Distributed network service system performance margin analysis method

Abstract

The invention relates to the technical field of distributed network service performance evaluation, in particular to a distributed network service system performance margin analysis method, which comprises the following steps: regularly acquiring QPS, CPU and RT data of each component, fitting to obtain a CPU-QPS curve and an RT-QPS curve, and determining inflection point response time of the RT-QPS curve of each component; obtaining a response time threshold value of each component based on the flow routing probability between the components, the inflection point response time of each component and an entrance response time upper limit, and determining a flow threshold value of each component according to the CPU-QPS curve; adjusting the slope coefficient of the RT-QPS curve based on the proportional relation between the calculation resource unit number and the expected adjustment resource unit number, and determining whether the current component is the main margin bottleneck or not based on the inflection point flow and the slope coefficient of the RT-QPS curve before and after adjustment; the method can dynamically judge the margin and judge the performance improvement degree of the component after the resource is changed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed network service performance evaluation, and particularly relates to a distributed network service system performance margin analysis method. BACKGROUND

[0002] With the continuous expansion and complexity of the Internet service system, the performance margin management of each component in the distributed network service system becomes an important content to ensure the stability of the service. The commonly used capacity analysis methods at present mainly include: a method based on machine learning regression combined with artificial definition of threshold, a method of detecting system component margin through stress testing, and a method of manually setting a fixed CPU monitoring threshold. For example, machine learning regression combined with RT (response time) and Util (utilization) thresholds are used, but the setting of these thresholds highly depends on artificial experience, and there are generally problems of inaccuracy and unreliability. For another example, the margin of bottleneck components is detected by using stress testing and manually setting RT threshold, but this method also needs artificial experience, and stress testing can only cover part of the core bottleneck components in the system, and it is difficult to fully reflect the overall performance margin. Another way is to manually set a fixed CPU monitoring threshold, such as a CPU monitoring threshold of not exceeding 30%, which is simple and easy to implement, but it can easily lead to underestimation of the actual margin of system performance, and thus cause resource waste and increase in operation and maintenance costs.

[0003] In view of the above problems, some patent technologies propose a turning point detection method based on an approximate queue model and Bayesian fitting for estimating the performance limit value of the component. However, in actual application, this type of method relies on the approximate queue model and simple Bayesian proportion for turning point detection, and the estimated component capacity limit value does not really establish a systematic correlation with the performance SLA (Service Level Agreement) of the system entrance or the global performance margin threshold, which is not convenient for realizing the allocation and dynamic adjustment of the system-level margin. At the same time, the attribution of the performance bottleneck is also difficult to accurately depict through the simple proportion relationship between the capacity limit value and the Bayesian prior, and the performance improvement effect of the whole system when the component resource changes cannot be directly described.

[0004] In summary, the current capacity evaluation and allocation technologies all have the problems of strong dependence on artificial experience, insufficient system-level optimization capability, and inaccurate performance bottleneck positioning. Therefore, it is urgent to propose a margin analysis method that can establish the correlation between the component and the system-level performance margin, realize the dynamic allocation of the margin, and accurately determine the attribution of the bottleneck, so as to improve the reliability and resource utilization efficiency of the network service system. SUMMARY

[0005] In view of the above problems, the present application provides a distributed network service system performance margin analysis method, which solves the technical problems that the prior art cannot dynamically determine the margin and cannot determine the performance improvement degree after changing the resources of the components.

[0006] The present application provides a distributed network service system performance margin analysis method, which comprises the following steps:

[0007] Step S1, periodically collecting the traffic, computing resource utilization rate and response time data of each component of the distributed network service system, fitting to obtain the computing resource utilization rate and traffic relationship curve and the response time and traffic relationship curve, and determining the inflection point response time of the response time and traffic relationship curve of each component;

[0008] Step S2, obtaining the response time threshold of each component based on the traffic routing probability between each component, the inflection point response time of each component and the upper limit of the entry response time, and determining the traffic threshold of each component according to the computing resource utilization rate and traffic relationship curve;

[0009] Step S3, when the component traffic exceeds the traffic threshold, obtaining the current component computing resource unit number and the expected adjustment resource unit number;

[0010] Step S4, adjusting the slope coefficient of the response time and traffic relationship curve based on the proportional relationship between the computing resource unit number and the expected adjustment resource unit number, and determining whether the current component is a main margin bottleneck based on the inflection point traffic and the slope coefficient of the response time and traffic relationship curve before and after adjustment.

[0011] Preferably, step S1 specifically comprises:

[0012] Step S1-1, collecting the traffic, computing resource utilization rate and response time of each component of the distributed network service system every fixed time to obtain sampling data;

[0013] Step S1-2, obtaining the computing resource utilization rate and traffic relationship curve based on the multi-order polynomial fitting of the relevant traffic, and determining the slope coefficient of the computing resource utilization rate and traffic relationship curve;

[0014] Step S1-3, performing Bayesian fitting on the slope coefficient of the computing resource utilization rate and traffic relationship curve as prior information to obtain the response time and traffic relationship curve;

[0015] Step S1-4, determining the transition point of the increase of the response time in the response time and traffic relationship curve of each component, and determining the corresponding response time at the transition point as the inflection point response time.

[0016] Preferably, in step S1-3, the expression of the response time and traffic relationship curve is:

[0017] RT = f(QPS) = P(QPS) / (QPS-b) + c

[0018] wherein RT represents the response time, f(QPS) represents a function of QPS, P(QPS) represents a polynomial of an arbitrary order related to QPS, and b and c are fitting parameters;

[0019] The expression of the slope coefficient of the response time and traffic relationship curve is:

[0020]

[0021] wherein d is the slope coefficient, and is the maximum and minimum of the QPS of component i in the observable range;

[0022] The expression of the loss function of the Bayesian fitting is:

[0023]

[0024] or

[0025] wherein L is the loss function of the Bayesian fitting, λ is the weight of prior information, min(·) represents taking the minimum value, is the parameter function in optimization.

[0026] Preferably, step S2 specifically comprises:

[0027] Step S2-1, obtaining the upper limit of the system entry response time RT SLA ;

[0028] Step S2-2, obtaining the traffic routing probability P j between each component and the path traffic routing probability from the system entry to each component J is the component dependency relationship index, and i is the component index;

[0029] Step S2-3, obtaining the response time threshold of each component from the ratio α of the upper limit of the entry response time RT SLA and the entry component inflection point response time , the path traffic routing probability from the system entry to each component , and the response time inflection point of the component

[0030] Step S2-4, determining the traffic threshold of each component according to the response time threshold of each component and the inverse function of the response time and traffic relationship curve ​

[0031] Preferably, in step S2-2, the flow routing probability between each component is obtained by calculating the ratio of the total flow of the upstream component to its flow to the downstream component.

[0032] The path flow routing probability of the system entrance to each component is expressed as:

[0033]

[0034] wherein, l is the path index, L is the total number of paths, the path flow probability of the entrance to component i is represents the flow routing probability of path l of component i, path i , l represents the set of all component dependencies of path l of component i.

[0035] Preferably, in step S2-3, the response time threshold is expressed as:

[0036]

[0037] In step S2-4, is expressed as:

[0038]

[0039] wherein, represents the flow threshold of component i, represents the inverse function of the RT-QPS relationship curve of component i.

[0040] Preferably, step S4 specifically comprises:

[0041] Step S4-1, adjusting the slope coefficient by calculating the ratio of the resource unit number R and the expected adjusted resource unit number R';

[0042] Step S4-2, obtaining the adjusted response time and flow relationship curve f'(QPS) by the adjusted slope coefficient d' and the response time and flow relationship curve f(QPS); respectively obtaining the inflection points QPS knee and QPS knee '

[0043] Step S4-3, based on QPS knee , QPS knee ', d and d', performing threshold judgment to obtain whether the current component is a main margin bottleneck.

[0044] Preferably, in step S4-2, the expression of the adjusted slope coefficient d' is:

[0045]

[0046] The expression for f′(QPS) is:

[0047]

[0048] Where a, b, and c are the fitting parameters.

[0049] Preferably, step S4-3 specifically includes:

[0050] judge The relationship between the value of β and the value of β, where β is the threshold for judging the impact of resource adjustments on the performance of the current component;

[0051] like Then determine that the current component is the main margin bottleneck. Then the downstream component of the current component is determined to be the main margin bottleneck.

[0052] Compared with the prior art, the present invention has at least the following beneficial effects:

[0053] (1) This invention realizes a system-level performance margin threshold calculation method by calculating the fitting curve of the relationship between the response time of each component and the traffic. Compared with the existing method that only estimates the inflection point of a single component, this method can decompose the system entry SLA constraint layer by layer to each component, and combine information such as traffic routing probability and inflection point response time to comprehensively determine the performance margin of each component. This ensures that the margin threshold allocation of all system components is consistent with the overall service goal of the system, effectively improving the scientificity and accuracy of performance margin analysis and optimization of distributed network service systems.

[0054] (2) This invention overcomes the shortcomings of existing technologies that simply describe the upstream and downstream dependencies between components by using a quantitative modeling method. By incorporating the impact of adjusting computing resource units on the response time performance of each node into the system model, it can describe the specific impact of downstream resource changes on upstream and even the entire system performance. This allows the system to consider not only the bottlenecks of individual components when planning resource allocation, but also the overall synergistic effect after resource adjustments of each component, thereby achieving automatic, proactive, and refined resource planning and performance optimization of the distributed system.

[0055] (3) By adjusting the slope coefficient of the response time curve after adjusting computing resources, this invention can dynamically assess the impact of resource changes on performance bottlenecks. This allows for the quantification of the impact of computing resource changes on the performance of a network service in a distributed system, enabling proactive design and planning of system resource allocation based on performance margin estimation. Attached Figure Description

[0056] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0057] Figure 1 The flowchart of the performance margin analysis method for distributed network service systems provided by this invention is shown.

[0058] Figure 2 A flowchart for the unified estimation of the full component margin threshold provided by this invention.

[0059] Figure 3 This is a schematic diagram of the routing ratio of the network service system provided by the present invention.

[0060] Figure 4 A schematic diagram of the relationship curves of the network service system provided by the present invention.

[0061] Figure 5 This is a schematic diagram of the inflection point response time of the network service system provided by the present invention.

[0062] Figure 6 Flowchart of the performance impact judgment model provided by this invention.

[0063] Figure 7 This is a schematic diagram showing the relationship between component performance before and after resource adjustment, as provided by the present invention. Detailed Implementation

[0064] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0065] The following lists the meanings of the proper nouns used in this invention:

[0066] Network service: A network service can be defined as a system node that receives requests and returns responses.

[0067] System: A system can be defined as a system composed of multiple network services.

[0068] Component: A component can be defined as a collection of network services in a system that perform a unified service function.

[0069] Component dependency: Component dependency can be defined as the dependency relationship between components in a system, such as A->B->C, where A depends on B, B depends on C, A is an upstream of B, and B is an upstream of C. In network service systems, dependency relationships can be defined as the upstream component's request processing requiring all or part of the response content from the downstream component. A direct dependency relationship between two components can be defined as a dependency edge.

[0070] Flow (QPS): Flow can be defined as the flow of traffic to components in a system.

[0071] Routing probability: Routing probability can define the probability that a request from an upstream component will require a response from a specific downstream component.

[0072] Computational resource utilization (Util, utilization, CPU, etc.): Computational resource utilization can be defined as the utilization rate of the computational resources of components in the system.

[0073] Response time (RT): Response time can be defined as the response time of a component in a system, that is, the time from when the relevant network service receives a request to when it returns a response.

[0074] Component performance: Component performance is a metric related to component reliability, specifically, the extent to which performance degradation may lead to insufficient component reliability. Current component performance metrics primarily consider response time, which may deteriorate with increasing traffic.

[0075] Component margin: Component margin can be defined as the distance between the performance index of a component in the system and the performance threshold of that component. If the performance index exceeds the threshold, the system is considered unreliable.

[0076] This invention first estimates the margin thresholds of system components in a distributed network service system, then uniformly analyzes the margin thresholds of all system components, and combines the overall system perspective with the component fitting equations to comprehensively determine the performance margin thresholds linked to the network service entry SLA. It also provides a model for judging the impact of upstream and downstream resource changes on the response time performance of the current node, enabling the proactive design and planning of system resource allocation based on performance margin estimation.

[0077] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment, such as... Figure 1 As shown, a method for performance margin analysis of a distributed network service system is disclosed, and the specific implementation steps are as follows:

[0078] Step S1: Periodically collect traffic, computing resource utilization and response time data of each component of the distributed network service system, fit the computing resource utilization and traffic relationship curves as well as the response time and traffic relationship curves, and determine the inflection point response time of each component's response time and traffic relationship curves.

[0079] The distributed network service system of this invention is a system composed of multiple network services. Each set of network services that performs a unified service function is considered a component, such as a microservice, a database, or a message queue. Specific steps include:

[0080] (1) The present invention collects the traffic (QPS), computing resource utilization (CPU utilization) and response time (RT) of each component of the distributed network service system at fixed intervals to obtain the corresponding sampling data.

[0081] In some embodiments, OpenTelemetry's metrics tool can be used to sample and obtain, for example, RT, QPS, and CPU utilization data at 15-second intervals.

[0082] (2) By fitting the RT-QPS relationship curve using RT and QPS data, a polynomial of any order related to QPS can be used for fitting, and the expression is:

[0083] RT=f(QPS)=P(QPS) / (QPS-b)+c

[0084] Where RT represents the response time, f(QPS) represents a function of QPS, P(·) represents a polynomial of arbitrary order related to QPS in the RT-QPS relationship curve, and b and c are fitting parameters.

[0085] In some embodiments, P(QPS) can be a first-order polynomial, expressed as: P(QPS)=θ*QPS+φ, where θ and φ are first-order polynomial fitting parameters.

[0086] (3) By fitting the CPU-QPS relationship curve using CPU utilization and QPS data, a polynomial of any order related to QPS can be used for fitting, and the expression is:

[0087] CPU = g(QPS)

[0088] Where g(·) is a polynomial of arbitrary order relating QPS to the CPU-QPS relationship curve.

[0089] In some embodiments, g(QPS) can be a second-order polynomial, expressed as: g(QPS)=γ*QPS 2 +δ*QPS+∈, where γ, δ,∈ are second-order polynomial fitting parameters.

[0090] (4) Calculate the slope of the resource utilization rate and flow rate relationship curve as prior information for Bayesian fitting to obtain the response time-flow rate RT-QPS relationship curve. Specific steps include:

[0091] The slope of the CPU-QPS relationship, i.e., the slope of the CPU-QPS relationship curve within the observable range, is determined by the following expression:

[0092]

[0093] Where d is the slope of the CPU-QPS relationship curve. and The maximum and minimum QPS of component i within the observable range

[0094] The loss function L for Bayesian fitting is defined as follows:

[0095]

[0096] Where L is the loss function of Bayesian fitting, and λ is the prior information weight, which can be any positive real number, with a recommended value of 0.002. Let be the parameter function in the optimization, and let its optimization objective be min θ L.

[0097] As another technical solution for calculating the loss function, this invention can also use uncertainty theory instead of probability theory to perform Bayesian fitting, and the expression for the loss function L is:

[0098]

[0099] Here, min(·) represents taking the minimum value.

[0100] Through the above steps, this invention determines the RT-QPS relationship curve, as follows: Figure 4 As shown, the QPS-RT curve typically exhibits the following behavior: initially, as the request volume (QPS) increases, the response time (RT) grows steadily; however, once the request volume reaches a certain point, the response time increases rapidly. This turning point where the response time increases rapidly is called the inflection point.

[0101] In some embodiments, to accurately identify the inflection point, the following method can be used: First, calculate the change in the RT growth rate between two adjacent sampling points. When a sudden increase in the RT growth rate corresponding to a certain QPS value is detected relative to previous sets of data, i.e., the growth rate exceeds a preset threshold, then that QPS value is determined as an inflection point. Typically, the preset threshold can be set empirically based on historical performance test data. For example, when the response time growth rate significantly exceeds the average rate of the previous stage, an inflection point can be identified. Alternatively, the Kneed algorithm can be used to solve the relationship curve to directly obtain the inflection point location. Using the above method, the critical point of system performance change under different loads can be obtained objectively and accurately, providing a reliable basis for subsequent capacity planning and system optimization.

[0102] Then, the RT value at the inflection point is determined from the QPS-RT curve as the inflection point response time, and the QPS value at the inflection point is taken as the inflection point flow rate.

[0103] Step S2: Based on the traffic routing probability between components, the inflection point response time of each component, and the upper limit of the ingress response time, obtain the response time threshold of each component, and determine the traffic threshold of each component according to the calculation resource utilization and traffic relationship curve.

[0104] In this step, the present invention maps the system entry performance SLA to the specific performance monitoring threshold of each component from the perspective of the entire system. This process considers the performance characteristics of each component and the routing probability between components, thus avoiding the unreliability of manually set thresholds. Figure 2 As shown, the specific steps are as follows.

[0105] (1) Obtain the upper limit of the system ingress response time RT specified by the SLA. SLA The upper limit of the entry response time is the overall system margin threshold and is determined by the system entry performance SLA. For example, if the SLA of an e-commerce website is to ensure that 99% of requests are completed within 100ms, then the upper limit of the system entry response time is 100ms.

[0106] (2) Obtain the traffic routing probability between each component and the path traffic routing probability from the system entry to each component.

[0107] The probability of traffic routing between system components is represented by P. j j = 1, 2, ..., m, where j is the component dependency index and m is the total number of component dependencies. For example, if the system has components A, B, and C, and all dependencies are A->B->C, then m = 2, P 1 Let P be the path traffic routing probability from A to B. 2 The probability of routing traffic along the path from B to C.

[0108] In some embodiments, traffic routing probabilities can be obtained using tracing tools such as OpenTelemetry. The routing probability can be obtained by calculating the ratio of the total traffic to the upstream component to the traffic to the downstream component.

[0109] The path traffic routing probability from the system entry point to each component is represented as follows: Where i is the component index and n is the total number of components. For example, if the system has components A, B, and C, then... The flow rate of A / the total flow rate of A The flow rate of A / the total flow rate of A.

[0110] Assuming the entry component has L paths l leading to component i, then Where l is the path index, L is the total number of paths, and the path traffic probability from the entry point to component i is... This represents the probability of traffic routing to path 1 for component i. path i，l This represents the set of all component dependencies along path 1 of component i.

[0111] (3) Obtain the inflection point response time of the entry component Upper limit of ingress response time RT SLA The ratio, expressed as:

[0112]

[0113] Where α represents the ratio of the inflection point response time of the entry component to the upper limit of the entry response time.

[0114] The ratio of the inflection point response time of the entry component to the SLA upper limit α, and the path routing probability of each component are used. Component response time inflection point Get the response time threshold of the component The expression is:

[0115]

[0116] (4) The function and inverse function of the relationship between the response time RT of each component and the flow rate QPS are expressed as fi(QPS). Based on the response time threshold of each component The traffic threshold for each component is determined by the following expression:

[0117]

[0118] in, This represents the traffic threshold for component i. This represents the inverse function of the RT-QPS relationship curve of component i.

[0119] like Figure 3 , Figure 5 As shown in the example, this invention illustrates a distributed system consisting of five sub-components, each with its own response time inflection point and traffic routing probability. With a system ingress SLA of 2 seconds and an ingress RT inflection point value of 1.8, the traffic threshold for component Q5 is calculated as follows:

[0120] 1. Obtain the RT-QPS relationship function f5(QPS) of component Q5 and its response time RT inflection point value of 0.6s;

[0121] 2. Obtain the maximum response time limit for the system entry point, which is 2 seconds.

[0122] 3. The path traffic routing probability for Q1->Q2 is 0.8, the path traffic routing probability for Q1->Q3 is 0.4, the path traffic routing probability for Q2->Q5 is 0.3, and the path traffic routing probability for Q3→Q5 is 0.6.

[0123] 4. Obtain the ratio of the inflection point response time of the entry component to the upper limit of the SLA.

[0124] 5. Obtain the path traffic routing probability from the system entry point to component Q5.

[0125] 6. Obtain the margin threshold of component Q5

[0126] 7. The RT-QPS relationship of Q5 is the inverse function. The inverse function is generally a numerical solution operation for fitting curves;

[0127] 8. Obtain the traffic threshold for component Q2.

[0128] Step S3: When the component traffic exceeds the traffic threshold, obtain the current number of component computing resource units and the expected number of resource units to be adjusted;

[0129] In the previous steps, the present invention determined the traffic threshold of each component to determine whether there is a performance problem in the current system. In this step, it can be determined whether the traffic of each component exceeds the traffic threshold. When the traffic of a component exceeds the traffic threshold, the computing resources of the current component are adjusted.

[0130] The current number of computing resource units is R, and the expected number of resource units to be adjusted is set to R′. For example: the current component has 100 microservice instances with 8-core CPUs, and the expected adjustment is to 120 microservice instances of the same specifications, then R = 100, R′ = 120.

[0131] The expected adjustment judgment of this invention can also be regarded as a method for judging whether the margin threshold has actually been exceeded, rather than just the expected impact. For example, when the traffic of node 1 exceeds the estimated threshold, but it is judged that the resource adjustment has little impact on the traffic margin threshold of node 1, then it is possible that it is not node 1 that has exceeded the margin threshold, but rather that a downstream node has exceeded the threshold and affected node 1. Therefore, this invention makes a judgment in subsequent steps.

[0132] Step S4: Adjust the slope coefficient of the response time and flow rate relationship curve based on the ratio of the number of computational resource units to the expected number of adjusted resource units. Determine whether the current component is the main margin bottleneck based on the inflection point flow rate and slope coefficient of the response time and flow rate relationship curve before and after adjustment.

[0133] like Figure 6 As shown, this step specifically includes:

[0134] (1) Obtain the slope coefficient d of the linear relationship between the current component resource utilization rate and the flow rate. The slope coefficient d has been obtained in step S1.

[0135] (2) The slope coefficient d′ is obtained by adjusting the ratio of R to R′. The expression is: For example, when R = 100 and R' = 120...

[0136] (3) The adjusted RT-QPS relationship curve f′(QPS) is obtained by using the relationship curve f(QPS) between the adjusted slope coefficient d′ and RT-QPS. Wherein,

[0137]

[0138] Where k is the reciprocal of the slope d of the linear relationship between CPU utilization and traffic QPS, k = 1 / d. And k' = 1 / d'.

[0139] Obtain the inflection point QPS of the RT-QPS relationship curve before and after adjustment. knee With QPS knee ′, inflection point QPS knee With QPS knee f(QPS) and f'(QPS) are obtained by solving for f(QPS) and f'(QPS) using the kneed algorithm.

[0140] (4)Judgment The relationship between the value of β and the value of β, where β is the threshold for judging the impact of resource adjustments on the performance of the current component, such as β = 0.1.

[0141] like Then determine that the current component is the main margin bottleneck. Then the downstream component of the current component is determined to be the main margin bottleneck.

[0142] Figure 7 This shows the component performance when the current component is the primary margin bottleneck.

[0143] The current component is the main bottleneck, so adjusting its resources will improve its performance. The performance of the component changed significantly before and after the resource adjustment. The inflection point before the resource adjustment was 40.4 QPS, the predicted inflection point after the adjustment was 29.4 QPS, and the actual performance inflection point after the resource change was 31 QPS, consistent with the prediction. Therefore, the current component is indeed the bottleneck.

[0144] If the downstream component is the main bottleneck, then adjusting the current component resources will improve the current component performance, with only a small change in component performance before and after the resource adjustment. The inflection point before and after the adjustment was 30.0 QPS, while the actual performance inflection point after the resource change was 29.2 QPS, which is consistent with the prediction. The downstream component is the bottleneck, and adjusting the current component resources will not improve the current component performance.

[0145] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for performance margin analysis of a distributed network service system, characterized in that, Includes the following steps: Step S1: Periodically collect traffic, computing resource utilization and response time data of each component of the distributed network service system, fit the computing resource utilization and traffic relationship curves as well as the response time and traffic relationship curves, and determine the inflection point response time of each component's response time and traffic relationship curves. Step S2: Based on the traffic routing probability between components, the inflection point response time of each component, and the upper limit of the ingress response time, obtain the response time threshold of each component, and determine the traffic threshold of each component according to the calculation resource utilization and traffic relationship curve. Step S3: When the component traffic exceeds the traffic threshold, obtain the current number of component computing resource units and the expected number of resource units to be adjusted; Step S4: Adjust the slope coefficient of the response time and flow rate relationship curve based on the ratio of the number of calculated resource units to the expected number of adjusted resource units. Determine whether the current component is the main margin bottleneck based on the inflection point flow rate and slope coefficient of the response time and flow rate relationship curve before and after adjustment.

2. The performance margin analysis method for distributed network service systems according to claim 1, characterized in that, Step S1 specifically includes: Step S1-1: Collect the traffic, computing resource utilization and response time of each component of the distributed network service system at fixed intervals to obtain sampling data; Step S1-2: Based on the multi-order polynomial fitting of the relevant flow rate, obtain the curve of the relationship between the calculated resource utilization rate and the flow rate, and determine the slope coefficient of the curve of the relationship between the calculated resource utilization rate and the flow rate. Step S1-3: Use the slope coefficient of the calculated resource utilization rate and flow relationship curve as prior information to perform Bayesian fitting and obtain the response time and flow relationship curve. Step S1-4: Determine the turning point where the response time increases in the response time-flow relationship curve of each component, and determine the response time corresponding to the turning point as the inflection point response time.

3. The performance margin analysis method for distributed network service systems according to claim 2, characterized in that, In steps S1-3, the expression for the response time-flow rate relationship curve is: RT=f(QPS)=P(QPS) / (QPS-b)+c Where RT represents the response time, f(QPS) represents a function of QPS, P(QPS) represents a polynomial of arbitrary order relating QPS, and b and c are fitting parameters. The expression for the slope coefficient of the curve relating resource utilization rate and flow rate is as follows: Where d is the slope coefficient. and The maximum and minimum QPS of component i within the observable range; The loss function expression for the Bayesian fitting is: Where L is the loss function of Bayesian fitting, λ is the prior information weight, and min(·) represents taking the minimum value. For the parameter function in the optimization.

4. The performance margin analysis method for distributed network service systems according to claim 3, characterized in that, Step S2 specifically includes: Step S2-1: Obtain the upper limit of system entry response time RT SLA ; Step S2-2: Obtain the traffic routing probability P between each component. j and the probability of path traffic routing from the system entry point to each component. j is the component dependency index, and i is the component index; Step S2-3, from the upper limit of the ingress response time RT SLA Inflection point response time of entry component The proportion α, the path traffic routing probability from each system entry point to each component and the inflection point of component response time Obtain the response time threshold of each component Step S2-4: Based on the response time threshold of each component The inverse function of the response time-flow relationship curve determines the flow threshold of each component.

5. The performance margin analysis method for distributed network service systems according to claim 4, characterized in that, In step S2-2, the traffic routing probability between the components is obtained by calculating the ratio of the total traffic of the upstream component to the traffic to the downstream component; The path traffic routing probability from the system inlet to each component The expression is: Where l is the path index, L is the total number of paths, and the path traffic probability from the entry point to component i is: The path represents the probability of traffic routing along path l of component i. i，l This represents the set of all component dependencies along path l of component i.

6. The performance margin analysis method for distributed network service systems according to claim 5, characterized in that, In steps S2-3, the response time threshold The expression is: In steps S2-4, The expression is: in, This represents the traffic threshold for component i. This represents the inverse function of the RT-QPS relationship curve of component i.

7. The performance margin analysis method for distributed network service systems according to claim 6, characterized in that, Step S4 specifically includes: Step S4-1: Adjust the slope coefficient by calculating the ratio of the number of resource units R and the expected adjustment number of resource units R′; Step S4-2: Obtain the adjusted response time-flow rate relationship curve f′(QPS) from the adjusted slope coefficient d′ and the response time-flow rate relationship curve f(QPS); obtain the inflection point QPS of f(QPS) and f′(QPS) respectively. knee With QPS knee ′ Step S4-3, based on QPS knee QPS knee Thresholds are used to determine whether the current component is the main margin bottleneck.

8. The performance margin analysis method for a distributed network service system according to claim 7, characterized in that, In step S4-2, the expression for the adjusted slope coefficient d′ is: The expression for f′(QPS) is: Where a, b, and c are the fitting parameters.

9. The performance margin analysis method for a distributed network service system according to claim 8, characterized in that, Step S4-3 specifically includes: judge The relationship between the value of β and the value of β, where β is the threshold for judging the impact of resource adjustments on the performance of the current component; like Then determine that the current component is the main margin bottleneck. Then the downstream component of the current component is determined to be the main margin bottleneck.

Images