ES2370424B1 - Method for determining in real time the causes of deterioration of service quality in converged networks. - Google Patents

Abstract

Method for the determination in real time of the causes of the deterioration of the quality of the service in convergent networks. # In this invention it is tried to identify between the original critical metrics of the services and of the network (that originate yield drops and changes in the related metrics) in convergent networks. The process is highly automated, presents an almost real-time response to changes in the CIPN, and does not require human intervention. The invention extends the concept of correlation to obtain the real causality relationships between performance deterioration and metrics in large CIPN, in which the rate of incoming events can easily overload the system's processing power.

Description

Method for the determination in real time of the causes of the deterioration of the quality of the service in convergent networks.

Technical sector

The present invention is related to the field of telecommunications network management and, more particularly, to the management of the distribution of integrated services in Next Generation Networks (NGN), where the Telecommunications and Technology Services of the Information is integrated into Converged IP Networks (CIPN).

Specifically, this invention helps to identify between the original critical services and network metrics, in converged networks.

Metric means a value that is assigned to a telecommunication route for a particular network interface that identifies the cost associated with the use of that route. For example, the metric can be assessed based on the speed of the link, the number of hops or the time delay.

In this description, a cluster of data means a set of data grouped by a causal relationship.

State of the art

Allowing the distribution of integrated services in the Next Generation Networks (NGN), where Telecommunications and Information Technology services are being integrated into Converged IP Networks (CIPN), is a complicated task given the underlying heterogeneity in the services offered on CIPN networks. An example of the prototype tool to achieve this integration objective is EzWeb [1]. EzWeb provides a Graphical User Interface (GUI) for a unified and homogeneous service. This platform is used for the integration of mega-services, services with millions of users, creating a market scenario, where the services can be registered and made in the web browser by the end users.

EzWeb will allow Telecommunications operators to go beyond the traditional distribution of voice / data services, allowing the integration of a variety of services that change permanently (often, over CIPN networks and Multimedia Subsystems over IP). For example, content distribution services (such as Imagenio [2]), purchase channels, Voice over IP capabilities, etc., allow the user of light terminals (for example PDA, notebook, etc ...) to have all services and information that are accessible from the network itself.

This CIPN-based scenario requires the deployment of various infrastructure resources (networks, computers, firewalls, load balancers, etc.) that have to be provided by one or more data centers or network nodes. The management of a Data Center or node for the provision of services requires an agile and efficient management directed by the assurance of the Quality of Service (QoS) policy.

The QoS guarantee is automatically managed through the supervision of some Service metrics, often referred to as Key Performance Indicators (KPI). There are various technologies to monitor services and networks such as probes, agents, log analysis, etc., that are capable of monitoring many metrics. These technologies include simple network sniffers, such as Wireshark [3], tools that add remote management, administration and distribution reports for simple network monitoring [4] or the traditional Simple Network Management Protocol (SNMP), which forms part of the range of Internet protocols as defined by the Internet Engineering Task Force (IETF). In addition, large-scale data collection and monitoring platforms are also available, built for Telecommunications and services companies [5].

However, it is crucial to identify the relevant metrics that have a direct influence on the perception of QoS. This task is not trivial in the scenario described, as there are many interactions between services and network components that are not obvious to operating technicians. The objective of this solution is to reveal the hidden dependencies between the metrics, which allow us to identify the root causes of service failures, QoS crashes, etc. Once these dependencies have been found, operators will be able to detect and prevent those situations in advance.

There are a large number of publications that addressed this problem, but could not establish real causal relationships [6, 7], which is a way of assessing the relevance of a given metric. In most systems, software helps the operator build such relationships [8, 9] or users have to learn these relationships for themselves [10]. The correlation is used to obtain the potential causality between metrics. In the United States, advanced semi-automated correlation techniques based on time series have been patented for business processes that provide additional certainty to the causal relationship [11]. The method of the present invention is different from said techniques due to several reasons. In the first place, the grouping of clusters formed on the basis of a causal relationship (which allows to discover hidden models) is performed, according to the value of the correlation coefficients. Furthermore, the present invention extends the proposal of said techniques [11] by optimizing the algorithms for an oppositely different application domain. Secondly, the process of extracting information entails a heavy calculation overhead to discover statistical relationships [12]. In the present invention a rapid indexing technique is proposed to alleviate this problem, helping to guide the discovery process in an optimal order [13]. The problem is further complicated because the services and, therefore, the measurements and their relationships will be continuously updated in the CIPN, which makes the present invention different from other publications related by causality [12].

Technical problem raised

The traditional solution to detect optimal metrics usually depends on the expertise of domain experts and is laborious and error-prone. About 20% of the work to identify a success metric comes from an install-observe-adjust cycle [14,15,16]. This trial and error procedure may not be feasible given the complexity, distribution and increasing size of the CIPN. In modern CIPN, new services that interact with previously existing services are continuously introduced into the network. Therefore, even the most experienced operators have little or no experience in their networks, making expert knowledge quickly outdated.

Tradition business intelligence systems [17,18] mainly provide predefined historical metrics (or backward indicators) during a long- or medium-term statistical analysis [19] and lack the necessary flexibility to support metrics or data collection in evolution over time for real-time analysis. At any given time, the metric that reveals a network failure may be different, so the process must be iterative.

Other techniques are available to detect important metrics [20], in which they apply a priority weighting vector (PWV) and integrate it into an SOM neural network (the normal states of the system are described with a self-organized map identi fi ed to from the data A large deviation of the data samples in the SOM nodes is detected as an anomalous behavior). Some authors use a scattered plot analysis in order to determine the point at which performance begins to decline and use the Winter Additive Smoothing Model to obtain the statistical trend and predict the evolution of the system [16]. However, these approaches are used to detect abnormal functioning of a single cell in the network and are based on Euclidean distance methods (not very suitable for detecting correlations in time series). Therefore, CIPN require more complex correlation techniques than mobile networks [21].

In modern CIPN, services can reside worldwide, and cause anomalous operations at the other end of the globe. Currently available techniques often deal with local data, suffering from a lack of global integration of the entire network.

Brief Description of the Invention

The objective of this invention is to help identify between the original critical services and network metrics (which result in performance drops and changes in related metrics) in converged networks. The process is highly automated and presents an almost real-time response to changes in the CIPN, and does not require human intervention. This invention extends the concept of correlation to obtain the real causal relationships between performance deterioration and metrics in large CIPNs, in which the rate of incoming events can easily overload the system process power.

It should be noted that these critical metrics may change during the lifetime of the CIPN as new services are added. Consequently, the process proposed here is highly adaptable to identify changes in the status of the CIPN on a global scale. This scalability is essential in distributed environments where potential interactions of services and network elements can take place all over the world, but degrade systems locally.

The process is highly automated, which avoids the costly trial and error procedures in the detection of the optimal metrics (those that really degrade the QoS), thanks to the specific causality correlation modules and time series that are included. In addition, the procedure is capable of dealing with the exponential increase in the number of services (and metrics) of interaction in the CIPN.

The invention presents an almost real-time response to changes in the CIPN, and does not require human intervention. It should be noted that these critical metrics may change during the lifetime of the CIPN, as new services are added to it and depend heavily on the behavior of the added services. The process proposed in this invention is highly adaptable to identify changes in the status of the CIPN on a global scale, where the data massively reaches the process nodes.

The proposed process improves the calculation of correlations of the metrics by calculating time shifts with non-metric distance functions, improving the Euclidean distance methods and the coefficients of simple correlation.

The invention is based primarily on techniques for reducing dimensionality and reinforcing them by adding a method of classification that allows the system to prioritize measurements by establishing a highly overloaded supervision system, such as that existing in the CIPN.

The invention provides a method for determining in real time the causes of deterioration of service quality in converged networks, which comprises the following basic stages:

to.

reduce the dimension of incoming events;

b.

make an orderly correlation of these events to find the causality relationships between the prioritized and high correlation data, building at least one causality distance matrix, and obtaining the potential causality between metrics, and

C.

grouping the data into clusters, formed on the basis of a causality relationship, forming data clusters with a high causal relationship based on said previously constructed causality distance matrix,

wherein said distance matrix of step b) provides the hierarchical relationships by means of a list of potentially operable indicators or dendrogram.

According to a preferred embodiment of the invention, the actual number of clusters in step c) is determined by the Akaike Information Criteria, in which the probability of log-likelihood of the model is increasingly penalized with the number of parameters .

Brief description of the drawings

To complete the description and in order to help a better understanding of the characteristics of the present invention, a set of figures is attached as an integral part of this description, where, as an illustration and not limitation, the following has been represented:

Figure 1 represents the proposed method to extract the relevant metrics; Y

Figure 2 represents a sample embodiment of the proposed method for extracting the relevant metrics.

Detailed description of the invention

In Figure 1, you can see how raw data and integrated data are arranged together (steps 1 and 2). Given the enormous volume of metrics information, it is necessary to make a dimensionality reduction (thus avoiding “the curse of dimensionality” http://en.wikipedia.org/wiki/Curse_of_dimensionality, present in large data sets of the systems of data extraction) (step 3). In this invention, dimensionality reduction techniques are used, such as Principal Component Analysis (PCA) [26], Decomposition in Singular Values (SVD) [22], or the Accumulation Approach of Pieces (PAA) [23 ], to filter the least significant metrics. These or other data reduction techniques are useful for separating less relevant metrics. In addition, a model of the optimal distribution process (PAA) [23] is made, such as the classic vertex cover problem in graph theory, which is a complete NP problem [25].

Once the dimensionality of the data is reduced, a metric correlation (step 4) is performed to find a relationship between the data. Several correlation techniques are available in printed information [24-27]. However, traditional metric distance functions, such as Euclidean distance and correlation coefficients, are not suitable for detecting correlation between time series [28]. There are other non-metric distance functions that calculate the time shifts between various metrics, which help determine the time order of these metrics (for example, Dynamic Time Deformation [29], DTW).

Correlation measurements do not imply a causal relationship (which is what is sought in this invention, to determine those metrics on which action can be taken to modify the overall system performance). But fortunately, one can rely on causality theories (step 5) to provide the solid statistical basis for checking the causal direction of various time series (for example, Granger's theory [30], Bayes' solutions that use values specific metrics to transform the previous estimate of probability into a subsequent estimate of the probability of the causes of the metric, or the Eigen functions of the Hilbert Transform [31]).

However, any metric can have many potentially determining metrics, so that the directional complexity increases with the number of metrics (therefore, the importance of dimensionality reduction is evident). This problem can be solved by accumulating the metrics (the metrics in a cluster have a higher correlation); in this way the temporarily preceding metrics in the cluster are supposed to determine the result of the others in the same cluster, with a higher relative weight than the rest of other clusters (step 6). The modification of the cumulative clustering technique is necessary in this case to cut edges of uncertainty when the set of clusters generated is too complex.

This section presents a concrete embodiment of the general method presented above (whose summary can be found in Figure 2). The raw data is generated from network probes and software agents that reside in the network infrastructure (1). This data is accumulated afterwards (a compromise between the response capacity and the reduction of data in the network must be achieved) using simple statistics (mean and medium) and sent to the processing center (2).

Then, PCA and SVD techniques are applied to reduce the dimensionality of massively generated data and heuristic selection of peaks (greedy vertex) is used based on the degree of the peak, and two simple algorithms are proposed to generate an estimated range (indexation) at the peaks (i.e. measurement points), based on the edges (existing statistical relationships) that affect them. The two algorithms are based on random sampling as in the reference [25]. This step (3) is therefore crucial to handle the huge amount of incoming events.

Once the dimension has been considerably reduced and incoming events have been prioritized, an orderly correlation (DTW [29]) is made, according to the previously obtained ranking (4). These steps allow you to search for causal relationships (applying Granger causality) between prioritized and high correlation data (5). Finally, data clusters with a high causal relationship (6) are formed, based on the previously constructed causality distance matrix, to obtain hierarchical relationships in a dendrogram, that is, a list of potentially operable indicators. However, the traditional cumulative cluster has the disadvantage of finding the edges of the different clusters. To solve this problem, the actual number of clusters can be determined by the Akaike Information Criteria, in which the probability of log-likelihood of the model is increasingly penalized with the number of parameters.

Abbreviations used

References

[1] EzWeb Mash-up platform. http://ezweb.morfeo-project.org/lng/en.

[2] Imagenio Description. http://en.wikipedia.org/wiki/Imagenio.

[3] Wireshark Net Sniffer. http://www.wireshark.com

[4] LANSurveyor. http://www.solarwinds.com/products/LANsurveyor/.

[5] Monitoring Genie. http://www.jilroy.com/MP Introduction.htm.

[6] Method and System for Optimizing the Performance of a Network. http://www.wipo.int/pctdb/en/wo.jsp?IA= EP2001012374 & WQ = 2003037018 & DI SPLAY = CLAIMS.

[7] Network Analyzing Method and Apparatus for Optimal Performance of Network and a Recording Medium Having Programs to Conduct Said Methods. http://www.wipo.int/pctdb/en/wo.jsp?IA=KR2002002302&wo=2003055251 & DI SPLAY = DESC.

[8] Van Grembergen; R.W. Saull; S. De Haes; “Linking the IT Balanced Scorecard to the Business Objectives at a Major Canadian Financial Group,” Journal of Information Technology Cases and Applications, 2003.

[9] Malina MA, Selto FH. Causality in Performance Measurement Model. Social Science research Networks. 2004. http://papers.ssrn.com/sol3/papers.cfm?abstract id = 488144.

[10] Chan APC, Scott D, Chan APL. Factors Affecting the Success of a Construction Project. J. Constr. Engrg and Mgmt. 2004: 1 (130): 153-155.

[11] Semi-automatic system with an iterative learning method for uncovering the leading indicators in business processes. United States Patent Application 20080201397.

[12] Liu Y, Arnold A, Abe N. Temporal Causal Modeling with Graphical Granger Methods. Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining. 2007

[13] Hui Z, Haifeng C, Guofei J, Xiaoqiao M, Yoshihira K. Fast statistical relationship discovery in massive monitoring data. IEEE Conference on Computer Communications Workshops, 2008.

[14] Pearson K. On Lines and Plans of Closest Fit to Systems and / or customer base grows the submitted book types are very of Points in Space. Philosophical Magazine, 2 (6): 559-572, likely to change dramatically.

[15] The Application of Trial Methodologies for Measuring and Verifying GPRS KPIs in the Railway Environment. 2004. http://www.bechtel.com/communications/assets/ fi les / TechnicalJournals / Septe mber2004 / Article2.pdf.

[16] Rhee CH, Park SJ, Lee YH, Kwon B, Yu JH. Deployment Method for Real-Time Wireless Network Optimizer in CDMA Network. http: //research.ac.upc.ed u / EW2004 / papers / 37.pdf.

[17] Kaplan RS, Norton, DP. The Balanced Scorecard -Translating Strategy Into Action, Harvard Business School Press, Boston, Massachusetts, 1996.

[18] Leidner, DE., Elam, JJ. Executive Information Systems: Their Impact on Executive Decision Making. Journal of Management Information Systems, 1994; 10: 139-156.

[19] TGG International White Paper. About Network Performance Monitoring & Benchmarking In a Fast Changing Environment. http://www.ttgint.com/UserFiles/File/Telecom/White%20paper-About%20network% 20performance% 20monitoring% 20and% 20Benchmarking.pdf.

[20] Laiho, J., Kylväjä, M., Hoglund, A. Utilization of Advanced Analysis Methods in UMTS Networks. Proceedings of the Vehicular Technology Conference 2002: 726-730. http: //lib.tkk. fi / Diss / 2002 / isbn9512259028 / article9.pdf.

[21] Kreher, R. UMTS Performance Measurement: A Practical Guide to KPIs for the UTRAN Environment. July 2008. ISBN-10: 0-470-03249-9.

[22] Golub, GH., Van Loan, CF. Matrix Computations 3rd ed., Johns Hopkins University Press, Baltimore, 1996.

[23] Keogh, E J., Chakrabarti, K., Pazzani, MJ., Mehrotra, S. Dimensionality reduction for fast similarity search in large time series databases. Knowledge and Information Systems, 2001; 3: 263-286.

[24] Time Series Analysis Overview http://www.statsoft.com/textbook/sttimser.html.

[25] Papadimitriou, S., Jimeng S., Yu, PS. Local Correlation Tracking in Time Series. Proceedings of the Sixth International Conference on Data Mining, 2006; 18: 456-465.

[26] Peter Robinson, 2007. Correlation testing in time series, spatial and cross-sectional data. CeMMAP working papers CWP01 / 07, Center for Microdata Methods and Practice, Institute for Fiscal Studies.

[27] Zhang, P., Huang, Y., Shekhar, S., Kumar, V. Correlation Analysis of Spatial Time Series Datasets: A Filterand-Refin ne Approach. Proc. of the 7thPAKDD 2003.

[28] Sakurai, Y., Papadimitriou, S., Faloutsos, C. BRAID: Stream Mining through Group Lag Correlations, SIGMOD, 599-610, 2005.

[29] Myers, CS., Rabiner, LR. A comparative study of several dynamic timewarping algorithms for connected word recognition. The Bell System Technical Journal, 1981; 60: 1389-1409.

[30] Granger, CWJ. Investigating causal relations by econometric models and cross-spectral methods. Econometrics, 1969; 37: 424-438.

[31] Knockaert, L., Dhaene, T. Causality Determination and Time Delay Extraction by Means of the Eigenfunctions of the Hilbert Transform. Proceedings of the 12th IEEE Workshop on Signal Propagation on Interconnects 2008: 1-4.

Claims (9)

1. Method for the determination in real time of the causes of the deterioration of the quality of the service in convergent networks, characterized in that it comprises:

to.

reduce the dimension of incoming events;

b.

make an orderly correlation of these events to find the causality relationships between the prioritized and high correlation data, building at least one causality distance matrix, and obtaining the potential causality between metrics, and

C.

grouping the data into clusters, formed on the basis of a causality relationship, forming data clusters with a high causal relationship based on said previously constructed causality distance matrix,

2.

Method for determining in real time the causes of the deterioration of the quality of the service in convergent networks, according to claim 1, characterized in that in step a) PCA and SVD techniques are applied to reduce the dimensionality of the massively generated data, uses heuristic peak selection (greedy vertex) based on the degree of the peak, and random sampling algorithms are applied to generate an estimated range (rapid indexing) at the peaks or measurement points, based on the edges (existing statistical relationships ) that affect them.

3.

Method for determining in real time the causes of deterioration of the quality of the service in convergent networks, according to claim 1, characterized in that in said step b) the causality of Granger is applied, assessing the causality by checking the causal direction of several series Temporary

Four.

Method for determining in real time the causes of deterioration of the quality of service in convergent networks, according to claim 1, characterized in that said distance matrix constructed in step b) provides the hierarchical relationships by means of a list of potentially operable indicators or dendrogram

5. Method for determining in real time the causes of deterioration of the quality of service in convergent networks, according to claim 1, characterized in that the real number of clusters in step c) is determined by the Akaike Information Criterion , in which the probability of log-likelihood of the model is increasingly penalized with the number of parameters.

6.

Method for the determination in real time of the causes of the deterioration of the quality of the service in convergent networks, according to any one of the preceding claims, characterized in that it directly uses the data of the networks that are in operation to have the adaptability in Quasi-real time to changes in these networks.

7.

Method for the determination in real time of the causes of the deterioration of the quality of the service in convergent networks, according to any one of the preceding claims, characterized in that a change is made in the metrics that are identi fi ed as new services are added to a network .

8.

Method for determining in real time the causes of deterioration of the quality of the service in convergent networks, according to claim 1, characterized in that prior to said stage a) the raw data and the integrated data of said events are arranged together Incoming

SPANISH OFFICE OF THE PATENTS AND BRAND Application no .: 200900406 SPAIN Date of submission of the application: 13.02.2009 Priority Date: REPORT ON THE STATE OF THE TECHNIQUE 51 Int. Cl.: H04W24 / 00 (2009.01) RELEVANT DOCUMENTS

Category

Documents cited Claims Affected

To

WO 03055251 A1 (SK TELECOM CO LTD et al.) 03.07.2003, summary; page 3, line 15 - page 5, line 16; page 6, line 19 - page 12, line 8; Figure 3. US 2004266442 A1 (FLANAGAN ADRIAN et al.) 30.12.2004, summary; paragraphs 18-31,38-40; Figure 1. 1-8 1-8

Category of the documents cited X: of particular relevance Y: of particular relevance combined with other / s of the same category A: reflects the state of the art O: refers to unwritten disclosure P: published between the priority date and the date of priority submission of the application E: previous document, but published after the date of submission of the application

This report has been prepared • for all claims • for claims no:

Date of realization of the report 24.11.2011

Examiner J. Santaella Vallejo Page 1/4

REPORT OF THE STATE OF THE TECHNIQUE Application number: 200900406 Minimum documentation sought (classification system followed by classification symbols) H04W Electronic databases consulted during the search (name of the database and, if possible, terms of search used) INVENES, EPODOC, INTERNET State of the Art Report Page 2/4  WRITTEN OPINION Application number: 200900406 Date of Written Opinion: 24.11.2011 Statement

Novelty (Art. 6.1 LP 11/1986)

Claims Claims 1-8 IF NOT

Inventive activity (Art. 8.1 LP11 / 1986)

Claims Claims 1-8 IF NOT

The application is considered to comply with the industrial application requirement. This requirement was evaluated during the formal and technical examination phase of the application (Article 31.2 Law 11/1986).  Opinion Base.- This opinion has been made on the basis of the patent application as published. State of the Art Report Page 3/4  WRITTEN OPINION Application number: 200900406 1. Documents considered.- The documents belonging to the state of the art taken into consideration for the realization of this opinion are listed below.

Document

Publication or Identification Number publication date

D01

WO 03055251 A1 (SK TELECOM CO LTD et al.) 03.07.2003

D02

US 2004266442 A1 (FLANAGAN ADRIAN et al.) 12-30-2004

2. Statement motivated according to articles 29.6 and 29.7 of the Regulations for the execution of Law 11/1986, of March 20, on Patents on novelty and inventive activity; quotes and explanations in support of this statement For the realization of this written opinion the claims contained in the application have been used. The request presents a method of determining in real time the causes of the deterioration of the quality of the service in convergent networks. To do this, it takes a series of indicators and processes them through the following steps: it reduces the dimensionality, correlates them to obtain a distance matrix of causal relationships and with this causality criterion they are grouped. The document of the state of the art closest to the invention is D01 deals with a method and an apparatus for analyzing networks independently of the propagation means to evaluate the operation of the network. For this, it analyzes the data that is collected in real time (summary) of the type of key performance indicators (KPI) and performs statistical processes (page 3, lines 23-29), among these analyzes is the correlation (page 4, line 16 ). The difference between D01 and the application is the concretion of the statistical methods. In D01, it is not specific if there is a reduction in the dimensionality of the data collected or if a correlation is made on these data to find the causality distances since it indicates that statistical analyzes or correlations are performed in general. The expert in the field in view of this document could not reach D01 since the range of techniques within the "statistical analysis" is very large and each technique has its advantages and disadvantages. Document D02 discloses a method and a system to optimize network performance. The invention collects data from the key performance indicators (KPI). These parameters are configured to establish a cost function, some parameter is varied and compared with the previously calculated function. There are several differences between D02 and the request: D02 is focused on improving network conditions, while the request is aimed at determining the causes of deterioration. The next difference is found in the second step of the method that instead of calculating a causality correlation as in the request a cost function is established and this is varied to establish the most similar parameters. Finally D02 does not establish a cluster procedure if it does not simply notify. None of the cited documents, taken alone or in combination, disclose the invention defined in claims 1-8 Furthermore, in the cited documents there are no suggestions that direct the person skilled in the art towards the invention defined by the claims. Thus, the invention claimed in claims 1-8 is, with reference to documents D01 and D02, is new and is considered to imply inventive activity as set forth in articles 6 and 8 of Patent Law 11/86. State of the Art Report Page 4/4