is called individuality.Directly measure time delay with nearest H time and be designated as Sa ₁, Sa ₂... Sa _H, and with Sa ₁, Sa ₂... Sa _HBe called a sample from overall Sample, H is a sample size.Can calculate sample average and sample variance by this sample, and calculate the confidential interval of population mean based on the t testing model.Because being the nothing of population mean, sample average estimates partially, can be with the confidential interval of the population mean of calculating, check next credibility of sample average constantly, if next moment sample time delay observed value is in the confidential interval of the population mean of being calculated, then use this time delay sample observation, as the input delay value of setting up network coordinate system; If this time delay sample observation is not in the confidential interval of the population mean of being calculated, the RS-TDM algorithm adopts the maximum likelihood estimator of the sample average of previous calculations as population mean, as the input delay value of network distance half measure space.

The selection strategy of RS-TDM algorithm sample size: the simple random sampling capacity is H (H 〉=3), and upgrade this simple random sampling with the individual observed value of up-to-date direct measurement time delay, with the individual time delay observed value in this sample, as the basis for estimation information of smoothing processing output time delay value.

The selection strategy of RS-TDM algorithm time delay population mean confidential interval: in the simple random sampling, preserve nearest H time the time delay individual information directly measured.At this capacity is in the simple random sampling of H, calculates its time delay sample average and time delay sample variance, and calculates the confidential interval of time delay population mean with it under the t testing model.

The smoothing processing result of RS-TDM algorithm: estimate partially because sample average is the nothing of population mean, can be with the confidential interval of the population mean of calculating, check next credibility of sample average constantly, if next moment sample time delay observed value is in the confidential interval of the population mean of being calculated, then use this time delay sample observation, as the input delay value of setting up network coordinate system; If this time delay sample observation is not in the confidential interval of the population mean of being calculated, the RS-TDM algorithm adopts the maximum likelihood estimator of the sample average of previous calculations as population mean, as the input delay value of network distance half measure space.Be characterized in: if the individual time delay observed value fluctuation range of institute's reference sample is big, to cause the population mean confidential interval to become big at interval, the RS-TDM algorithm allows that the output time delay value changes in the population mean confidential interval, and the coboundary of confidential interval can suppress to export the variation that delay data is crossed over the order of magnitude, forms stable smoothing processing output result.

Specific embodiment:

As Fig. 1～shown in Figure 4, in the non-direct latency measurement system of IP based network coordinate, t each node N constantly keeps a table, and this table has two territories, [ID, RTT _ID] _N, wherein ID is the identification number of the neighbor node of node N, RTT _IDFor node N and this neighbor node carry out the time delay value that measure space embeds.

Based on the delay data smoothing processing algorithm (RS-TDM) of t detection model is preprocess method at network coordinate system input delay, this algorithm is by historical time delay record value, come the input delay value of smoothing processing, embed the theoretical requirement that its input delay value stabilization is changed and can in time reflect current network conditions to satisfy network distance half measure space as network coordinate system core algorithm (as the Vivaldi algorithm).

Delay data smoothing processing algorithm (RS-TDM) based on the t detection model

Be convenient and describe that we get two node A and B in the network coordinate system.Wherein A is a local node, carries out the renewal process of network coordinate, and B is the reference neighbor node of A, the time delay value between node AB, and as the input delay of RS-TDM algorithm, and with obtaining the input delay of smoothing processing result as network coordinate system.

(1) variable description

1) .Sample is the time delay value that limited number of time is directly measured between node AB, comprise whole possible time delay values between node AB in this time delay formation, be called overall, this is to carry out the measure distance space to embed required time delay set, with the form of one dimensional numerical, to carry out the input delay data of smoothing processing as the RS-TDM algorithm.。

2) .SA={Sa ₁, Sa ₂... Sa _HBe simple random sampling from overall Sample, by the individual Sa that directly measures time delay between node AB recently for H time ₁, Sa ₂... Sa _HForm its individual Sa ₁, Sa ₂... Sa _HIt is observed result from time delay information among the overall Sample.Here, sample size (being historical record time delay window size) is H (H 〉=3), and, sample SA will along with among the overall Sample up-to-date obtain the time delay individuality directly measured arrival and upgrade.

3) .AVER is simple random sampling SA ({ Sa among the overall Sample ₁, Sa ₂... Sa _H) sample average, in the RS-TDM algorithm, with the maximal possibility estimation of this sample average as population mean.

4) .MAXV is the confidential interval upper bound of population mean, with simple random sampling SA ({ Sa ₁, Sa ₂... Sa _H) sample average and sample variance as its independent variable.

$\mathrm{MAXV}=\stackrel{\‾}{X}+\frac{S}{\sqrt{n}}{t}_{\mathrm{\α}}(n-1)$

5).

Be the sample average of simple random sampling SA, S is the sample standard deviation of simple random sampling SA, and N is the sample size of SA, and what use in the RS-TDM algorithm is the t method of inspection, and 1-α is called confidence level.

6) .RTT _IDBe the output result of RS-TDM algorithm, as the input that network distance half measure space embeds, be used to set up network coordinate system with this time delay value.

(2) algorithmic procedure

Algorithm input: sample observation SA

Algorithm output: smoothing processing is RTT as a result _ID

Algorithm steps:

1). for the time delay individuality of up-to-date direct measurement, judge the overall Sample that this time delay individuality belongs to according to " neighbor node ID value " in its format fields, " the original time delay value between node " that extract in the individual format fields of this time delay treats the time delay value of smoothing processing as this.

2). according to simple random sampling the SA ({ Sa in the history window between this node ₁, Sa ₂... Sa _H) information, calculate sample average

With sample standard deviation S, thereby further calculate the confidential interval upper bound MAXV of population mean.Wherein

Be all the sample average of simple random sampling SA with AVER,

Be used to calculate MAXV, and AVER is as the maximal possibility estimation of population mean.

3) if. newly in individual format fields " original time delay value between node " greater than MAXV, it is unusual to think that in the RS-TDM algorithm this time delay observed value exists, make smoothing processing output RTT as a result _IDEqual the maximum likelihood estimator AVER of population mean.Otherwise, make RTT _IDEqual " original time delay value between node " in the individual format fields of this new time delay.

4). with " original time delay value between node " in the individual format fields of this time delay, more time delay individual information in the history window of new samples SA guarantees that sample size remains H (H 〉=3).

5). export RTT as a result with smoothing processing _ID,, under network coordinate system core algorithm (as Vivaldi), upgrade the coordinate figure of this node as the input delay value of network distance half measure space.

Wait for new individual time delay observed value,, jump to step 1 if there is new individual time delay observed value to arrive; Otherwise, continue to wait for.

Claims (1)

1. based on the network coordinate system input delay pre-treatment method of t detection model, it is characterized in that, each network node writes down between itself and part neighbor node recently H and directly measures time delay value, under the t testing model, according to the time delay observed value of this time delay formation, estimate this internodal next confidential interval of time delay observed value constantly, to detect and to suppress unusual time delay observed value, obtain its level and smooth output time delay result, concrete steps are as follows:

1. variable-definition:

A, Sample are the time delay values that limited number of time is directly measured between node A, B, A and B are two nodes in the network coordinate system, wherein A is a local node, carry out the renewal process of network coordinate, B is the reference neighbor node of A, comprises whole possible time delay values between node AB in this time delay formation, be called overall, this is to carry out the measure distance space to embed required time delay set, with the form of one dimensional numerical, as the input delay data that will carry out smoothing processing;

B, SA={Sa ₁, Sa ₂... Sa _HBe simple random sampling from overall Sample, by the individual Sa that directly measures time delay between node AB recently for H time ₁, Sa ₂... Sa _HForm its individual Sa ₁, Sa ₂... Sa _HBe the observed result from time delay information among the overall Sample, sample size is that historical record time delay window size is H, H 〉=3, and, sample SA will along with among the overall Sample up-to-date obtain the time delay individuality directly measured arrival and upgrade;

C, AVER are simple random sampling SA ({ Sa among the overall Sample ₁, Sa ₂... Sa _H) sample average, with the maximal possibility estimation of this sample average as population mean;

D, MAXV are the confidential interval upper bounds of population mean, with simple random sampling SA ({ Sa ₁, Sa ₂... Sa _H) sample average and sample variance as its independent variable:

$\mathrm{MAXV}=\stackrel{\‾}{X}+\frac{S}{\sqrt{n}}{t}_{\mathrm{\α}}(n-1);$

E,

Be the sample average of simple random sampling SA, S is the sample standard deviation of simple random sampling SA, and N is the sample size of SA, and what use is the t method of inspection, and 1-α is called confidence level;

F, RTT _IDBe the output result, as the input that network distance half measure space embeds, be used to set up network coordinate system with this time delay value;

2. processing procedure:

A, for the time delay individuality of up-to-date direct measurement, judge the overall Sample that this time delay individuality belongs to according to " neighbor node ID value " in its format fields, " the original time delay value between node " that extract in the individual format fields of this time delay treats the time delay value of smoothing processing as this;

B, according to simple random sampling the SA ({ Sa in the history window between this node ₁, Sa ₂... Sa _H) information, calculate sample average

With sample standard deviation S, thereby further calculate the confidential interval upper bound MAXV of population mean, wherein Be all the sample average of simple random sampling SA with AVER, Be used to calculate MAXV, and AVER is as the maximal possibility estimation of population mean;

If c newly in individual format fields " original time delay value between node " greater than MAXV, it is unusual to think that this time delay observed value exists, make smoothing processing output RTT as a result _IDEqual the maximum likelihood estimator AVER of population mean; Otherwise, make RTT _IDEqual " original time delay value between node " in the individual format fields of this new time delay;

D, with " original time delay value between node " in the individual format fields of this time delay, more time delay individual information in the history window of new samples SA guarantees that sample size remains H;

E, with smoothing processing output RTT as a result _ID,, under the network coordinate system core algorithm, upgrade the coordinate figure of this node as the input delay value of network distance half measure space;

F, the new individual time delay observed value of wait if there is new individual time delay observed value to arrive, jump to step a; Otherwise, continue to wait for.

Images