Cloud service provider dynamic pricing method, system and device and storage medium

Abstract

The invention provides a dynamic pricing method, system and device for a cloud service provider, and a storage medium, and relates to the field of network resource evaluation. According to the characteristics of different cloud services in a cloud federation computing platform, a meta-heuristic algorithm and a pricing strategy of deep reinforcement learning are provided for universal cloud services and special cloud services respectively. According to the method, multiple factors influencing cloud service pricing are fully considered, time sequence features, continuous features and classification features are involved, and the optimal pricing strategy is dynamically adjusted according to historical data and real-time environment states of services. According to the method, a proper pricing strategy is adopted for different types of cloud services, a distribution and profit balance point is sought, the service quality of the cloud federation is improved, resource use is optimized, and the profit of a cloud service provider in the federation is improved to the maximum extent.

Description

Technical Field

[0001] The present invention relates to the field of network resource evaluation, and particularly to a dynamic pricing method, system, device and storage medium for cloud service providers. Background Art

[0002] In recent years, cloud computing has provided various services through the Internet, changed the paradigm of computing, and become a new model for providing flexible and comprehensive services to users and various companies. Cloud service providers provide infrastructure to customers in the form of virtual machines, and connect the infrastructures of different providers by forming a cloud federation to achieve resource sharing and provide QoS that meets user satisfaction. This cloud federation can make more effective use of idle server resources and improve the fault tolerance of the system.

[0003] However, there are several providers in the cloud federation, and the service prices and qualities provided are different. Therefore, how to manage resources more effectively is an important task for the cloud federation. Resource management in the cloud federation includes four parts: discovery, selection, pricing, and resource allocation. Pricing is one of the processes of resource management, which can promote the maximization of the profits of cloud federation members and achieve the optimization of resource utilization, so that the federation members can provide shared resources according to requests and avoid over - supply or scarcity of resources.

[0004] The difference between the cloud federation platform and a general cloud service platform is that the services provided by multiple different cloud service providers are also diverse. A single pricing strategy cannot be applied to different types of cloud services, and the resources required by cloud services themselves fluctuate due to policies and upstream supply chains. Therefore, the pricing strategy for the cloud federation platform needs to dynamically adapt to the market environment, effectively price various different cloud services reasonably, and maximize the profits of cloud federation members. Summary of the Invention

[0005] Object of the Invention: In order to formulate a reasonable pricing strategy for various different cloud services in the cloud federation cloud computing platform, the present invention proposes a dynamic pricing method for cloud service providers, and further proposes a system, device and storage medium for implementing this method, which divides the cloud services of the cloud federation computing platform into general services and proprietary services. For general services with less resource requirements and relatively stable prices, a method based on linear programming is proposed to allocate requests among federation members and use a meta - heuristic algorithm for service pricing. For proprietary services with large resource requirements and relatively large impacts from the supply chain market, a pricing strategy based on deep reinforcement learning is proposed, fully considering various factors such as user needs, market environment, and historical pricing. By adopting appropriate pricing strategies for different types of cloud services, seeking the equilibrium point between allocation and profit, improving the service quality of the cloud federation, optimizing resource use, and maximizing the profits of cloud service providers in the federation.

[0006] In the first aspect of the present invention, a dynamic pricing method for cloud service providers is proposed, and the steps are as follows:

[0007] Obtain the inventory resource information of virtual machines from cloud service providers, and send resource requests according to costs and prices;

[0008] For each resource request, calculate the resource allocation situation of the required resources in the cloud service provider, and give an allocation array;

[0009] For each resource allocation situation in the allocation array, calculate the balance point of the allocation and revenue coefficient of the cloud service provider by using linear regression;

[0010] Divide each resource request into a general service request or a dedicated service request according to the resource consumption;

[0011] For each general service request within a period, based on the balance point of the allocation and revenue coefficient, use a meta-heuristic algorithm for pricing;

[0012] For each dedicated service request within a period, on the basis of the balance point of the allocation and revenue coefficient, comprehensively consider the historical sales time series characteristics, resource inventory continuous characteristics, and promotion method classification characteristics, and use reinforcement learning for dynamic pricing adjustment;

[0013] Output the pricing result.

[0014] In a further embodiment of the first aspect, obtaining the resource information and inventory resources of virtual machines from cloud service providers specifically includes:

[0015] Request the resources and costs of all cloud service providers on the cloud federation platform, and construct a shared resource pool G:

[0016] G = CSP(1) ∪ CSP(2) ∪ … ∪ CSP(n)

[0017] In the formula, CSP(n) represents the amount of resources that the nth cloud service provider can provide; each cloud service provider CSP includes the total CPU resource amount, the total GPU resource amount, the memory resource amount, the network resource amount, and the number of online servers.

[0018] In a further embodiment of the first aspect, for each resource request L, calculate the resource allocation situation of the required resources in the cloud service provider, and the remaining resource data array G of the shared resource pool after each allocation ′ , and calculate the allocation array after allocating the resource request L to all cloud service providers CSP that meet the requirements:

[0019] G ′ = G - L(x1,x2…xn)

[0020] CSP(i) ′ = CSP(i) - L

[0021]

[0022] Set(G) = {G1, G2…Gn}

[0023] Wherein, G1 represents the situation where the resource request L is allocated to CSP(1); G2 represents the situation where the resource request L is allocated to CSP(2); Gn represents the situation where the resource request L is allocated to CSP(n); CSP(i) represents the amount of resources that the i-th cloud service provider can provide; CSP(i) ′ represents the remaining amount of resources after the i-th cloud service provider allocates the required resources to the resource request L; Set(G) represents the allocation array formed by all possible allocation sets.

[0024] In a further embodiment of the first aspect, the allocation and revenue coefficient equilibrium point p of the cloud service provider is calculated by means of linear regression:

[0025]

[0026] Wherein, Provision represents the provision amount of the CSP; ALCR represents the supply balance; MN represents the allocation and revenue coefficient; β is the proportionality coefficient.

[0027] In a further embodiment of the first aspect, each resource request is divided into a general service request or a dedicated service request according to the resource consumption, specifically including:

[0028] If the situation Gi where the resource request L is allocated to the i-th cloud service provider is less than the preset value, it is defined as the general service request G geneal ; if the situation Gi where the resource request L is allocated to the i-th cloud service provider is greater than or equal to the preset value, it is defined as the dedicated service request G spec ; G geneal ∪G spec = G.

[0029] In a further embodiment of the first aspect, for each general service request G during the period geneal , based on the allocation and revenue coefficient equilibrium point p, when the allocation and revenue coefficient MN is less than 0.1, the service provider is re-evaluated:

[0030] if(MN < 0.1) reAllocate(L)

[0031] Otherwise, directly complete the allocation with the current allocation and revenue coefficient equilibrium point p and return the pricing:

[0032]

[0033] Wherein, p(L) represents that the resource request L is allocated according to the current allocation and the equilibrium point p of the income coefficient; reAllocate(L) represents that the resource request L re-evaluates the service provider.

[0034] In a further embodiment of the first aspect, for each dedicated service request within a period, based on the equilibrium point of the allocation and income coefficient, comprehensively considering the historical sales time series characteristics, the continuous characteristics of resource inventory, and the classification characteristics of promotion methods, reinforcement learning is used to perform dynamic pricing adjustment;

[0035] The goal of reinforcement learning is to learn the policy π, and the policy π dynamically adjusts the service price according to the current state component characteristics to maximize the expected cumulative revenue:

[0036]

[0037] Wherein, τ represents the complete pricing sequence data; r(τ) represents the total revenue of the pricing sequence data;

[0038] Based on the request pricing of the current round of cycle, a feature vector is constructed to train the pricing model of the next round of cycle; the pricing of the current cycle is then predicted by the model after constructing the feature vector.

[0039] In the second aspect of the present invention, a dynamic pricing optimization system for a cloud federation platform is proposed, and the system includes:

[0040] A resource request module, configured to obtain the inventory resource information of virtual machines from a cloud service provider, and send a resource request according to the cost and price;

[0041] A service provider evaluation module, configured to calculate the resource allocation situation of each resource request required in the cloud service provider, and give an allocation array; for each resource allocation situation in the allocation array, calculate the equilibrium point of the allocation and income coefficient of the cloud service provider by using linear regression;

[0042] A service provider scheduling module, configured to divide each resource request into a general service request or a dedicated service request according to the resource consumption;

[0043] A general service pricing module, configured to perform pricing on each general service request within a period based on the equilibrium point of the allocation and income coefficient, and output a pricing result by using a meta-heuristic algorithm;

[0044] A dedicated service pricing module, configured to perform dynamic pricing adjustment on each dedicated service request within a period based on the equilibrium point of the allocation and income coefficient, comprehensively considering the historical sales time series characteristics, the continuous characteristics of resource inventory, and the classification characteristics of promotion methods, and output a pricing result.

[0045] In a third aspect of the present invention, an electronic device is proposed. The electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the dynamic pricing method of the cloud service provider described in the first aspect is implemented.

[0046] In a fourth aspect of the present invention, a computer-readable storage medium is proposed. At least one executable instruction is stored in the storage medium. When the executable instruction runs on an electronic device, the electronic device is caused to execute the dynamic pricing method of the cloud service provider described in the first aspect.

[0047] Beneficial effects: The present invention proposes a dynamic pricing method for cloud service providers. According to the characteristics of different cloud services in the cloud federation computing platform, meta-heuristic algorithms and deep reinforcement learning pricing strategies are respectively proposed for general-purpose cloud services and dedicated cloud services. It fully considers various factors affecting cloud service pricing, including temporal features, continuous features, and categorical features, and dynamically adjusts the optimal pricing strategy according to the historical data of the service and the real-time environmental state. By adopting appropriate pricing strategies for different types of cloud services, the present invention seeks the equilibrium point between allocation and profit, improves the service quality of the cloud federation, optimizes resource utilization, and maximizes the profit of cloud service providers in the federation. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic flow chart of the dynamic pricing method of the cloud service provider in the embodiment.

[0049] Figure 2 It is a schematic diagram of the allocation and coefficient income balance in the embodiment.

[0050] Figure 3 It is a schematic diagram of the dynamic pricing rebalancing in the embodiment.

[0051] Figure 4 It is an architecture diagram of the cloud service provider dynamic pricing optimization system in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] In the following description, a large number of specific details are given to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, some well-known technical features are not described to avoid confusion with the present invention.

[0053] Before elaborating on the embodiments in detail, some proprietary terms that may appear later are first explained.

[0054] QoS (Quality of Service) refers to the ability of a network to use various underlying technologies to provide better service capabilities for specified network communications.

[0055] IaaS (Infrastructure as a Service) refers to a service model in which IT infrastructure is provided as a service over a network and billed according to the actual usage or occupancy of resources by users.

[0056] SaaS (Software-as-a-service) is a software application model that provides software services based on the Internet.

[0057] CSP (Cloud Service Provider) refers to a manufacturer that provides cloud services in a cloud federation platform.

[0058] A3C (Asynchronous Advantage Actor-Critic) is a deep reinforcement learning algorithm based on Actor-Critic that provides a general asynchronous and concurrent reinforcement learning framework and can be trained in parallel using multi-core CPUs or distributed computing resources.

[0059] LSTM (Long Short-Term Memory) is a variant of the Recurrent Neural Network (RNN) used for modeling and predicting sequence data and time series data.

[0060] This embodiment proposes a dynamic pricing method for cloud service providers, and its framework is as Figure 1As shown. This method is a three-stage algorithm that uses linear programming, meta-heuristic algorithms, and deep reinforcement learning to optimize resource usage and maximize the profit of cloud service providers in the federation. The first stage is the resource request stage, in which resource information (CPU Core, Memory, GPU) of virtual machines and the inventory resources of service providers are obtained from service providers, and requests are sent based on costs and prices. The second stage is the evaluation stage, in which each service provider is evaluated according to multiple criteria such as resource availability, response time, and reliability. The third stage is the pricing stage. For general services, a meta-heuristic algorithm is used to provide prices to service providers using the evaluation results, and the quantity of shared resources of providers is managed proportionally according to requests and prices. For proprietary services, in addition to the evaluation results of service providers, temporal features such as historical sales volume, continuous features such as resource inventory, and categorical features such as promotion methods will also be comprehensively considered, and dynamic pricing is performed using reinforcement learning. By specifying appropriate pricing strategies for different types of services, resource usage is effectively optimized, and the profit of service providers in the cloud federation is maximized.

[0061] The evaluation period of the periodic evaluation is determined by the agent, where requests are pre-sent by the agent to cloud service providers (CSPs). Each CSP has some virtual machines of CPU and GPU types, which are declared to the agent at the start of the period, including the corresponding costs and prices. The agent allocates CSP virtual machines to requests and converts them into CSP-based revenues. The relationship between the revenues obtained during the period and the resources provided is expressed as Equation 1, and the revenue allocation corresponds to the reserved resources of the CSP. Here, Brk refers to the resource requirements for which the agent obtains requests from users and sends the requests to the CSP.

[0062] Allocation=Provision+Brk (1)

[0063] As long as the revenue is not always equal to the investment and the entire investment resources may not be sold, the formula needs to be modified as shown in Equation 2. alpha is a constant, which is a ratio in the economy and is obtained from the linear regression between revenue and investment balance over a period of time, that is, the investment balance is determined after each request, and then the revenue is obtained based on the linear regression of the requests, and it is inversely proportional to the revenue.

[0064] Allocation=Provision-alpha*coefficient_income (2)

[0065] After that, the expected CSP revenue for the same request is defined according to the theory in macroeconomics, as shown in Equation 3. The ratio of the reserve to the average price is equal to the funds that the CSP expects to receive, which is k times the revenue.

[0066]

[0067] On the other hand, since there is not always an equality relationship in the above equation, we calculate it as Equation 4 through the beta constant coefficient and linear regression.

[0068]

[0069] It can be seen that the coefficient income has a direct relationship with the distribution. Therefore, the investment and price average can be regarded as constants over a period of time, and the relationship between the distribution and the coefficient income is represented by two independent relationships, where the distribution has a direct and inverse relationship with the coefficient income. Then, the equilibrium coefficient-income is obtained from the intersection of the two, where a positive equilibrium coefficient represents a resource surplus and a negative equilibrium coefficient represents a resource shortage. On the one hand, if it is intended to reduce the resource surplus, the evaluation process can be continued to balance the resources in the main period in other periods to achieve the ideal supply. On the other hand, when there is a resource shortage, the resource price can be increased in accordance with the equilibrium ratio to control the demand.

[0070] The architecture of the method proposed in this embodiment is divided into three layers: cloud, agent, and user, as Figure 1 shown. The cloud layer includes cloud service providers that share service resources. The agent layer is the middle part. At the same time, the agent determines the time interval of the evaluation, obtains shared resources from the federation members and evaluates the federation members during the period, and then determines the evaluation results for the service providers at the end of the period, and proposes the best price for the future period to the service providers respectively.

[0071] At the end of each request, Formula 5 can be obtained, which is established between the allocation (ALC) and the provision. α is a constant value determined proportionally to the investment value, and the coefficient income is determined according to it. Generally, α is considered to be a value that makes the supply balance (ALCR) lower after each allocation, that is, the allocation is inversely proportional to the supply balance, and the relationship between the allocation and the coefficient income is also inversely proportional, which is determined by linear regression over a period of time, as Figure 2 shown by the blue line in.

[0072] ALC = Provision - α * ALCR (5)

[0073] The currency demand of the service provider is equal to the ratio of the reserve profit to the average profit. The average profit calculates the profit of the virtual machine. Therefore, the currency demand is the ratio of the expected profit of the virtual machine unit calculated through investment and average profit to the expected profit. If the currency demand is defined as the ratio of the average profit of the virtual machine to the profit obtained from the provision, then the currency demand is expected to be several times higher than the allocation. Therefore, after each request, the capital demand expected by the provider is calculated as shown in Equation 6. The reserve balance after the request is represented by the beta constant and the coefficient income, and the relationship between the allocation and the coefficient income (MN) is obtained using linear regression, as Figure 2 shown by the red line in

[0074]

[0075] At the end of each cycle, the agent first calculates the relationship between the allocation and the coefficient income from the demand income part, and then calculates the relationship between the allocation and the coefficient income using the currency demand part. The intersection of the two is the balance point of the coefficient income and the allocation.

[0076] The pricing algorithm proposed by this method consists of three stages. The first stage is the management resource request stage, in which the agent obtains information about the service provider and the user, determines the evaluation period, and distributes the request among the federation members. The second stage of the algorithm is the evaluation stage, in which the middleman evaluates the federation members according to economic concepts and determines the income equilibrium coefficient of the federation members. The third stage is pricing, which determines the optimal resources of the CSP and proposes the optimal price of the CSP.

[0077] Furthermore, the present invention designs and implements a dynamic pricing optimization system for a cloud federation platform. There are various types of cloud services in the cloud federation platform, and a single pricing strategy cannot well adapt to the characteristics of different cloud services. Therefore, in the system designed by the present invention, cloud services are divided into general-purpose cloud services and dedicated cloud services. This division method takes into account on the one hand the different characteristics of the services themselves. Low-overhead resources do not involve high-performance hardware and have stable upstream prices. On the other hand, it also takes into account reducing the overhead of the pricing system itself. The price of general-purpose services is relatively stable, and the heuristic algorithm is used to calculate with low overhead. While dedicated cloud services are greatly affected by the real-time environment. In addition, the user groups of these two types of services also have differences. Customers of general-purpose cloud services pay more attention to the solution of the service itself and are not sensitive to resources. While dedicated cloud services often require exclusive resources and are more sensitive to the price of the resources of the cloud service itself.

[0078] The system module architecture proposed by the present invention is as Figure 4As shown in the figure, it mainly includes a resource request module, a service provider evaluation module, a cloud service scheduling module, a general cloud service pricing module, and a dedicated cloud service pricing module. The introduction of each module is as follows:

[0079] Resource request module: The resource request module is responsible for requesting virtual machine resources, computing resources such as CPUs and GPUs, storage resources such as Memory, uplink and downlink bandwidth, network resources such as public IPs, and the current cost prices of each resource from different service providers in the cloud federation computing platform, and representing them in the form of a shared resource array. The pseudocode for this process is as follows:

[0080] begin

[0081] 1. Define the evaluation period, which contains N requests within the period

[0082] 2. Accept each CSP resource and save it locally

[0083] 3. Accept the cost and price of each CSP

[0084] 4. while(period < N)

[0085] 5. Accept the customer request and save it in the storage

[0086] 6. Calculate the request distribution of the CSP

[0087] 7. Set the allocation according to the VM storage number and type of each CSP

[0088] 8. Update the shared resources, that is, after the allocation, update the remaining number of virtual machines in the array

[0089] 9. Set the shared resources, that is, after the allocation, save the remaining number of virtual machines in the array

[0090] 10. end while

[0091] 11. return(allocation array, shared resource array)

[0092] Service provider evaluation module: The service provider evaluation module is responsible for evaluating multiple evaluation criteria such as the availability, response time, and reliability of the virtual machine resources provided by the service provider, evaluating the system revenue of the service provider according to economic concepts, using an algorithm to calculate the two-line equations of each provider using linear regression, and calculating and storing the equilibrium array of each provider using the intersection of the two lines to obtain the balance point of the allocation and coefficient revenue. The pseudocode for this process is as follows:

[0093] begin

[0094] 1. Read the allocation array and the shared resource array

[0095] 2. while (both arrays are not empty)

[0096] 3. Calculate the CSP allocation according to the allocation array

[0097] 4. Calculate the reserve of the CSP based on the shared resource array

[0098] 5. Calculate the coefficient income allocation for each CSP

[0099] 6. Calculate the reserve for each CSP

[0100] 7. Calculate and set the array of (savings allocation, coefficient income)

[0101] 8. end while

[0102] 9. for each CSP

[0103] 10. Regress the allocation and coefficient income according to Formulas 5 and 6

[0104] 11. Calculate the equilibrium coefficient income and equilibrium reserve

[0105] 12. Set the set of equilibrium coefficient income and equilibrium reserve, and save the equilibrium array

[0106] 13. end

[0107] Cloud service scheduling module: The cloud service scheduling module is responsible for classifying cloud service types and scheduling them to the general service pricing module and the dedicated service pricing module for pricing according to the cloud service category. The classification method of the service refers to the computing and storage overhead of the service itself, whether the service is an online service or an offline service, and whether it occupies exclusive resources.

[0108] For general services, they are not sensitive to resources and have low overhead. For such services, users value the solution of the service itself more. Service providers can adopt the tidal strategy or elastic scaling to enable multiple cloud services of the same type to share resources. Therefore, the pricing is relatively stable and not easily affected by the external environment.

[0109] For dedicated services, in addition to valuing the solution of the cloud service itself, users have additional requirements for either computing or storage, are greatly affected by external environments such as the supply chain, and often need to occupy exclusive resources.

[0110] General service pricing module: The specific algorithm process of the general service pricing module is shown in the following pseudocode:

[0111] begin

[0112] 1. Read the equilibrium array

[0113] 2. while (the equilibrium array is not empty)

[0114] 3. Calculate the CSP allocation based on the allocation array

[0115] 4. if (equilibrium coefficient return < 0.10)

[0116] 5. Call Algorithm 2 to recalculate and evaluate the new shared resources

[0117] 6. end while

[0118] 7. Calculate and return the pricing

[0119] This algorithm only provides prices to service providers using the results of service provider evaluations. This algorithm manages the quantity of shared resources of providers according to the ratio of requests to prices, and provides the most favorable prices to cloud service providers based on evaluation results to maximize their profits.

[0120] For general-purpose services, the price is relatively stable and less affected by the resources themselves. The cloud federation computing platform side is more inclined to keep the price in a relatively stable state to reduce customers' perception of cloud service resources.

[0121] In addition, the calculation overhead of the general service pricing module is also relatively lower than that of the dedicated service pricing module.

[0122] Dedicated service pricing module: The dedicated service pricing module is for cloud services with obvious resource requirements. In addition to considering the evaluation results of service providers, it also needs to consider factors such as the historical price change trend of the service and the external environment. The pricing model of dedicated services is Figure 3 as shown, a dynamically changing process. In view of the characteristics of dedicated services, the present invention designs the features in Table 1 below for pricing:

[0123] Table 1: Corresponding table of features and feature types

[0124]

[0125] The input of the deep reinforcement learning dynamic pricing algorithm is each state component in the Markov decision process. The feature types of the state components defined in the present invention include historical features, sequence features, and binary features. By integrating the three types of features, it not only pays attention to the historical data of cloud services but also the current environmental state.

[0126] In the cloud federation computing platform pricing scenario faced by the present invention, the goal of reinforcement learning is to learn a policy π that can dynamically adjust the service price according to the current state component features to maximize the expected cumulative return:

[0127]

[0128] Among them, τ represents the complete pricing sequence data, and r(τ) represents the total return of the pricing sequence data. To maximize the return, the core lies in the calculation of R(τ). The present invention uses the currently excellent reinforcement learning evaluation framework A3C (Asynchronous advantage actor-critic) for the calculation. The A3C algorithm framework uses the N-step cumulative discounted return to estimate r(τ):

[0129]

[0130] The features are encoded using LSTM to retain the sequence position information of historical data. The cell structure of LSTM is defined as follows:

[0131] Input gate: i t = σ(W i x t + U i h t-1 + b i )

[0132] Forget gate: f t = σ(W f x t + U f h t-1 + b f )

[0133] Output gate: o t = σ(W i x t + U i h t-1 + b i )

[0134] Candidate state:

[0135] Memory cell:

[0136] External state: h t = o t ⊙ tanh(c t )

[0137] Based on A3C and LSTM, a neural network can be constructed for the dynamic pricing of dedicated services. The input features of the network include historical features, sequential features, and binary features. Among them, the historical features contain the sales volume change sequence of cloud services and the price change sequence in the past month. These time-series data are processed using LSTM for feature extraction to retain the sequential dependencies of information. The binary features and sequential features include the duration from the pricing point to the sales end point, the stock of basic resources of each cloud service at the time of pricing, such as CPU, GPU, and Memory, etc., and the service provider's evaluation results, which are extracted through entity embedding. Finally, a multi-layer feedforward neural network is used for non-linear mapping to comprehensively extract high-order feature representations, considering the impacts of different types of features to make the pricing strategy more robust.

[0138] Furthermore, the dynamic pricing adopts a periodic adjustment strategy. For each round of pricing and evaluation cycle, which includes several cloud service resource requests, the following pricing process steps will be carried out:

[0139] 1. Resource Request Module: Each evaluation cycle contains N resource requests, and the evaluation cycle will be adjusted regularly according to the activities and time of the cloud platform itself. All resource requests within the evaluation cycle will be incorporated into the dynamic pricing adjustment.

[0140] 1) Request the resources and costs of all cloud service providers CSP on the cloud federation platform to construct a shared resource pool G:

[0141] G = CSP(1) ∪ CSP(2)…CSP(n) (9)

[0142] The details of CSP i are shown in Table 2:

[0143] Table 2: Detailed list of items included in CSP i

[0144] CPU total GPU total Memory total Network Online service1 Online server2 x1i x2i x3i x4i x5i x6i

[0145] The details of G(CSP total) are shown in Table 3:

[0146] Table 3: Detailed list of items included in G(CSP total)

[0147]

[0148] 2) For each resource request L from a customer, calculate the resource allocation of its corresponding resources in the cloud service provider and the remaining resource data array G of the shared resource pool after each allocation ′ :

[0149] G ′ = G - L(x1,x2…xn) (10)

[0150] And calculate the allocation array after allocating L to all cloud service providers CSPs that meet the requirements:

[0151] CSPx ′ = CSPx - L (11)

[0152] G1 = G ′ = CSP(1)' ∪ CSP(2)…CSPx…CSP(n)

[0153] G2 = G ′ = CSP(1) ∪ CSP(2)'…CSPx…CSP(n)

[0154] Set(G) = {G1, G2…Gn}

[0155] Among them, G1 represents the situation where L is allocated to CSP1, G2 represents the allocation to CSP2, and Set(G) represents all possible allocations.

[0156] 2. Service provider evaluation module: Calculate the equilibrium array based on the pricing framework of economic concepts

[0157] First, calculate the resources after CSP allocation according to the allocation array, and then calculate the reserves of CSPs based on each different resource allocation (resources are allocated to different CSP combinations) and shared resources based on the shared resource array.

[0158]

[0159] Among them, Provision represents the reserve, and mean_price represents the average price in the previous period.

[0160] Subsequently, for each service provider, use linear regression to calculate the balance point of the supplier allocation and revenue coefficient:

[0161]

[0162] As Figure 2 shown, where ALC represents the resource allocation (i.e., all requests), Provision represents the provision of CSP, and ALCR represents the supply balance.

[0163] 3. Service provider scheduling module: For each resource request within the cycle, divide it according to the required cloud service category and then perform price pricing.

[0164] G geneal ∪G spec = G

[0165] The division standard conditions are as follows:

[0166] General-purpose service G geneal: Offline resources, with relatively low resource overhead, allowing for elastic scaling, relatively stable pricing, and placing more emphasis on complete solutions.

[0167] Dedicated service Gspec: Exclusive resources, high guarantee, and the price is greatly affected by the resource stock.

[0168] 4. General service pricing module: Price each general service request within a cycle.

[0169]

[0170] General service pricing only uses the results evaluated by the service provider to provide prices. On the premise of meeting the reserve and equilibrium coefficient for its revenue balance, when the revenue coefficient is greater than 0.1, the price is returned. When the revenue coefficient does not meet the conditions, the request will not be scheduled to this CSP. When all CSPs do not meet the request, the service provider is re-evaluated.

[0171] 5. Dedicated service pricing module: Dynamically adjust the pricing for each dedicated service request within a cycle.

[0172] Collect historical features, sequence features, and binary features for the service as described in the dedicated service pricing module, and construct a feature vector. Subsequently, use A3C, LSTM, and MLP to construct a neural network. The LSTM part is used to process sequence data, and the A3C part is used to make pricing decisions based on the processed features.

[0173] The reward function is defined as maximizing the cumulative revenue:

[0174]

[0175] Construct a feature vector based on the request pricing of the current round of the cycle to train the pricing model for the next round of the cycle. The pricing for the current cycle is obtained by model prediction after constructing the feature vector:

[0176] p(L) = model_predicat(τ)

[0177] All the processes of the cloud service provider dynamic pricing method disclosed in the above embodiments can be embedded in an electronic device for operation. The electronic device includes a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface complete communication with each other through the communication bus. The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute all the process steps of the cloud service provider dynamic pricing method disclosed in the above embodiments, which will not be elaborated here.

[0178] The electronic device can also communicate with one or more external devices (such as a keyboard, a pointing device, a Bluetooth device, etc.), and can also communicate with one or more devices that enable a user to interact with the electronic device, and / or communicate with any device that enables the electronic device to communicate with one or more other computing devices (such as a router, a modem, etc.). Such communication can be carried out through an input / output (I / O) interface. Moreover, the electronic device can also communicate with one or more networks (such as a local area network LAN), a wide area network WAN, and / or a public network, such as the Internet, through a network adapter. The network adapter communicates with other modules of the electronic device through a bus. It should be understood that although not shown in the figures, other hardware and / or software modules can be used in combination with the electronic device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0179] It should be noted that the above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center that contains one or more sets of available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0180] It should be understood that in various embodiments of the present application, the magnitudes of the serial numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0181] It should be understood that in various embodiments of the present application, the magnitudes of the serial numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

Claims

1. A cloud service provider dynamic pricing method, characterized in that: The steps include: Obtain the inventory resource information of virtual machines from the cloud service provider and send resource requests based on cost and price; For each resource request, calculate the resource allocation of the required resources in the cloud service provider and give an allocation array; For each resource allocation situation in the allocation array, a balance point of allocation and revenue coefficient of the cloud service provider is calculated by linear regression; Classify each resource request into a general service request or a dedicated service request according to the amount of resource overhead; For each general service request within the period, a meta-heuristic algorithm is used to perform pricing based on the distribution and revenue coefficient balance point; For each dedicated service request within the cycle, based on the balance point of the allocation and revenue coefficient, dynamic pricing adjustment is performed using reinforcement learning, taking into account the historical sales time series characteristics, resource inventory continuity characteristics and promotion method classification characteristics; Output pricing results.

2. A cloud service provider dynamic pricing method according to claim 1, characterized in that: Obtain virtual machine resource information and inventory resources from the cloud service provider, including: Request resources and costs from all cloud service providers on the cloud federation platform to build a shared resource pool G: G=CSP(1)∪CSP(2)∪…∪CSP(n) In the formula, CSP(n) represents the amount of resources that the nth cloud service provider can provide; each cloud service provider CSP includes the total CPU resources, total GPU resources, memory resources, network resources, and the number of online servers.

3. A cloud service provider dynamic pricing method according to claim 1, characterized in that: For each resource request L, calculate the resource allocation of the required resources in the cloud service provider, as well as the remaining resource data array G in the shared resource pool after each allocation ′ , and calculate the allocation array after allocating the resource request L to all satisfied cloud service providers CSP: G ′ =G-L(x1,x2…xn) CSP(i) ′ =CSP(i)-L Set(G)={G1,G2…Gn} Where G1 represents the case where resource request L is allocated to CSP(1); G2 represents the case where resource request L is allocated to CSP(2); Gn represents the case where resource request L is allocated to CSP(n); CSP(i) represents the amount of resources that the i-th cloud service provider can provide; CSP(i) ′ It represents the amount of resources remaining after the i-th cloud service provider allocates the required resources to the resource request L; Set(G) represents the allocation array formed by all possible allocation sets.

4. A cloud service provider dynamic pricing method according to claim 3, characterized in that: The equilibrium point p of the cloud service provider's distribution and revenue coefficient is calculated using linear regression: Where Provision represents the amount provided by CSP; ALCR represents the supply balance; MN represents the allocation and income coefficient; β is the proportionality coefficient.

5. A cloud service provider dynamic pricing method according to claim 3, characterized in that: Each resource request is divided into a general service request or a dedicated service request according to the amount of resource overhead, including: If the resource request L allocated to the i-th cloud service provider Gi is less than the preset value, it is defined as a general service request G geneal ; If the resource request L is allocated to the i-th cloud service provider Gi is greater than or equal to the preset value, it is defined as a dedicated service request G spec ; G geneal ∪G spec =G.

6. A cloud service provider dynamic pricing method according to claim 5, characterized in that: For each general service request G in the cycle geneal , based on the distribution and income coefficient balance point p, when the distribution and income coefficient MN is less than 0.1, the service provider is re-evaluated: if(MN<0.1)reAllocate(L) Otherwise, the allocation is completed directly based on the current allocation and the income coefficient balance point p, and the pricing is returned: Where p(L) means that the resource request L is allocated according to the current allocation and the revenue coefficient balance point p; reAllocate(L) means that the resource request L re-evaluates the service provider.

7. A cloud service provider dynamic pricing method according to claim 1, characterized in that: For each dedicated service request within the cycle, based on the balance point of the allocation and revenue coefficient, dynamic pricing adjustment is performed using reinforcement learning, taking into account the historical sales time series characteristics, resource inventory continuity characteristics and promotion method classification characteristics; The goal of reinforcement learning is to learn a strategy π that dynamically adjusts the service price according to the current state component characteristics to maximize the expected cumulative benefits: In the formula, τ represents the complete pricing series data; r(τ) represents the total revenue of the pricing series data; Based on the request pricing of the current cycle, a feature vector is constructed to train the pricing model for the next cycle; the pricing of the current cycle is obtained by constructing a feature vector and then predicting it by the model.

8. A dynamic pricing optimization system for a cloud federation platform, characterized in that: include: The resource request module is used to obtain the inventory resource information of the virtual machine from the cloud service provider and send resource requests based on the cost and price; The service provider evaluation module is used to calculate the resource allocation of each resource request in the cloud service provider and provide an allocation array; for each resource allocation in the allocation array, the cloud service provider's allocation and revenue coefficient balance point is calculated by linear regression; The service provider scheduling module is used to classify each resource request into a general service request or a dedicated service request according to the amount of resource overhead; A general service pricing module, for pricing each general service request within a period using a meta-heuristic algorithm based on the distribution and revenue coefficient balance point, and outputting a pricing result; The dedicated service pricing module is used to perform dynamic pricing adjustment using reinforcement learning for each dedicated service request within the cycle, based on the balance point of the allocation and revenue coefficients, taking into account the historical sales time series characteristics, resource inventory continuity characteristics and promotion method classification characteristics, and outputting the pricing results.

9. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the cloud service provider dynamic pricing method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The storage medium stores at least one executable instruction, and when the executable instruction is executed on the electronic device, the electronic device executes the cloud service provider dynamic pricing method according to any one of claims 1 to 7.

Images