Cloud service management platform resource configuration method and system

Abstract

The present invention discloses a cloud service management platform resource configuration method and system, which relates to cloud service resource configuration and management technology. The system includes a user demand prediction module, a capacity planning decision module and a cloud service computing resource online allocation module. By predicting the user demand of the system, capacity planning decisions are made, and then online allocation of cloud service computing resources is performed through decision simulation. The technical solution provided by the present invention can improve the allocation rate of the central processing unit of the public cloud resource management system and reduce the number of virtual machine outages.

Description

Technical Field

[0001] The present invention relates to cloud service resource configuration and management technology, and specifically to a resource configuration method and system based on modeling and simulation of a public cloud resource management platform, specifically modeling and simulation and causal analysis of a cloud resource management system, which is a technology that integrates demand forecasting, capacity planning decisions and online allocation. Background Art

[0002] Most of the existing work in the field of cloud service resource configuration and management technology focuses on studying the allocation of cloud computing resources and how to effectively supply cloud computing resources to meet users' quality of service (QoS) requirements. Reference [1] (Jain, Tarun, Jishnu Hazra. 2019. Hybrid cloud computing investment strategies. Production and Operations Management 28 (5) 1272–1284) studies the capacity matching and investment strategies of cloud resources from the user's perspective, and reference [2] (Arbabian, Mohammad Ebrahim, Shi Chen, Kamran Moinzadeh. 2021. Capacity expansions with bundled supplies of attributes: An application to server procurement in cloud computing. Manufacturing & Service Operations Management 23 (1) 191–209) explores the optimal procurement strategy of cloud service providers for two supply resource packages in response to the demand for two non-substitutable attributes (such as CPU and RAM). However, the existing cloud computing resource allocation technology has the following shortcomings: first, it only focuses on short-term (e.g., hourly, daily) decisions, and does not make comprehensive plans from the perspective of the entire life cycle; second, it only considers the expansion of cloud resources, but not the reduction strategy; third, long-term, short-term, macro, micro, and forecasting decisions are independent of each other. Therefore, the present invention aims to make comprehensive plans from the perspective of the entire life cycle, adjust the forecasting model according to the needs of the decision model, make long-term expansion and reduction decisions, and realize the integration of forecasting and decision-making. Summary of the invention

[0003] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a resource configuration method and system based on public cloud resource management platform modeling and simulation, which is a cloud resource management system resource planning and evaluation scheme. By simulating the cloud service resource configuration operation under a certain user demand curve, key indicators such as utilization rate, allocation rate, and hot spot risk are calculated, thereby exposing potential problems in corresponding supply chain decisions, allowing decision makers to compare multiple plans and make the best decision.

[0004] The technical solution of the present invention is as follows:

[0005] A resource configuration method based on modeling and simulation of a public cloud resource management platform, which predicts user demand of the system, makes capacity planning decisions, and then performs online allocation of cloud service computing resources through decision simulation; the system includes a supply side, a supply and demand matching side, and a decision side; the resource configuration method includes the following steps:

[0006] A. Carry out user demand forecast; and obtain the outlier boundary through residual normal analysis, identify the outliers and process them, and obtain the cloud server virtual machine demand forecast sequence;

[0007] The implementation method of demand forecasting adopts the time series forecasting model (PROPHET model);

[0008] A1. Data processing: The data involved in demand forecasting is cloud server demand data, including customer demand data for cloud server virtual machines. The data structure contains two columns: each date of the virtual machine and the number of virtual machines required for the corresponding date. First, perform data cleaning, simple analysis and other data processing work to convert the demand data into an input format acceptable to the model - that is, set it to the column name corresponding to the PROPHET forecasting model, and divide it into training set and test set data;

[0009] A2.PROPHET prediction: Substitute the training set data into the PROPHET prediction model to perform time series prediction on a long time scale. In specific implementation, the long scale is to predict one year backward every seven days, that is, to predict the demand value and its distribution of virtual machines for the next 365 days every seven days from the starting point of the test set data, and output the predicted value of demand as the mean and standard deviation of the normal distribution.

[0010] A3. Through residual normal analysis, determine whether to improve the model from the perspective of residuals:

[0011] The residual is obtained by subtracting the predicted value of the demand for virtual machines in the test set (i.e., the test set data) from the actual value of the demand for virtual machines. The residual is analyzed normally. If the residual obeys the normal distribution, it is considered that the model cannot be improved from the perspective of the residual, and the process jumps to A5. If the residual does not obey the normal distribution, it is processed. The specific process is as follows:

[0012] According to the law of large numbers, the initial outlier point limit is set to the absolute value of the minimum residual, and the outlier point limit is continuously lowered (for example, the initial outlier point limit value is reduced by 1 each time);

[0013] For the residual data under different abnormal point boundaries, calculate the kurtosis coefficient and skewness coefficient of the residual data. If both indicators are less than the set threshold (such as 1.96), the result of the normal analysis of the current residual obeys the normal distribution. If the virtual machine demand on a certain date is greater than the current abnormal point boundary value (that is, the virtual machine demand corresponding to each date, a total of two columns of data points), it is an abnormal point, and the abnormal point boundary is stopped; that is: if the skewness coefficient and kurtosis coefficient of the residual under a certain abnormal point boundary are both less than the set threshold 1.96, it is considered that the abnormal point boundary has been found, and the point where the virtual machine demand is greater than the abnormal point boundary is called an abnormal point;

[0014] A4. Perform outlier processing on the outliers found in A3;

[0015] Find the abnormal point and use the virtual machine demand corresponding to the previous moment as the virtual machine demand corresponding to the current abnormal point moment, and analyze whether the prediction accuracy has improved.

[0016] The calculation formula for prediction accuracy is 1-SMAPE (Symmetric Mean Absolute Percentage Error), where y i Represents the true value of the data, Represents the data prediction value, and n represents the number of data. Output the prediction sequence and pass it to the supply side and the supply and demand matching side; where the prediction accuracy is 1-SMAPE; i is the sequence number;

[0017] A5. Output the forecast sequence obtained by the PROPHET forecast model and outlier processing. The forecast sequence is the forecast data of virtual machine demand forecasted every other week for one year, and transmits it to the capacity planning decision module.

[0018] B. Make capacity planning decisions;

[0019] The present invention applies the multi-period newsboy model to cloud service resource management, that is, the implementation scheme of cloud service resource capacity planning decision is constructed and the multi-period newsboy model dynamic planning is adopted;

[0020] The life cycle of a cloud server consists of the growth phase, maturity phase, and decline phase. Decision makers need to expand capacity during the growth phase and maturity phase, and reduce capacity during the decline phase. The specific steps for capacity planning decisions are as follows:

[0021] B1. Model construction: Construct a multi-period newsboy model, define parameters such as cost composition, and clearly define the order of occurrence of sequential decision-making events in each period.

[0022] The life cycle of a cloud server consists of a growth period, a maturity period, and a decline period. It is necessary to expand the capacity of cloud servers during the growth and maturity periods, and reduce the capacity of cloud servers during the decline period.

[0023] During the capacity expansion phase, the purchase, assembly, and shelf-loading of hosts all take time. We call the time from placing an order to the shelf-loading of hosts the capacity expansion delivery time. In practice, the average delivery time is about one month, so capacity expansion decisions need to be made in advance. In addition, the capacity expansion process requires some procedures, transportation costs, etc. We model them as fixed costs for capacity planning. The level of fixed costs will affect the frequency of cloud server capacity adjustment. On the other hand, for each capacity expansion, if the capacity is expanded excessively, idle computing power will incur holding costs, and the existing capacity is insufficient to meet market demand. Unmet demand will be lost and result in penalty costs. During the recession period, capacity reduction is carried out to save holding costs. We assume that it will take several periods before the corresponding capacity is completely removed from the shelves. This is called the "capacity reduction preparation time." Similarly, capacity reduction plans should be decided in advance.

[0024] We construct a general dynamic model (multi-period newsboy model) to describe the three different stages of the life cycle, and the total length of the life cycle is recorded as This is the decision time domain, in which each period can make decisions on expanding or reducing the resource pool capacity. t represents the number of CPUs ordered by the cloud computing resource pool in period t, which is equivalent to the capacity change in period t. t >0 indicates expansion, while q t <0 means capacity reduction. t =(q t-L+1 ,…,q t-1 ) T is the in-transit order in period t. Assume that the delivery time for capacity expansion and the preparation time for capacity reduction are both L, and L is the number of waiting periods required for capacity change (actual delivery of capacity expansion or capacity reduction decision). If a capacity expansion or capacity reduction decision is made in period t, the CPU volume of the capacity expansion or capacity reduction decision will arrive in period t+L. Then in period t, t=1…T, the order of sequential decision events in each period is the following three processes.

[0025] 1. At the beginning of period t, orders placed before period L are received, and the current total resource pool NI is observedt-1 +q t -L=NI t Because the delivery time has L periods, use I t Represents the inventory level, which is the sum of the total resource pool on hand and orders in transit where 1 is a (L-1)-dimensional vector, and q i is the number of CPUs ordered by the cloud computing resource pool in the i-th period, which is equivalent to the capacity change in the i-th period; N is the net inventory; q t-L The number of CPUs ordered for the cloud computing resource pool in the tLth period is equivalent to the capacity change in the tLth period;

[0026] 2. According to the current total amount of resource pool, decide the capacity planning of resource pool and order q t The corresponding unit CPU purchase cost is c t , the order will be delivered in period t+L, and as long as the order quantity is not 0, the capacity planning fixed cost K will be incurred t On the other hand, in the capacity reduction phase, the unit CPU processing cost of capacity reduction is r t , capacity planning fixed cost K t .

[0027] 3. Observe the total demand of all users D t . According to the total NI of the resource pool t and total demand D t There are two costs for the matching degree: unsatisfied demand will incur a penalty cost b per CPU t , the redundant resource pool will generate a unit CPU holding cost h t The model assumes that all unsatisfied demands are lost. Without loss of generality, the penalty cost will be higher than the holding cost, i.e. b t >h t Always holds true. Also consider the cost discount factor α, 0≤α≤1.

[0028] The discounted total cost of a cloud computing server over its entire life cycle, i.e., the cost function of the multi-period newsboy model, has the following form: the cost consists of 1) capacity expansion cost, 2) capacity reduction cost, 3) fixed cost of capacity change, 4) holding cost of redundant resources, and 5) penalty cost of unmet demand.

[0029]

[0030] in, Indicates the size of the expansion capacity. Indicates the size of the reduced capacity, (x) +=max{x,0} means taking the maximum value between 0 and the value in the brackets, x represents the value of a random variable, (x) - =max{-x,0} means taking the maximum value of 0 and the opposite of the value in the brackets; Represents the indicator function, which takes the value 1 when the conditions in the brackets are met; α t represents the discount factor for period t, c t represents the unit purchase cost, r t represents the unit reduction cost in period t, K t represents the fixed cost of capacity change in period t, h t represents the unit holding cost of redundant resources in period t, NI t represents the total amount of cloud computing resources in the resource pool in period t, D t represents the total demand of all users in period t, b t represents the unit penalty cost in period t, and ∑ represents the summation symbol.

[0031] The uncertainty in the life cycle of cloud computing servers mainly lies in the total demand of users in each period. Next, we consider two situations where decision makers have different levels of knowledge about demand: a deterministic capacity planning scenario and an online random optimization scenario.

[0032] B2. Deterministic capacity planning scenario:

[0033] Obtain the total demand for each period of the life cycle of the cloud computing server, that is, the capacity in the deterministic scenario. Include the upper limit of the solution of the multi-period newsboy model in stochastic optimization. The optimization problem of the total demand for the life cycle of the cloud computing server in the deterministic scenario is expressed as a nonlinear mixed integer programming problem (objective function), and a solver such as Gurobi can be used to directly solve the capacity planning.

[0034] Deterministic capacity planning scenario: The company is a prophet who can perfectly predict the demand over the entire life cycle, that is, the company knows exactly the total demand in each period. The peak demand will occur at the tth m The optimization problem under certain circumstances is a nonlinear mixed integer programming problem, as follows:

[0035]

[0036] The objective function is to minimize the total cost of the entire decision time domain T, which is composed of the cost of expansion, the cost of reduction, the fixed cost of capacity change, the holding cost of redundant resources, and the penalty cost of unmet demand. Constraint 1 represents the change of the cloud computing capacity listed. Constraints 2 and 3 will continuously order the quantity q t and 0-1 variable Y tConstraints 4 and 5 define the decision variable, Y t It is a 0-1 variable, indicating whether to make a decision on capacity change. t It is the sum of listed cloud computing capabilities and is always positive.

[0037] in, Indicates the size of the expansion capacity. Indicates the size of the reduced capacity, (x) + =max{x,0} means taking the maximum value between 0 and the value in the brackets, x represents the value of a random variable, (x) - =max{-x,0} means taking the maximum value of 0 and the opposite of the value in the brackets; Represents the indicator function, which takes the value 1 when the conditions in the brackets are met; α t represents the discount factor for period t, c t represents the unit purchase cost, r t represents the unit reduction cost in period t, K t represents the fixed cost of capacity change in period t, h t represents the unit holding cost of redundant resources in period t, NI t represents the total amount of cloud computing resources in the resource pool in period t, D t represents the total demand of all users in period t, b t represents the unit penalty cost in period t, ∑ represents the summation symbol, and Y t is a 0-1 variable. If q t =0, then Y t =0, if q t ≠0, then Y t =1.

[0038] In actual application, the deterministic capacity planning scenario assumes that the point estimate obtained by demand forecasting is completely accurate and can be combined with demand forecasting for iterative capacity planning solution. In each decision cycle, the demand forecast value for the subsequent time is obtained through forecasting.

[0039] B3. Online random optimization scenario: At the beginning of each decision cycle, the decision maker can obtain the distribution of subsequent total demand.

[0040] The online stochastic optimization scenario is modeled as a Markov decision process and the Bellman equation is constructed.

[0041]

[0042] in, Indicates the size of the expansion capacity. Indicates the size of the reduced capacity, (x) +=max{x,0} means taking the maximum value of 0 and the value in the brackets, and x represents the value of a random variable; represents the indicator function, which takes the value 1 when the conditions in the brackets are met; α represents the discount factor, c t represents the unit purchase cost, r t represents the unit reduction cost in period t, K t represents the fixed cost of capacity change in period t, h t represents the unit holding cost of redundant resources in period t, NI t represents the total amount of cloud computing resources in the resource pool in period t, D t represents the total demand of all users in period t, b t represents the unit penalty cost in period t, f t represents the distribution of data at time t, Indicates the expected sign.

[0043] In theory, the optimal strategy in the online stochastic optimization scenario has a (s,S) structure, specifically:

[0044] 1. In the tth period of the expansion phase, if the unoccupied CPU inventory level (including in-transit orders) is lower than the threshold s t , that is, expansion is carried out, and after the expansion, the unoccupied CPU inventory level will reach the threshold S t Otherwise, no expansion will be performed.

[0045] 2. In the tth period of the capacity reduction phase, if the unoccupied CPU inventory level (including in-transit orders) is higher than the threshold S t , the capacity will be reduced, and after the capacity reduction, the unoccupied CPU inventory level will reach the threshold s t Otherwise, no capacity reduction will be performed.

[0046] In terms of algorithm, the present invention proposes an approximate algorithm AAAB (Adaptive-Approximation-Algorithm-Based) strategy. This strategy does not decompose the cost by decision period, but by decision behavior. That is, the cost of the tth period is divided by the capacity change q in the tth period. t Associated, once q t It is determined that these costs will not be affected by any future decisions, but only by future demand. We will use superscript A to refer to the AAAB algorithm in the future. Since the scope T, i.e., the life cycle of the cloud computing server, is also the total decision cycle, is finite, and the cost parameter is time-dependent, α is the time-dependent cost parameter, and we can assume that α is equal to 1 without loss of generality.

[0047] First, define the capacity change q in period t tDirectly associated effective holding costs and penalty costs It is expressed as:

[0048]

[0049] in, represents the effective holding cost in period t, q t represents the capacity change in period t, c t represents the unit purchase cost, h j represents the unit holding cost of redundant resources in period j, D j represents the total demand of all users in period j, I t represents the inventory level in period t. The superscript + indicates the maximum value between 0 and the value in brackets. t+L represents the unit holding cost of redundant resources in period t+L, D t+L represents the total demand of all users in period t+L, represents the penalty cost in period t, b t+L represents the unit penalty cost in period t+L, r t represents the unit reduction cost in period t, b j represents the unit penalty cost in the jth period, and ∑ represents the summation symbol.

[0050] The capacity planning decision of the AAAB strategy includes the expansion phase and the reduction phase, as follows:

[0051] 1. Calculate the balance That is, when the holding cost and the penalty cost are equal, the corresponding q t The amount is expressed as:

[0052]

[0053] Among them, f t represents the distribution of data at time t; θ t To balance the price;

[0054] in, represents the effective holding cost in period t, represents the penalty cost in period t, represents the capacity change in period t that makes the expected effective holding cost in period t equal to the expected penalty cost in period t, f t represents the distribution of data at time t, represents the expected sign, θ t To balance the price;

[0055] 2. In the expansion phase, calculate the holding cost K and other quantities ( That is, when the expected holding cost is equal to K, the corresponding q t Amount)

[0056]

[0057] And calculate and

[0058] in, represents the effective holding cost in period t, Indicates that the expected holding cost in period t is equal to K t Capacitance change at time, f t represents the distribution of data at time t, Indicates the expected sign.

[0059] 3. In the capacity reduction stage, calculate the capacity reduction cost K and other quantities (When the expected cost of capacity reduction is equal to K, the corresponding q t Amount)

[0060]

[0061] And calculate and

[0062] in, represents the penalty cost in period t, Indicates that the expected reduction cost in period t is equal to K t Capacitance change at time, f t represents the distribution of data at time t, Indicates the expected sign.

[0063] 4. Based on the calculation results, there are two specific capacity planning situations:

[0064] [Case 1.] If the equilibrium price (the expected value of holding cost or penalty cost when holding cost and penalty cost are equal) is fixed cost (K t ) is large, that is, θ t ≥K t , then the capacity change of AAAB strategy decision is the balance amount

[0065] [Case 2.] If the equilibrium price is less than the fixed cost, that is, θ t <K t , then the AAAB strategy has p t The probability of ordering K equal amounts is 1-p t The probability of not ordering, where the probability p t satisfy

[0066]

[0067] Among them, p t represents the probability of ordering K equal quantities, represents the penalty cost in period t, Indicates that the expected capacity reduction (or expansion) cost in period t is equal to K t Capacitance change at time, f t represents the distribution of data at time t, Indicates the expected sign.

[0068] The AAAB strategy always balances the expected effective holding cost and the expected effective penalty cost. It can be theoretically proven that the expected cost of the AAAB strategy is at most three times the expected cost of the optimal strategy.

[0069] C. Online allocation of cloud service computing resources through decision simulation; the implementation methods of online allocation are: single decision simulation analysis and end-to-end decision simulation analysis;

[0070] Single decision simulation analysis:

[0071] C1. Data reading and preparation: First, read and prepare the test data (including distribution rate data). The test data is the creation and deletion event records of the existing virtual machines on the cloud resource platform, including different distribution rate data; the distribution rate data is the actual allocation and usage of the current host on the cloud resource platform; the distribution rate data is the actual usage of the host on the cloud resource platform on the day, that is, the total number of cores of the distributed virtual machines divided by the total number of cores that can be placed on the host.

[0072] C2. Initialize host status: Initialize the NUMA (Non Uniform Memory Access) node load of each host according to the total load and the number of empty hosts. Each host has two NUMA nodes. Different NUMA nodes on the host are used for placement during actual allocation, and the needs of the virtual machine can only be completely placed on one NUMA of the host. The distribution rate data set is mainly used here; the initialization of the host status mainly depends on the distribution rate data set mentioned in C1, and the current usage and remaining amount of the host are initialized through the distribution rate information in the data set.

[0073] C3. Virtual machine event processing: For each virtual machine creation or deletion event, processing is performed based on the demand and the binary attributes of creation / deletion. The test dataset is mainly used here; the test dataset includes records of virtual machine demands that arrive and leave over a period of time. Through this dataset, we can simulate the arrival and departure process of each virtual machine.

[0074] C4. Output results: Finally, the information of allocated and unallocated virtual machines is recorded, and the total amount of unallocated vCPUs (virtual Central Processing Unit, specifying the CPU cores allocated to the virtual machine) (demand) is counted; after building the initial state of the host and preparing the arrival and departure data sets of the virtual machines, they are placed in chronological order to simulate the placement and departure process of these virtual machines. The most classic Firstfit algorithm in the packing problem is used in the specific implementation.

[0075] C5. Result analysis: Compare the impact of different allocation algorithms on capacity expansion decisions by the current allocation algorithm (Firstfit algorithm, 92% allocation rate) and some baseline methods of resource allocation (including Nextfit, Bestfit, etc., 82% allocation rate). The final results show (see Table 1) that the present invention can effectively improve the capacity expansion strategy.

[0076] End-to-end decision simulation analysis:

[0077] Starting from the initial time, the host profile at the initial time is used as the initial host distribution (the initial number of hosts and the placement of virtual machines on each host).

[0078] From the first day (initial time) to the last day (final time), the event record data of real virtual machine creation (when demand increases, virtual machines are placed according to the allocation algorithm) and deletion (when demand decreases, virtual machines on the host are deleted) in the system are used to perform placement on a long time scale (simulation on a long time scale in years).

[0079] According to the information provided by the expansion decision in part B (how many cores to expand each day and the number of hosts to expand), empty hosts are added into the system at the expansion time point, thereby innovatively achieving end-to-end simulation of the entire simulation process.

[0080] The simulated allocation results output the unallocated amount, allocation rate, demand, and demand satisfaction rate on a weekly basis. Whether the calculated allocation rate is close to the actual allocation rate can be used to determine whether the allocation algorithm and the macro expansion decision are reasonable. If the allocation rate is too low, it means that the expansion part has ordered more hosts, resulting in a certain waste of resources. The end-to-end decision simulation analysis is a simplified simulation of the actual allocation of the existing network over a long time scale (a long time scale in years).

[0081] In specific implementation, the present invention realizes a resource configuration system for modeling and simulating a public cloud resource management platform, including: a user demand prediction module, a capacity planning decision module and a cloud service computing resource online allocation module; wherein: the user demand prediction module is used to predict virtual machine demand data on a long-term scale (forecasting the next year every other week) to provide input for the capacity planning decision module; the capacity planning decision module is used to make long-term cloud resource expansion and reduction decisions under different scenarios (future demand is determined and only future demand distribution is known); the cloud service computing resource online allocation module is used to simulate and analyze the online allocation of cloud resources.

[0082] Compared with the prior art, the beneficial technical effects of the present invention are:

[0083] The present invention provides a resource configuration method based on a public cloud resource management platform modeling and simulation, which improves the allocation rate of the central processing unit CPU of the public cloud resource management system and reduces the number of virtual machine outages by adopting the cloud service resource management platform modeling and simulation and causal analysis methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0085] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0086] Figure 1 The flow of the method of the present invention is illustrated. First, the historical business data and public data are processed and transmitted as input to the demand side and supply side of the public cloud resource management system, and finally the function / regression relationship between "input data + key decision" and key indicators is obtained.

[0087] On the demand side, during the specific implementation, taking a certain cloud resource type c6_2 as an example, the data before 2020-07-01 is used as the training set, and the data from 2020-07-01 to 2024 is used as the test set (including the demand value for computing resources). The demand value and its distribution for the next 365 days are predicted every seven days (output as the mean and standard deviation of the normal distribution), and the actual value of the test set is subtracted from the predicted value to obtain the predicted residual. According to the law of large numbers, the initial outlier point limit is set to the absolute value of the minimum residual value, and the p-value (Probability Value) of the residual is solved. If the p-value is less than the set threshold (such as 0.05), it is considered that the residual follows the normal distribution, and the model cannot be adjusted from the perspective of the residual. Otherwise, the outlier point limit is continuously lowered, and the skewness coefficient of the residual of the prediction point stored after the limit is adjusted is calculated. and kurtosis coefficient The skewness Kurtosis δ γ1 represents the standard deviation of skewness, δ γ2 Represents the standard deviation of kurtosis, E represents the expected sign, μ represents the mean of the data, δ represents the standard deviation of the data, and X represents the value of the data. If the values ​​of the two coefficients are both between -1.96 and 1.96, it is considered that the residual follows a normal distribution. This boundary is the outlier demarcation boundary. Points greater than the current outlier boundary are outliers, and the outlier boundary is stopped from being lowered. An outlier analysis is performed on the outliers found, that is, the outlier types are divided into international, domestic, and regional political, economic, commercial, scientific and technological, and sports events. It is found that major events occurred near the date of the outlier. The minimum and maximum values ​​of different types are used as the value range of the type. If the time is 3 days before and after the outlier, a date is randomly selected within the time range, and a value is randomly selected within the value range of the corresponding type and added to the original predicted value of that day. It is found that the prediction accuracy is increased by 0.5%. The calculation formula for the prediction accuracy is 1-SMAPE (Symmetric Mean Absolute Percentage Error, symmetric mean absolute percentage error). y i Represents the true value of the data, Represents the data prediction value, and n represents the number of data. Output the prediction sequence and pass it to the supply side and the supply and demand matching side.

[0088] On the decision-making side: The specific implementation plan for capacity planning decisions is:

[0089] First, build the model:

[0090] The order of events in each decision cycle t is defined as follows:

[0091] a. At the beginning of period t, the company receives an order placed L periods before the delivery time and observes the current total resource pool NI t .

[0092] b. According to the current total amount of resource pool, decide the capacity planning of resource pool and order q t The amount of CPU, q t Indicates the in-transit orders in period t The corresponding unit CPU purchase cost is c t , the order will be delivered in period t+L, and as long as the order quantity is not 0, the capacity planning fixed cost K will be incurred t On the other hand, during the capacity reduction phase, the unit CPU processing cost of the company's capacity reduction is r t , capacity planning fixed cost K t .

[0093] c. The company observes the total demand of all users D tDepending on the matching degree between the total resource pool and the total demand, there are two kinds of costs: Unmet demand will incur a penalty cost per CPU b t , the redundant resource pool will generate a unit CPU holding cost h t The model assumes that all unmet demand is lost. Consider the cost discount factor α, which is greater than 0 and less than or equal to 1.

[0094] The discounted total cost of a cloud computing server over its entire life cycle has the following form: the cost is composed of 1) capacity expansion cost, 2) capacity reduction cost, 3) fixed cost of capacity change, 4) holding cost of redundant resources, and 5) penalty cost of unmet demand.

[0095]

[0096] The state transition is defined as:

[0097]

[0098] Secondly, the scenarios are divided into deterministic capacity planning scenarios and uncertain capacity planning scenarios.

[0099] Deterministic capacity planning scenario: The company is a prophet who can perfectly predict the demand over the entire life cycle, that is, the company knows exactly the total demand in each period. The peak demand will occur at the tth m The optimization problem under certain circumstances is a nonlinear mixed integer programming problem, as follows:

[0100]

[0101] The objective function is to minimize the total cost of the entire decision time domain T, which is composed of the cost of expansion, the cost of reduction, the fixed cost of capacity change, the holding cost of redundant resources, and the penalty cost of unmet demand. Constraint 1 represents the change of the cloud computing capacity listed. Constraints 2 and 3 will continuously order the quantity q t and 0-1 variable Y t Constraints 4 and 5 define the decision variable, Y t It is a 0-1 variable, indicating whether to make a decision on capacity change. t It is the sum of listed cloud computing capabilities and is always positive.

[0102] in, Indicates the size of the expansion capacity. Indicates the size of the reduced capacity, (x) + =max{x,0} means taking the maximum value between 0 and the value in the brackets, x represents the value of a random variable, (x) -=max{-x,0} means taking the maximum value of 0 and the opposite of the value in the brackets; Represents the indicator function, which takes the value 1 when the conditions in the brackets are met; α t represents the discount factor for period t, c t represents the unit purchase cost, r t represents the unit reduction cost in period t, K t represents the fixed cost of capacity change in period t, h t represents the unit holding cost of redundant resources in period t, NI t represents the total amount of cloud computing resources in the resource pool in period t, D t represents the total demand of all users in period t, b t represents the unit penalty cost in period t, ∑ represents the summation symbol, and Y t is a 0-1 variable. If q t =0, then Y t =0, if q t ≠0, then Y t =1.

[0103] The above problem can be solved directly using a solver such as Gurobi.

[0104] In actual application, the deterministic capacity planning scenario assumes that the point estimate obtained by demand forecasting is completely accurate and can be combined with demand forecasting for iterative capacity planning solution. In the tth period, the demand forecast value of the subsequent time is obtained through forecasting, and the discounted total cost after the tth period is taken as the optimization target, that is,

[0105]

[0106] Among them, t′ refers to the tth to Tth period of the life cycle; represents the size of the expansion capacity in period t′, represents the size of the capacity reduction in period t′, (x) + =max{x,0} means taking the maximum value between 0 and the value in the brackets, x represents the value of a random variable, (x) - =max{-x,0} means taking the maximum value of 0 and the opposite of the value in the brackets; represents the indicator function, which takes the value 1 when the conditions in the brackets are met and takes the value 0 when they are not met; αt′-t represents the discount factor for the t′-t period, c t′ represents the unit purchase cost, r t′ represents the unit reduction cost in period t′, K t′ represents the fixed cost of capacity change in period t′, h t′ represents the unit holding cost of redundant resources in period t′, NI t′represents the total amount of cloud computing resources in the resource pool in period t′, D t′ represents the total demand of all users in period t′, b t′ represents the unit penalty cost in period t′, and ∑ represents the summation symbol.

[0107] By optimizing the above formula, we can get the capacity planning strategy for the tth period and beyond. In the t+1th period, based on the total demand actually observed in the tth period, we update the forecast of the subsequent total demand, and then combine it with the current inventory of CPUs obtained by the previous capacity planning strategy, and take the discounted total cost after the t+1th period as the optimization target to solve it, and update it in this reciprocating order.

[0108] Take a cloud resource type c6_2 as an example, and make capacity planning decisions for the subsequent entire domain at t=0. In specific implementation, make decisions on a weekly basis, take every Friday as the decision time point, and do not make capacity planning decisions on the six days in between. Set the discount factor α to 0.9, c to 1, h to 1, r to 1, b to 500, K to 100, L to 9, and the initial inventory value to the actual supply (host) CPU total amount in the previous L periods. In this case, the ideal average allocation rate of CPU can reach 89.81%, and there will be no supply interruption, which is 8.27% higher than the current actual situation.

[0109] Online random optimization scenario: At the beginning of each decision cycle, the decision maker can obtain the distribution of subsequent total demand.

[0110] The online stochastic optimization scenario is modeled as a Markov decision process and the Bellman equation is constructed.

[0111]

[0112] in, Indicates the size of the expansion capacity. Indicates the size of the reduced capacity, (x) + =max{x,0} means taking the maximum value of 0 and the value in the brackets, and x represents the value of a random variable; represents the indicator function, which takes the value 1 when the conditions in the brackets are met; α represents the discount factor, c t represents the unit purchase cost, r t represents the unit reduction cost in period t, K t represents the fixed cost of capacity change in period t, h t represents the unit holding cost of redundant resources in period t, NI t represents the total amount of cloud computing resources in the resource pool in period t, D t represents the total demand of all users in period t, b t represents the unit penalty cost in period t, f tRepresents the distribution of data at time t.

[0113] In theory, the optimal strategy in the online stochastic optimization scenario has a (s,S) structure, specifically:

[0114] 3. In the tth period of the expansion phase, if the unoccupied CPU inventory level (including in-transit orders) is lower than the threshold s t , that is, the company will expand its capacity, and after the expansion, the unoccupied CPU inventory level will reach the threshold S t Otherwise, the company will not expand.

[0115] 4. In the tth period of the capacity reduction phase, if the unoccupied CPU inventory level (including in-transit orders) is higher than the threshold S t , the company will reduce the capacity, and after the reduction, the unoccupied CPU inventory level will reach the threshold s t Otherwise, the company will not reduce capacity.

[0116] Due to the curse of dimensionality (the dimensions of the parameters and variables involved are too large), the two thresholds s that constitute the optimal strategy t and S t It is difficult to calculate directly.

[0117] In terms of algorithm, we propose an approximate algorithm AAAB strategy, which follows the classic EOQ idea to balance nonlinear fixed costs and holding costs with penalty costs. The AAAB strategy does not decompose costs by decision period, but by decision behavior, that is, the cost is divided into the capacity change q in the tth period. t Associated with, once q t It is determined that these costs will not be affected by any future decisions, but only by future demand, and the superscript A is used to refer to the AAAB algorithm. Since the horizon T is finite and the cost parameters are time-dependent, we can assume α equals 1 without loss of generality.

[0118] First, define the capacity change q in period t t Directly associated effective holding costs and penalty costs

[0119]

[0120] in, represents the effective holding cost in period t, q t represents the capacity change in period t, c t represents the unit purchase cost, h j represents the unit holding cost of redundant resources in period j, D j represents the total demand of all users in period j, I trepresents the inventory level in period t. The superscript + indicates the maximum value between 0 and the value in brackets. t+L represents the unit holding cost of redundant resources in period t+L, D t+L represents the total demand of all users in period t+L, represents the penalty cost in period t, b t+L represents the unit penalty cost in period t+L, r t represents the unit reduction cost in period t, b j represents the unit penalty cost in period j, ∑ is the summation symbol;

[0121] The capacity planning decisions for the specific AAAB strategy are as follows:

[0122] 5. Calculate the balance Right now

[0123]

[0124] in, represents the effective holding cost in period t, represents the penalty cost in period t, represents the capacity change in period t that makes the expected effective holding cost in period t equal to the expected penalty cost in period t, f t represents the distribution of data at time t, represents the expected sign, θ t To balance the price;

[0125] 6. During the expansion phase, calculate (holding cost) K and other quantities

[0126]

[0127] And calculate and

[0128] in, represents the effective holding cost in period t, Indicates that the expected holding cost in period t is equal to K t Capacitance change at time, f t represents the distribution of data at time t, Indicates the expected sign.

[0129] In the capacity reduction stage, calculate (capacity reduction cost) K and other quantities

[0130]

[0131] And calculate and

[0132] in, represents the penalty cost in period t, Indicates that the expected reduction cost in period t is equal to K t Capacitance change at time, f t represents the distribution of data at time t, Indicates the expected symbol

[0133] 7. Based on the calculation results, there are two specific capacity planning situations:

[0134] [Case 1.] If the equilibrium price is greater than the fixed cost, that is, θ t ≥K t , then the capacity change of AAAB strategy decision is the balance amount

[0135] [Case 2.] If the equilibrium price is less than the fixed cost, that is, θ t <K t , then the AAAB strategy has p t The probability of ordering the expected holding cost K is equal to There is 1-p t The probability of not ordering, where the probability p t satisfy

[0136]

[0137] Among them, p t represents the probability of ordering K equal quantities, represents the penalty cost in period t, Indicates that the expected capacity reduction (or expansion) cost in period t is equal to K t Capacitance change at time, f t represents the distribution of data at time t, Indicates the expected sign.

[0138] The AAAB strategy always balances the expected effective holding cost and the expected effective penalty cost. It can be theoretically proven that the expected cost of the AAAB strategy is at most three times the expected cost of the optimal strategy.

[0139] C. Online allocation of cloud service computing resources through decision simulation; the implementation methods of online allocation are: single decision simulation analysis and end-to-end decision simulation analysis.

[0140] Take a cloud resource type c6_2 as an example, and make capacity planning decisions for the subsequent entire domain at t=0. The specific implementation is to make decisions in weekly units, take every Friday as the decision time point, and do not make capacity planning decisions on the 6 days in between. Set the discount factor α to 0.9, c to 1, h to 1, r to 1, b to 500, K to 100, L to 9, and the initial inventory value to the actual supply (host) CPU total amount in the previous L periods. In this case, the ideal average allocation rate of CPU can reach 86.08%, which is 4.54% higher than the current actual situation. There will be 6 supply outages, with a total supply of 46915 CPUs.

[0141] Table 1 Comparison of the results of online allocation using the method of the present invention on the data of 2021-2023

[0142]

[0143] On the online allocation side, first, read and prepare test data and distribution rate data. Test data is the record of creation and deletion events of existing virtual machines on the cloud resource platform. Distribution rate data is the actual allocation and usage of the current numerical experiment host on the day. Then, the NUMA node load of each host is initialized according to the total load and the number of empty hosts. Each host has two NUMA nodes. When actually allocating, different NUMAs on the host are used for placement, and the demand of the virtual machine can only be completely placed on a certain NUMA of the host. The distribution rate data set is mainly used here. For each virtual machine creation or deletion event, it is processed according to the demand and the amount of creation / deletion (binary attributes). The test data set is mainly used here. Finally, the information of allocated and unallocated virtual machines is recorded, and the total amount of unallocated vCPUs (demand) is counted. Compare the impact of different allocation algorithms on expansion decisions: According to Table 1, we can see that we conducted three experimental comparisons, and the values ​​in the table are the final allocation rates (percentages). The base column refers to the allocation rate without adjustment based on online allocation. 92 refers to the result of adjustment after we use the allocation algorithm with an allocation rate of 92% for a single simulation decision (that is, the allocation algorithm of double numa is not considered), and 82 refers to the result of adjustment after we use the allocation algorithm with an allocation rate of 92% for a single simulation decision (that is, the allocation algorithm of double numa is considered, which is also a more realistic algorithm). From the results, we can conclude that:

[0144] 1. Algorithms with different allocation rates have a great impact on macro expansion decisions. The results of the allocation algorithm with an allocation rate of 92% and the allocation algorithm with an allocation rate of 82% are significantly different;

[0145] 2. The online allocation part has a very significant impact on the macro expansion part, which explains the significance of the existence of this part, that is, to make appropriate adjustments to the macro expansion decision part.

[0146] It should be noted that the purpose of publishing the embodiments is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A cloud service management platform resource configuration method, It is characterized in that By predicting the user demand of the system, capacity planning decisions are made, and then online allocation of cloud service computing resources is performed through decision simulation; the steps include: A. Perform user demand forecasting; obtain the outlier boundary through analysis, identify the outliers and process them, and obtain the forecast sequence of cloud server virtual machine demand; B. Make dynamic capacity planning decisions for cloud server resources; including: B1. Construct a multi-period newsboy model, define the cost structure and state transition parameters, and the order of occurrence of decision events in each period t in the life cycle of the cloud server; including: Expand cloud server capacity during the growth and maturity stages of the cloud server life cycle, and reduce cloud server capacity during the decline stage; Decisions on expanding or reducing resource pool capacity at each stage of the life cycle; In the tth period, t = 1…T, the total length of the life cycle is That is, the decision time domain; the order in which sequential decision events occur includes the following three processes: First, at the beginning of period t, we receive orders placed before period L, and observe that the current total resource pool is NI t-1 +q t-L =NI t ; Among them, I t The total amount of the resource pool and the sum of orders in transit in period t is the inventory level where 1 is an L-1 dimensional vector, and N is the net inventory; the delivery time for capacity expansion and the preparation time for capacity reduction are both L; Q t =(q t-L+1 , ..., q t-1 ) T For orders in transit; i is the number of CPUs ordered by the cloud computing resource pool in the i-th period; q t-L is the number of CPUs ordered by the cloud computing resource pool in the tLth period, that is, the capacity change in the tLth period; Second, q t represents the capacity change in period t; according to the current total amount of resource pool, decide the capacity planning of resource pool and order q t The number of CPUs; where q t >0 means expansion, q t <0 means capacity reduction; the purchase cost of a unit CPU in period t is c t , the order is delivered in period t+L, and the capacity planning fixed cost K is generated t ; In the capacity reduction phase, the unit CPU processing cost of the capacity reduction in period t is r t , that is, the unit reduction cost is r t ; The fixed cost of capacity planning is K t ; Third, observe the total demand D of all users in period t t ; According to the total amount of resource pool NI in period t t and total demand D t There are two kinds of costs: unmet demand will incur a penalty cost per CPU b t , the redundant resource pool will generate a unit CPU holding cost h t ; The penalty cost is higher than the holding cost, that is, b t >h t ; The discounted total cost of a cloud computing server over its entire life cycle includes: capacity expansion cost, capacity reduction cost, capacity planning fixed cost, holding cost, and penalty cost; The discounted total cost of a cloud computing server over its entire life cycle is the cost function of the multi-period newsboy model, expressed as: in, is the discounted total cost function of the cloud computing server over its entire life cycle; Indicates the size of the expansion capacity. Indicates the size of the reduced capacity; () + The superscript + in the brackets indicates the maximum value between 0 and the value in brackets; () - The superscript - in the brackets indicates the maximum value between 0 and the opposite of the value in the brackets. Represents the indicator function, which takes the value 1 when the conditions in the brackets are met; α t represents the discount factor for period t; B2. The optimization of the total demand for the cloud computing server from the tth period to the Tth period in the life cycle is expressed as a nonlinear mixed integer programming objective function, and the capacity planning is directly solved by the solver; the optimization objective function of the total demand for the cloud computing server from the tth period to the Tth period in the life cycle is expressed as: Where t′ refers to any period from the tth period to the Tth period of the life cycle; α t′-t represents the discount factor for period t′-t; represents the size of the expansion capacity in period t′, represents the capacity reduction in period t′; c t′ represents the unit purchase cost in period t′, r t′ represents the unit reduction cost in period t′, K t′ represents the fixed cost of capacity planning in period t′, h t′ represents the unit holding cost of redundant resources in period t′, NI t′ represents the total amount of cloud computing resources in the resource pool in period t′, D t′ represents the total demand of all users in period t′, b t′ represents the unit penalty cost in period t′; B3. Establish an adaptive approximation algorithm AAAB; including: a. Decompose the cost according to the decision behavior and compare the cost with the capacity change q in the tth period of the life cycle t When the capacity change q in period t is t It is certain that the cost in period t is only related to future demand; Definition and capacity change q in period t t Directly associated effective holding costs and penalty costs b. Calculate the balance That is, when the holding cost and the penalty cost are equal, the corresponding q t The amount of c. Calculate the holding cost in the expansion phase and the reduction cost in the reduction phase; In the expansion phase, calculate the holding cost K and other quantities The holding cost K is equivalent to That is, when the expected holding cost is equal to K, the corresponding q t The amount of In the capacity reduction stage, calculate the capacity reduction cost K and other quantities The capacity reduction cost K is equivalent to That is, when the expected cost of capacity reduction is equal to K, the corresponding q t The amount of d. Perform capacity planning to balance the expected effective holding cost and the expected effective penalty cost so that the expected cost is at most three times the expected cost of the optimal strategy; C. Online allocation of cloud service computing resources through single decision simulation analysis and end-to-end decision simulation analysis; Through the above steps, resource configuration based on modeling and simulation of public cloud resource management platform is realized.

2. The cloud service management platform resource configuration method according to claim 1, It is characterized in that Step A specifically uses the time series prediction model PROPHET model to predict user demand, including: A1. Data processing: The data involved is cloud server demand data, including customer demand data for cloud server virtual machines; The data structure includes: the number of virtual machines required for each date and the corresponding date; First, perform data cleaning to convert the required data into an input format acceptable to the model, that is, set it to the column name corresponding to the PROPHET prediction model, and divide it into training set and test set data; A2. Perform PROPHET prediction: Substitute the training set data into the PROPHET forecasting model to perform time series forecasting on a long time scale; the long time scale is to forecast backward for one year every seven days, that is, to forecast the demand value and its distribution of virtual machines for the next 365 days every seven days from the starting point of the test set data, and output the forecast value of the demand with the mean and standard deviation of the normal distribution; A3. Through residual normal analysis, determine whether to improve the model from the perspective of residuals: The residual is obtained by subtracting the predicted value of the virtual machine demand from the actual value of the virtual machine demand in the test set. A normal analysis is performed on the residual. If the residual obeys the normal distribution, the process jumps to A5. If the residual does not obey the normal distribution, the process is performed as follows: Set the initial outlier limit to the absolute value of the minimum residual, and continuously lower the outlier limit; For the residual data under different outlier point limits, calculate the kurtosis coefficient and skewness coefficient of the residual data; If the kurtosis coefficient and the skewness coefficient are both less than the set threshold, the result of normal analysis of the current residual follows a normal distribution; If the demand for virtual machines on a certain date is greater than the current abnormal point limit, it is an abnormal point, and the abnormal point limit should stop being lowered; A4. Perform outlier processing on the outliers found in A3; Use the virtual machine demand corresponding to the previous moment as the virtual machine demand corresponding to the current abnormal point moment to analyze the prediction accuracy; The prediction accuracy is 1-SMAPE, where y i Represents the true value of the data, Represents the data prediction value, n represents the number of data; outputs the prediction sequence, which is passed to the supply side and the supply and demand matching side; i is the sequence number; A5. Output the forecast sequence obtained through the PROPHET forecast model and outlier processing; the forecast sequence is the forecast data of virtual machine demand forecasted for one year every other week.

3. The cloud service management platform resource configuration method according to claim 1, It is characterized in that In step B2, Gurobi is used as the solver.

4. The cloud service management platform resource configuration method according to claim 1, It is characterized in that In step B3, define the capacity change q in period t t Directly associated effective holding costs and penalty costs It is expressed as: Among them, A represents the adaptive approximation algorithm AAAB; T is the life cycle of the cloud computing server; represents the effective holding cost in period t, c t represents the unit purchase cost, h j represents the unit holding cost of redundant resources in period j, D j represents the total demand of all users in period j. The superscript + represents the maximum value between 0 and the value in brackets. t+L represents the unit holding cost of redundant resources in period t+L, D t+L represents the total demand of all users in period t+L, represents the penalty cost in period t, b t+L represents the unit penalty cost in period t+L, r t represents the unit reduction cost in period t, b j represents the unit penalty cost in period j, ∑ represents the summation symbol; Calculate the balance It is expressed as: Where E[ ] represents the expected function; f t represents the distribution of data at time t; θ t To balance the price.

5. The cloud service management platform resource configuration method according to claim 1, It is characterized in that In step C, the single decision simulation analysis includes: C1. Read and prepare test data and its distribution rate data; the test data is the creation and deletion event records of existing virtual machines on the cloud resource platform; C2. Initialize host status: Initialize the non-uniform memory access NUMA node load of each host according to the total load and the number of empty hosts. Each host has two NUMA nodes. Different NUMA nodes on the host are used for placement during actual allocation, and the needs of the virtual machine can only be completely placed on one NUMA of the host; C3. Virtual machine event processing: For each virtual machine creation or deletion event in the test data set, it is processed according to the demand and the amount of creation or deletion; C4. Output result: record the information of allocated and unallocated virtual machines, and obtain the total required amount of CPU cores of unallocated virtual machines.

6. The cloud service management platform resource configuration method according to claim 5, It is characterized in that Step C4 specifically adopts the Firstfit algorithm.

7. The cloud service management platform resource configuration method according to claim 1, It is characterized in that In step C, the end-to-end decision simulation analysis includes: Starting from the initial time, the host profile at the initial time is used as the initial host distribution, that is, the initial number of hosts and the virtual machines placed on each host; From the initial time to the final time, the event record data of the creation and deletion of real virtual machines in the system is used for long-term placement; the long-term scale is to predict the next year every week; when real virtual machines are created, that is, when demand increases, virtual machines are placed according to the allocation algorithm; when virtual machines are deleted, that is, when demand decreases, virtual machines on the host are deleted; Based on the expansion decision, the number of cores and hosts to be expanded each day is determined, and empty hosts are added to the system at the expansion time point, thereby realizing an end-to-end simulation process.

8. The cloud service management platform resource configuration method according to claim 5 or 7, It is characterized in that The allocation results of the simulation are based on weekly output of unallocated quantity, allocation rate, demand quantity and demand satisfaction rate.

9. A system using the cloud service management platform resource configuration method according to claim 1, It is characterized in that The system is used to configure the resources of the cloud service management platform, including: a user demand prediction module, a capacity planning decision module and a cloud service computing resource online allocation module; wherein: The user demand prediction module is used to predict the long-term virtual machine demand data and provide input for the capacity planning decision module; the capacity planning decision module is used to make long-term cloud resource expansion and reduction decisions in different scenarios; the cloud service computing resource online allocation module is used to simulate and analyze the online allocation of cloud resources.

Images