Computing power resource adjustment method and device, equipment, storage medium and program product

Abstract

The invention discloses a computing power resource adjustment method and apparatus, a device, a storage medium and a program product. The method comprises the steps of obtaining a tenant identifier and computing power resource information of any tenant of a cloud platform; wherein the computing power resource information comprises an initial load utilization rate and an initial computing power node number; determining an estimated computing power node number of the tenant according to the tenant identifier and the computing power resource information; sending the pre-estimated computing power node number to the cloud platform, and obtaining a reference load utilization rate after the cloud platform adjusts the computing power resource of the tenant based on the pre-estimated computing power node number; evaluating the reference load utilization rate to obtain an evaluation result; and when the evaluation result is that the number of the pre-estimated computing power nodes is not accurate, adjusting the number of the pre-estimated computing power nodes according to the evaluation result to obtain a target computing power node number, and sending the target computing power node number to the cloud platform. By adopting the embodiment of the invention, the regular or irregular fluctuation of the service can be accurately adapted, the service congestion is avoided, and meanwhile, the deployment and operation and maintenance cost of the cloud platform is reduced.

Description

Technical Field

[0001] This invention relates to the field of cloud computing, and more particularly to a method, apparatus, device, storage medium, and program product for adjusting computing resources. Background Technology

[0002] A cloud platform refers to an integrated environment built on cloud computing technology, providing users with infrastructure resources such as computing, storage, and networking, as well as related services. Elastic scaling is one of the core capabilities of cloud platform resource management. By automatically adjusting the number of resource nodes on demand, it can quickly meet the needs of fluctuating business loads while reducing the cost of idle computing resources. Currently, there are two main types of elastic scaling solutions: one is time-based scaling, which expands or shrinks capacity within a specified time period, suitable for scenarios where business volume changes have predictable time patterns or where there are stable requirements for the number of nodes; the other is load-based scaling, which expands or shrinks capacity according to changes in business volume. However, time-based scaling relies excessively on the stability of business patterns. In the event of unexpected situations (such as an early surge or delay in business volume), it can easily lead to insufficient resource supply and business congestion. While load-based scaling can adapt to irregular business fluctuations, it requires a high-cost automated management and monitoring system, increasing the deployment and maintenance costs of the cloud platform. Summary of the Invention

[0003] The purpose of this invention is to provide a method, apparatus, device, storage medium, and program product for adjusting computing resources, which can accurately adapt to the regular or irregular fluctuations of business, avoid business congestion, and reduce the deployment and maintenance costs of cloud platforms.

[0004] To achieve the above objectives, embodiments of the present invention provide a method for adjusting computing resources, including: Obtain the tenant identifier and computing resource information of any tenant on the cloud platform; wherein, the computing resource information includes the initial load utilization rate and the initial number of computing nodes; Based on the tenant identifier and the computing power resource information, determine the estimated number of computing power nodes for the tenant; The estimated number of computing nodes is sent to the cloud platform, and the reference load utilization rate after the cloud platform adjusts the computing resources of the tenant based on the estimated number of computing nodes is obtained. The reference load utilization rate is evaluated to obtain the evaluation results; When the evaluation result indicates that the estimated number of computing power nodes is inaccurate, the estimated number of computing power nodes is adjusted according to the evaluation result to obtain the target number of computing power nodes, and the target number of computing power nodes is sent to the cloud platform.

[0005] As an improvement to the above scheme, determining the estimated number of computing power nodes for a tenant based on the tenant identifier and the computing power resource information includes: Query the experience database corresponding to the tenant identifier; When the historical computing resource information corresponding to the computing resource information is found in the experience database, the estimated number of computing nodes for the tenant is determined based on the historical computing resource information. When no historical computing resource information corresponding to the computing resource information is found in the experience database, the estimated number of computing nodes for the tenant is calculated using the computing resource information and a preset load threshold.

[0006] As an improvement to the above solution, determining the estimated number of computing power nodes for the tenant based on the historical computing power resource information includes: Extract the corresponding historical final values ​​and corresponding probability distribution features from the historical computing power resource information; Based on the probability distribution characteristics of the historical final values, the estimated number of power nodes for the tenant is determined.

[0007] As an improvement to the above scheme, the load threshold includes a scaling-up threshold and a scaling-down threshold. The step of calculating the estimated number of computing power nodes for the tenant using the computing power resource information and the preset load threshold includes: When the initial load utilization rate is greater than the expansion threshold, the estimated number of computing nodes for the tenant is calculated based on the initial load utilization rate, the initial number of computing nodes, and the expansion threshold. When the initial load utilization is less than the scaling-down threshold, the estimated number of computing nodes for the tenant is calculated based on the initial load utilization, the initial number of computing nodes, and the scaling-down threshold.

[0008] As an improvement to the above scheme, the step of calculating the estimated number of computing nodes for the tenant based on the initial load utilization rate, the initial number of computing nodes, and the expansion threshold includes: The product of the initial load utilization rate and the initial number of computing nodes is calculated as the computing load reference value; Calculate the first ratio between the computing power load reference value and the expansion threshold; The first ratio is rounded up to obtain the estimated number of power nodes for the tenant.

[0009] As an improvement to the above solution, the step of calculating the estimated number of computing nodes for the tenant based on the initial load utilization, the initial number of computing nodes, and the scaling-down threshold includes: The product of the initial load utilization rate and the initial number of computing nodes is calculated as the computing load reference value; Calculate a second ratio between the computing power load reference value and the scaling-down threshold; The second ratio is rounded down to obtain the estimated number of power nodes for the tenant.

[0010] As an improvement to the above solution, before sending the estimated number of power nodes to the cloud platform, the method further includes: When the current computing power resource adjustment operation is an expansion operation, the total number of computing power nodes currently allocated to all tenants is counted, and the reference number of nodes to be expanded in this operation is summarized. Add the sum of the number of computing power nodes to the number of reference nodes to obtain the estimated total occupancy; If the estimated total occupancy is greater than or equal to the total resources of the cloud platform, the current computing power resource adjustment operation is stopped; if the estimated total occupancy is less than the total resources of the cloud platform, the step of sending the estimated number of computing power nodes to the cloud platform is executed.

[0011] As an improvement to the above scheme, the evaluation of the reference load utilization rate to obtain the evaluation result includes: When the current computing power resource adjustment operation is an expansion operation, if the reference load utilization rate is less than or equal to the expansion threshold, the evaluation result is that the estimated number of computing power nodes is accurate; if the reference load utilization rate is greater than the expansion threshold, the evaluation result is that the estimated number of computing power nodes is inaccurate.

[0012] As an improvement to the above scheme, adjusting the estimated number of computing power nodes based on the evaluation results to obtain the target number of computing power nodes includes: The target number of computing power nodes is obtained by adding a preset expansion step size to the estimated number of computing power nodes.

[0013] As an improvement to the above scheme, the evaluation of the reference load utilization rate to obtain the evaluation result includes: When the current computing power resource adjustment operation is a scaling-down operation, if the reference load utilization rate is greater than or equal to the scaling-down threshold, the evaluation result is that the estimated number of computing power nodes is accurate; if the reference load utilization rate is less than the scaling-down threshold, the evaluation result is that the estimated number of computing power nodes is inaccurate.

[0014] As an improvement to the above scheme, adjusting the estimated number of computing power nodes based on the evaluation results to obtain the target number of computing power nodes includes: The target number of computing power nodes is obtained by reducing the preset callback step size based on the estimated number of computing power nodes.

[0015] As an improvement to the above solution, before sending the estimated number of power nodes to the cloud platform, the method further includes: When the current computing power resource adjustment operation is an expansion operation, if the tenant does not belong to the preset high-priority whitelist, the tenant voting mechanism is activated to verify the preset number of computing power nodes. If the verification fails, stop the current computing power resource adjustment operation; if the verification passes, execute the step of sending the estimated number of computing power nodes to the cloud platform.

[0016] As an improvement to the above scheme, the tenant voting mechanism includes: Based on the tenant identifier of the target tenant, the reference computing power node number for each is generated by sampling from the experience database corresponding to the target tenant; wherein, the target tenant is the tenant other than the current tenant; If any reference computing power node number is greater than the estimated computing power node number, then vote in favor; otherwise, vote against. After all target tenants have completed their voting, calculate the percentage of affirmative votes out of the total votes. If the approval rate is greater than or equal to a preset threshold, the verification is deemed successful; if the approval rate is less than the threshold, the verification is deemed unsuccessful.

[0017] To achieve the above objectives, embodiments of the present invention also provide a computing power resource adjustment device, comprising: A generator is used to obtain the tenant identifier and computing resource information of any tenant on the cloud platform; wherein, the computing resource information includes the initial load utilization rate and the initial number of computing nodes; based on the tenant identifier and the computing resource information, the generator determines the estimated number of computing nodes for the tenant; the generator sends the estimated number of computing nodes to the cloud platform, and obtains the reference load utilization rate after the cloud platform adjusts the computing resources of the tenant based on the estimated number of computing nodes; A discriminator is used to evaluate the reference load utilization, obtain an evaluation result, and when the evaluation result indicates that the estimated number of computing power nodes is inaccurate, adjust the estimated number of computing power nodes according to the evaluation result to obtain the target number of computing power nodes, and send the target number of computing power nodes to the cloud platform.

[0018] As an improvement to the above solution, the computing power resource adjustment device further includes an interface module, which provides a data interaction channel between the computing power resource adjustment device and the cloud platform.

[0019] As an improvement to the above solution, the interface module includes: A first interface is used to transmit a first request message and a first response message; wherein the first request message carries the tenant identifier and the computing power resource information, and the first response message carries the estimated number of computing power nodes; The second interface is used to transmit a second request message and a second response message; wherein the second request message carries the reference load utilization rate, and the second response message carries the target number of computing power nodes; The first interface and the second interface verify data through a signature mechanism.

[0020] To achieve the above objectives, embodiments of the present invention also provide a computing resource adjustment device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the computing resource adjustment method as described in any of the above embodiments.

[0021] To achieve the above objectives, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the computing resource adjustment method as described in any of the above embodiments.

[0022] To achieve the above objectives, embodiments of the present invention also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the computing resource adjustment method as described in any of the above embodiments.

[0023] Compared to existing technologies, the computing power resource adjustment method, apparatus, device, storage medium, and program products disclosed in this invention dynamically acquire tenant load and node information, accurately determine the estimated number of computing power nodes by combining tenant identifiers, and conduct evaluation and secondary optimization based on the actual load feedback after adjustment, ultimately outputting the appropriate target number of computing power nodes. This can accurately adapt to the regular or irregular fluctuations of business operations, avoid business congestion, and reduce the deployment and operation and maintenance costs of the cloud platform. In addition, through the closed-loop mechanism of initial estimation and feedback adjustment, the allocation of computing power resources closely matches the actual load of tenants, avoiding resource surplus or shortage and improving resource utilization efficiency; the evaluation and adjustment based on real-time load utilization can quickly respond to fluctuations in tenant workload intensity and ensure the stability of business operations; combined with the historical characteristics or configuration rules associated with tenant identifiers, differentiated resource allocation is achieved, taking into account the needs of individual tenants and the overall resource balance of the cloud platform; through the automated evaluation and adjustment of the entire process, manual intervention is reduced, improving the efficiency and intelligence level of cloud platform computing power resource scheduling, and supporting fair and efficient resource sharing in multi-tenant scenarios. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the application of the generative adversarial network model provided in this embodiment of the invention; Figure 2 This is a flowchart of a computing resource adjustment method provided in an embodiment of the present invention; Figure 3 This is a flowchart of determining the pre-estimated force node provided in an embodiment of the present invention; Figure 4 This is a flowchart provided by an embodiment of the present invention for determining whether to send the estimated number of force nodes to the cloud platform; Figure 5 This is another application diagram of the generative adversarial network model provided in this embodiment of the invention; Figure 6 This is another flowchart provided by an embodiment of the present invention for determining whether to send the estimated number of force nodes to the cloud platform; Figure 7 This is a structural block diagram of the dynamically scalable DSCS and cloud platform provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of information interaction between the dynamically scalable DSCS and the cloud platform provided in an embodiment of the present invention; Figure 9 This is a structural block diagram of a computing resource adjustment device provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In the cloud computing field, when facing large amounts of real-time data and changing business loads, real-time monitoring and elastic scaling of cloud platform resources can improve resource utilization and user experience. Autoscaling needs to be coordinated with load balancing (or task balancing). The resources used for autoscaling are cloud hosts, physical servers, and other resources bound to the load balancer's backend service nodes. Autoscaling triggers scaling events based on monitoring metrics such as the average CPU (Central Processing Unit) load of the backend service nodes. These events create new servers according to the configured resource template or remove servers according to a removal policy. Autoscaling performs scaling actions by monitoring metrics such as cloud host CPU usage. It can also be achieved by setting time periods to create autoscaling policies, thus automatically adding or removing cloud hosts from the bound load balancer backend nodes.

[0027] It should be noted that automatic scaling technology includes the following basic concepts: 1) Auto-scaling instance: Used to manage and configure multiple scaling tasks. All scaling policies, scaling templates, and scaling records are contained in the auto-scaling instance; 2) Trigger / Scaling Policy: Used to monitor startup conditions, including monitoring policies for indicators such as CPU and memory, as well as set periodic scaling rules. When the startup conditions are met, the scaling policy will be triggered and corresponding actions will be executed. 3) Scaling Template: A configuration template for cloud servers, specifying parameters such as CPU, memory, image, system disk, data disk, and payment method for the cloud server. When expansion is triggered, a new cloud server is created according to the configuration of the scaling template. 4) Scale-up records: Scale-up and scale-down records are used to audit the execution of automatic scaling tasks.

[0028] This invention utilizes autoscaling technology to automate the adjustment of computing resources on a cloud platform. The controller within the cloud platform continuously monitors resource metrics and tasks in the schedule. Among the monitored metrics, when the average CPU load of cloud hosts in the load balancer increases to H (the expansion threshold), a scaling-up strategy is initiated; when the average load decreases to L (the shrinking threshold), a shrinking strategy is initiated. Monitoring is performed through a set scaling time period, and the set scaling strategy is also triggered when the time conditions are met. This invention supports custom monitoring methods, sending events to triggers when expected conditions are met, and then executing the scaling strategy. Simultaneously, the autoscaling management service is triggered, creating N cloud host instances according to the configuration template and recording them in the scaling record. These cloud host instances are bound to the backend service nodes of the load balancer, and the load balancer performs health checks on these N cloud host instances before starting request allocation, at which point the average load will decrease. If cloud hosts are to be deleted, the system also considers which cloud hosts to delete, with deletion strategies including prioritizing the deletion of the most recently created cloud hosts, prioritizing the deletion of the oldest created cloud hosts, and random deletion. After scaling, the scaling policy will be updated. For example, if the average CPU load of the cloud server increases or decreases, or if the next scaling cycle has not yet arrived, the scaling policy will be updated in real time and executed according to the latest scaling policy, and this process will continue. At the same time, all expansion and contraction operations will be recorded in the scaling record.

[0029] Furthermore, this invention introduces GAN (Generative Adversarial Network) into the automatic scaling technology. GAN is a generative model that learns by having two neural networks compete against each other, enabling it to learn generative tasks without using labeled data. A GAN consists of a generator and a discriminator. The generator randomly samples from the latent space as input, and its output should mimic real samples in the training set as closely as possible. The discriminator's input is either a real sample or the generator's output, and its goal is to distinguish the generator's output from real samples as much as possible. The generator and discriminator compete against each other and continuously learn, ultimately making it impossible for the discriminator to determine whether the generator's output is real.

[0030] This invention combines autoscaling technology with GANs, assuming the total CPU resources of the database cloud platform are M, and the allocated resources for each tenant are Pi, where i represents the tenant ID. Without considering resource over-allocation or reuse, the system's resource constraint at any given time is: M ≥ ∑Pi.

[0031] See Figure 1 , Figure 1 This is a schematic diagram illustrating the application of the Generative Adversarial Network (GAN) model provided in this embodiment of the invention. The generator takes the load index z of the cloud platform controller as input and samples it from the tenant's experience database Pdata to obtain sampled data P(z). After P(z) passes through the generator's neural network, it generates a predicted number of load nodes x, which serves as the initial prediction value. Based on the initial prediction value x, the controller scales up and down the tenant resources of the cloud platform. After the controller performs scaling, the discriminator judges the truth value x by checking the latest index of the tenant monitored by the controller. x is compared with the latest index. If it is a scaling-up scenario and the monitored index is below H, it is judged as true, and the final value of the sample is x; otherwise, it is false, and the final scaling value needs to be x+a (a preset scaling-up step size, which can be set empirically). If it is a scaling-down scenario and the monitored index is above L, it is true, and the final value of the sample is x; otherwise, it is false, and the final scaling value needs to be xb (a preset callback step size, which can be set empirically). The discriminator inserts the final value y into the experience database. Furthermore, the present invention updates the neural network parameters of the generator and discriminator using gradient descent based on the obtained final value information. In the subsequent game, the generator and discriminator learn continuously, so that the scaling value suggested by the generator can pass the discriminator's judgment with the highest probability and match the resource increase or decrease needs of the cloud database platform tenant.

[0032] See Figure 2 , Figure 2 This is a flowchart of a computing resource adjustment method provided in an embodiment of the present invention, which includes steps S1 to S5.

[0033] S1. Obtain the tenant identifier and computing resource information of any tenant on the cloud platform; wherein, the computing resource information includes the initial load utilization rate and the initial number of computing nodes.

[0034] For example, the tenant identifier is used to indicate the identity of different tenants in the cloud platform, ensuring that the computing resource adjustment operation for each tenant can be accurately located to a specific entity, avoiding confusion in resource scheduling between tenants; the computing resource information represents the current computing resource occupancy and load operation status of the tenant, providing a data basis for subsequent resource adjustments; the initial load utilization rate represents the actual load ratio of the computing nodes currently allocated to the tenant (such as the average CPU utilization rate), reflecting the busyness of existing resources; the initial number of computing nodes represents the total number of computing nodes currently allocated by the cloud platform to the tenant to carry its jobs, which is a benchmark reference value for resource adjustments.

[0035] It should be noted that the controller in the cloud platform is responsible for receiving resource requests estimated by the generator for tenants and performing scaling operations; simultaneously, it monitors tenant resource usage and outputs resource metrics to the generator and discriminator. The monitoring period is T, and each monitoring moment is recorded as t. The data recorded in the experience database is in vector format e, and its dimensions are determined as needed. The data in this invention is defined as (u,t,p,n,y), where u represents the tenant identifier; t is the date and time of occurrence; p is the initial number of computing power nodes allocated to the tenant at time t; n is the tenant's current initial load utilization rate n; and y is the required target number of computing power nodes, i.e., the new final resource value output by the generative adversarial network model. The vector uses p and n as dimensions to represent the features of the model. To improve timeliness, the most recent m records in the experience database can be taken as the sampling sample space. The dimension of the resource metric z monitored by the controller for tenants is the same as the sampling format of the experience database, which is also a vector format (u,t,p,n,0). Since it is monitoring data, the target number of computing power nodes in z is temporarily 0.

[0036] S2. Determine the estimated number of computing power nodes for the tenant based on the tenant identifier and the computing power resource information.

[0037] For example, the data input into the generative adversarial network model is in the form of a load index z=(u,t,p,n,0). The generative adversarial network model determines the estimated load node of tenant u based on this load index.

[0038] Further, see Figure 3 , Figure 3 This is a flowchart of determining the pre-estimated force node provided by an embodiment of the present invention, wherein step S2 specifically includes steps S21 to S23.

[0039] S21. Query the experience database Pdata corresponding to the tenant identifier.

[0040] For example, the experience database Pdata is a structured data set that stores the tenant's historical computing resource adjustment records. It includes information such as the target number of computing nodes and the corresponding load change results for the tenant at different times, with different initial load utilization rates and different initial number of computing nodes. During a query, the current tenant's identifier is used as an index to locate its dedicated experience database Pdata. Then, based on the feature values ​​of the current computing resource information (such as initial load utilization rate and initial number of computing nodes), a sample P(z) is sampled from the experience database Pdata. This P(z) carries historical computing resource information.

[0041] S22. When historical computing resource information corresponding to the computing resource information is found in the experience database, the estimated number of computing nodes for the tenant is determined based on the historical computing resource information.

[0042] For example, when the historical computing resource information corresponding to the computing resource information is found in the experience database Pdata, it means that the tenant has made computing adjustments under similar load conditions and resource configurations, and the historical data can be used as a reference for the current adjustment. At this time, the estimated number of computing nodes can be determined based on the historical computing resource information.

[0043] Further, in step S22, determining the estimated number of computing power nodes for the tenant based on the historical computing power resource information includes: extracting the corresponding historical final value and the corresponding probability distribution feature from the historical computing power resource information; and determining the estimated number of computing power nodes for the tenant based on the probability distribution feature of the historical final value.

[0044] For example, P(z) is passed through the generator neural network G(z) to obtain a new sample x, which is the probability distribution to be solved: x = G(P(z)). For instance, the generator first receives the current load index z from the controller. The load index z carries information including: tenant A, p = 3, n = 70%. Then, it filters "resource adjustment records of tenant A (or similar tenants) in the past when 'p≈3, n≈70%'". These records are records of reasonable final resource values ​​y that can be determined when similar situations are encountered in the past. For example, three records are filtered out, with y values ​​of y1 = 4, y2 = 4, and y3 = 5, forming a similar experience sample set P(z). The generator neural network G(z) learns from the similar experience sample set P(z) and analyzes it to find that when p = 3 and n = 70%, y is usually between 4 and 5, and in most cases it is 4. Based on this pattern, the generator neural network G(z) will output a value that best fits the pattern as the estimated number of nodes x for the current tenant (e.g., x=4). This x is not a guess, but is derived from the pattern of similar historical cases to ensure the rationality of the initial estimate.

[0045] In this embodiment of the invention, by extracting historical final values ​​and corresponding probability distribution characteristics from historical computing resource information, and determining the estimated number of computing nodes based on these probability distribution characteristics, the effective experience of tenants in past resource adjustments can be fully utilized. Furthermore, historical final values ​​are optimal resource configuration results verified in practice, and their probability distribution characteristics can reflect the patterns of resource demand under different load scenarios (e.g., frequently occurring final values ​​often correspond to more reasonable configurations). The estimation method based on probability distribution characteristics avoids the randomness of single historical data, captures the inherent correlation between resource demand and load characteristics through statistical regularities, and makes the determination of the estimated number of computing nodes more closely aligned with the tenant's actual business model, significantly improving the accuracy of the initial estimation, reducing the frequency of subsequent adjustments, and thus improving the efficiency and stability of cloud platform computing resource scheduling.

[0046] S23. When no historical computing resource information corresponding to the computing resource information is found in the experience database, the estimated number of computing nodes of the tenant is calculated using the computing resource information and the preset load threshold.

[0047] For example, if no historical computing resource information corresponding to the aforementioned computing resource information is found in the experience database, it indicates that the tenant's current load status (the combination of initial load utilization and initial number of computing nodes) is appearing for the first time, or there are insufficient similar adjustment cases in the historical records for direct reference. This could be due to the tenant having just joined the platform, significant changes in load characteristics, or the matching degree in historical data not reaching the preset threshold. In this case, estimation cannot be based on past experience. Instead, the estimated number of computing nodes for the tenant must be calculated using the aforementioned computing resource information and the preset load threshold. This ensures that even in the absence of historical data support, a reasonable estimated value can still be generated based on the platform's general rules, guaranteeing the continuity and adaptability of the resource adjustment process.

[0048] In this embodiment of the invention, when corresponding historical computing power resource information exists, the effective configurations previously verified by tenants are reused through an experience database to fully match their personalized load characteristics and improve estimation accuracy. When no matching historical data exists, standardized calculations based on preset load thresholds ensure that reasonable estimated values ​​can still be generated under new scenarios or sudden load changes, avoiding adjustment stagnation due to lack of experience. This approach fully utilizes historical data to mine business patterns and ensures process robustness through general rules, ultimately achieving efficient generation of estimated computing power nodes for different tenants and different load scenarios. This lays a reliable foundation for subsequent resource adjustments and improves the overall resource scheduling and adaptation capabilities of the cloud platform.

[0049] Further, in step S23, the load threshold includes an expansion threshold and a reduction threshold. Calculating the estimated number of computing power nodes for the tenant using the computing power resource information and the preset load threshold includes: when the initial load utilization rate is greater than the expansion threshold, calculating the estimated number of computing power nodes for the tenant based on the initial load utilization rate, the initial number of computing power nodes, and the expansion threshold; when the initial load utilization rate is less than the reduction threshold, calculating the estimated number of computing power nodes for the tenant based on the initial load utilization rate, the initial number of computing power nodes, and the reduction threshold.

[0050] For example, the expansion threshold H is used to determine whether the tenant's current load is too high and requires additional resources. When the initial load utilization exceeds this threshold, it indicates that the existing resources are insufficient to support the current workload, and an expansion assessment needs to be initiated. For example, H can be set to 80%, meaning that expansion should be considered when the CPU utilization exceeds 80%. This can be set based on empirical values. The reduction threshold L is used to determine whether the tenant's current load is too low and requires reduced resources. When the initial load utilization is below this threshold, it indicates that the existing resources are idle and wasted, and a reduction assessment needs to be initiated. For example, L can be set to 30%, meaning that reduction should be considered when the CPU utilization is below 30%. This can also be set based on empirical values. The initial load utilization rate n is compared with the expansion threshold H and the reduction threshold L because the load utilization rate directly reflects the matching degree between resources and job requirements. Only when n > H (overload) or n < L (underload) is it necessary to adjust the amount of resources (expansion or reduction) to optimize the matching efficiency. If the initial load utilization rate n% satisfies: L ≤ n ≤ H, it indicates that the current amount of resources and load requirements are in a reasonable balance, and no resource adjustment is required. At this time, the process ends and subsequent steps S3~S5 are not executed to avoid unnecessary resource fluctuations affecting business stability.

[0051] In this embodiment of the invention, by clearly distinguishing between expansion and contraction scenarios and matching corresponding calculation logic, it is ensured that the direction of resource adjustment is perfectly aligned with load requirements. When the load is too high, the number of nodes to be added is calculated based on the expansion threshold to avoid impacting business operations due to insufficient resources; when the load is too low, the number of nodes to be reduced is calculated based on the contraction threshold to prevent resource idleness and waste. In addition, the introduction of load thresholds sets clear trigger boundaries for resource adjustments, avoiding frequent adjustments caused by small fluctuations in load, ensuring the stability and efficiency of cloud platform resource scheduling, and achieving the resource management goal of on-demand allocation and precise control.

[0052] Furthermore, specific examples of expansion and reduction scenarios are provided in this embodiment of the invention, including the following cases 1) and 2).

[0053] Case 1): When the initial load utilization rate is greater than the expansion threshold, expansion is required. In this case, calculating the estimated number of computing nodes for the tenant based on the initial load utilization rate, the initial number of computing nodes, and the expansion threshold includes: calculating the product of the initial load utilization rate and the initial number of computing nodes as a computing load reference value; calculating a first ratio of the computing load reference value to the expansion threshold; and rounding the first ratio up to obtain the estimated number of computing nodes for the tenant.

[0054] For example, if n > H, then the controller needs to expand the computing resources of the tenant, and the total estimated number of computing nodes after expansion is... At this time, the number of force nodes is estimated. The calculation process satisfies the following formula: (1); in, This represents the floor function; This is a reference value for computing load. This is the first ratio. It should be noted that... The maximum value cannot exceed the preset node number threshold xmax. If it does, then the final value will be... =xmax.

[0055] For example, assume the initial load utilization is n=85%, the initial number of computing nodes is P=10, and the expansion threshold is H=80%. In this case, n(85%)>H(80%), satisfying the expansion condition. Substituting into the above formula (1), the calculation yields... The final estimated number of computing nodes is 11, which means that one more computing node needs to be added.

[0056] In this embodiment of the invention, the product of the initial load utilization rate and the initial number of computing power nodes is used as the computing power load reference value. This accurately reflects the tenant's current total load, rather than the load percentage of a single node, providing an objective data foundation for subsequent resource calculations and avoiding estimation errors caused by focusing only on the load of a single node while ignoring the total demand. By using the ratio of the load reference value to the expansion threshold, the theoretical number of nodes required to reduce the load to a reasonable threshold can be directly derived. Combined with rounding up, this ensures that the final estimated value can fully support the current total load, avoiding situations where the adjusted load still exceeds the threshold due to insufficient nodes. This satisfies the tenant's real-time resource requirements while preventing resource waste caused by excessive expansion. The method can quickly generate an appropriate estimated number of computing power nodes without relying on complex models, balancing the accuracy and efficiency of the calculation and providing a reliable guarantee for the rationality of expansion operations.

[0057] 2) When the initial load utilization rate is less than the scaling-down threshold, scaling down is required. In this case, the step of calculating the estimated number of computing nodes for the tenant based on the initial load utilization rate, the initial number of computing nodes, and the scaling-down threshold includes: calculating the product of the initial load utilization rate and the initial number of computing nodes as a computing load reference value; calculating a second ratio between the computing load reference value and the scaling-down threshold; and rounding down the second ratio to obtain the estimated number of computing nodes for the tenant.

[0058] For example, if n < L, then the controller needs to scale down the tenant's computing resources, and the total estimated number of computing nodes after scaling down is... At this time, the number of force nodes is estimated. The calculation process satisfies the following formula: ; in, represents the floor function; is the computing power load reference value; is the second ratio. It should be noted that The value of cannot be less than 1 at least. If it is less than, then finally = 1.

[0059] For example, assume that the initial load utilization rate is n = 25%, the initial number of computing power nodes is P = 10, the scaling threshold is L = 30%, and the preset minimum number of nodes is 1. At this time, n (25%) < L (30%), which meets the scaling condition. After substituting into the above formula (2), it is calculated that , and finally the pre-estimated number of computing power nodes is 8, meaning that 2 computing power nodes need to be reduced.

[0060] In the embodiment of the present invention, through the ratio of the load reference value to the scaling threshold, the theoretical number of nodes required to raise the load back to a reasonable threshold is directly deduced, and then through the floor operation, it is ensured that the final pre-estimated value can not only raise the load above the scaling threshold, avoiding the load being too low and resources still being idle due to excessive reduction of the number of nodes, but also can release redundant resources to the greatest extent, reduce unnecessary node occupancy, maintain the balance between the load and resources while avoiding resource waste, and effectively improve the overall utilization efficiency of the cloud platform resources.

[0061] Furthermore, referring to Figure 4 , Figure 4 is the flowchart for judging whether to send the pre-estimated number of computing power nodes to the cloud platform provided by the embodiment of the present invention; before executing step S3, the method further includes steps S201 to S203.

[0062] S201. When the current computing power resource adjustment operation is an expansion operation, count the total number of currently allocated computing power nodes of all tenants, and summarize the reference number of nodes to be expanded this time.

[0063] Exemplarily, count the total number of currently allocated computing power nodes of all tenants, that is, the total number of nodes currently occupied by each tenant on the cloud platform. For example, currently all tenants use a total of 500 nodes. And summarize the reference number of nodes to be expanded this time. This reference number of nodes is the difference between the pre-estimated number of computing power nodes of the current tenant and the initial number of computing power nodes. For example, if the initial number of nodes of the current tenant is 10 and the pre-estimated number of computing power nodes is 15, then the reference number of nodes is 5, so as to clarify the additional resource amount required for this expansion. Since multiple tenants may need to expand simultaneously, the reference number of nodes of all tenants needs to be summarized. Assume the reference number of nodes is 30.

[0064] S202. Add the total number of computing power nodes to the reference number of nodes to obtain the expected total occupancy.

[0065] For example, the sum of the above computing power nodes, 500, is added to the reference number of nodes, 30, to obtain the expected total occupancy of 530. This value represents the total resource occupancy of the cloud platform after expansion and is used to determine whether it exceeds the platform's carrying capacity.

[0066] S203. If the estimated total occupancy is greater than or equal to the total resources of the cloud platform, stop this computing power resource adjustment operation; if the estimated total occupancy is less than the total resources of the cloud platform, proceed to step S3.

[0067] For example, if the total number of resources on the cloud platform is 510, and the estimated total usage is 530≥510, it indicates that the resources are insufficient. The current expansion operation of some tenants (which can be selected as needed, such as prioritizing the tenants with the fewest expansion nodes) needs to be stopped to avoid exhausting the platform resources. If the total number of resources is 600, and the estimated total usage is 530<600, it indicates that the resources are sufficient, and subsequent adjustment steps can be executed.

[0068] In this embodiment of the invention, by statistically analyzing the current total occupancy and the planned expansion of resources, the resource load after expansion is accurately predicted. From a global perspective, this avoids the risk of platform resource exhaustion due to expansion by a single tenant, ensuring the overall stability of the cloud platform. Furthermore, expansion is only performed when resources are sufficient, ensuring the feasibility of resource allocation, preventing resource contention among tenants, and reserving reasonable resource space for other tenants. This balances the needs of individual tenants with the overall resources of the platform, improving the fairness and reliability of resource scheduling in multi-tenant scenarios.

[0069] Furthermore, this embodiment of the invention also introduces a fence mechanism to protect the experience database. Because cloud database platforms have numerous tenants with varying levels of technical skills, for example, some users write SQL statements with poor execution plans, leading to a sudden increase in CPU resource load during execution. If the metrics monitored by the controller are not limited, individual tenants may experience abnormally high resource consumption. In the worst case, multiple tenants executing such SQL statements simultaneously can exhaust cloud platform resources, resulting in less skilled tenants receiving more resources than more skilled tenants—this is unfair. This relates to the robustness of the computing power scaling solution model: examining the model's stability in the face of unexpected inputs, including its ability to handle out-of-distribution (OOD) data and its resilience against adversarial attacks. Improvements can be made through data cleaning and preprocessing: before training and testing the model, the data is cleaned and preprocessed to remove outliers and handle missing data, reducing the impact of imperfect data.

[0070] To avoid the aforementioned problems, existing technologies typically set an upper limit on the number of resources a tenant can scale up. This invention provides an alternative approach using ensemble learning. Ensemble learning combines the outputs of multiple models, using ensemble learning methods (such as voting or averaging) to obtain more robust and accurate results, thereby mitigating the problem of overconfidence. Expanding the empirical vector data e, increasing the number of tasks c, and letting e = (u, t, c, p, n, x), the computing power scaling model incorporates the task quantity c parameter into resource estimation. See also... Figure 5 , Figure 5 This is another application diagram of the generative adversarial network model provided in this embodiment of the invention, where V(x) is a voting function Vote, indicating that x evaluated by P(z) needs to be voted on by all tenants. By replacing u in the index z with the identifier ui of other tenants, and then sampling, the initial estimated value xi of other tenants can be obtained, and all tenants ui vote on x for u.

[0071] At this point, the load metric z of tenant u obtained by the controller and the data sampled from the generator's empirical data satisfy the following: z = (u, t, p, n, 0); x=G(P(z)); The load metrics zi of other tenant UIs and the data sampled from the generator's empirical data satisfy the following: zi = (ui, t, p, n, 0); xi=G(P(zi)).

[0072] It's important to note that if a tenant's historical CPU load spikes and resource demands are abnormally high due to poor SQL (e.g., a poor execution plan), for example, a normal tenant might require 2 nodes for the same task, but this tenant historically requires 10 nodes due to SQL issues, GAN will treat this abnormal behavior as normal. The generator will then output x=10, fully accepting this abnormal demand. GAN's goal is to match resource allocation to the tenant's actual load, not to determine the reasonableness of the load. It cannot distinguish between a tenant's reasonable business load and abnormal load caused by their own technical limitations. It will judge abnormal resource demands corresponding to abnormal loads as reasonable, outputting x that conforms to individual abnormal habits but is globally unreasonable. If multiple tenants simultaneously obtain high x values ​​through GAN due to abnormal loads, GAN's individual load matching logic cannot predict this. Each tenant's x may match their own load individually, but when combined, it will directly exhaust the cloud platform's total resources, causing all tenants' businesses to crash. Therefore, a fence mechanism called Vote is needed. The essence of Vote is to perform a second global review of the resource demands output by GAN based on the collective experience of all tenants.

[0073] See Figure 6 , Figure 6This is another flowchart provided by an embodiment of the present invention for determining whether to send the estimated number of power nodes to the cloud platform. Before executing step S3, the method further includes steps S211 to S212.

[0074] S211. When the current computing power resource adjustment operation is an expansion operation, if the tenant does not belong to the preset high-priority whitelist, the tenant voting mechanism is activated to verify the estimated number of computing power nodes.

[0075] For example, the high-priority whitelist refers to a pre-defined set of tenants with special resource scheduling permissions set by the cloud platform. This typically includes tenants with extremely high business continuity requirements, large resource usage, or those with exclusive service agreements with the platform. The expansion needs of these tenants are often time-sensitive and require high stability; therefore, they do not require verification through a tenant voting mechanism and can directly enter the subsequent resource adjustment process, ensuring their business is not affected by additional process delays. Ordinary tenants not on this whitelist, however, must initiate a voting mechanism for their expansion requests. Other tenants judge the reasonableness of the request based on their own experience, preventing excessive platform resource consumption due to abnormal expansion requests from individual ordinary tenants. This ensures the service quality of high-priority tenants while maintaining fairness in resource allocation among ordinary tenants.

[0076] It should be noted that the voting mechanism primarily targets expansion operations, not shrinking operations. The core reason is that shrinking operations do not pose a risk of excessive resource consumption and do not require constraint through tenant group voting. Abnormal expansion (such as a tenant requesting a large amount of resources due to a sudden increase in CPU caused by inefficient SQL statements) consumes the cloud platform's limited resources, potentially exhausting the total resource M and crowding out other tenants' reasonable resources, thus undermining allocation fairness. This is precisely the core problem that the voting mechanism in this invention aims to solve. Shrinking, on the other hand, reduces the number of nodes allocated to a tenant, essentially releasing idle resources. It does not consume additional resources but rather increases the cloud platform's idle resources, benefiting other tenants. There is no issue of excessive resource consumption or crowding out others' resources, therefore, it does not require constraint through a voting mechanism.

[0077] S212. If the verification fails, stop the current computing power resource adjustment operation; if the verification passes, execute the step of sending the estimated number of computing power nodes to the cloud platform.

[0078] For example, if the verification fails, it means that the current tenant's estimated number of power nodes has not been recognized by the experience of most target tenants, and there may be a risk of excessive resource requests, which could easily lead to an imbalance or waste of platform resources. In this case, the current computing power resource adjustment operation should be stopped to avoid unreasonable expansion occupying too many platform resources. At the same time, the current tenant should be informed of the reason for the verification failure, such as the requested resources exceeding the reasonable range for similar scenarios of similar tenants, and be guided to optimize job configuration or adjust resource requirements. If the verification passes, it means that the current tenant's estimated number of power nodes is in line with the reasonable resource range formed by the experience of most target tenants, which is consistent with the fairness principle of multi-tenant resource sharing on the platform, and is unlikely to have a negative impact on the resource use of other tenants. In this case, step S3 should be executed to promote the orderly development of the subsequent resource adjustment process, ensuring that the expansion operation meets the needs of the current tenant while taking into account the overall resource balance of the platform.

[0079] In this embodiment of the invention, the reasonableness of the current tenant's estimated number of capacity nodes can be effectively verified through voting verification by other tenants based on their own experience. This avoids the platform resources from being tilted towards a single entity due to excessive resource requests from individual tenants, ensuring a more balanced resource allocation among multiple tenants, and especially protecting the resource usage rights of ordinary tenants. In addition, if the verification fails, the operation is stopped, which can intercept abnormal expansion requests that may cause resource waste or imbalance in advance, reducing the ineffective consumption of cloud platform resource scheduling.

[0080] Furthermore, the tenant voting mechanism includes: sampling and generating reference computing power node numbers for each target tenant from the experience database corresponding to the target tenant based on the target tenant's tenant identifier; wherein, the target tenant is the remaining tenants excluding the current tenant; if any reference computing power node number is greater than the estimated computing power node number, a vote of approval is cast, otherwise a vote of disapproval is cast; after all target tenants have completed their voting, the approval ratio of the approval votes to the total number of votes is calculated; if the approval ratio is greater than or equal to a preset ratio threshold, the verification is deemed successful; if the approval ratio is less than the ratio threshold, the verification is deemed unsuccessful.

[0081] For example, tenant u always trusts its own experience and votes in favor of x. Given the same number of jobs, other tenants ui, based on their own experience, do not allow u to use excessive resources. If other tenants' sample data is empty, indicating a lack of experience, they do not participate in the vote; after successful sampling, if xi > x, they vote in favor; otherwise, they vote against. The total number of votes and the number of votes in favor are counted. If the proportion of votes in favor is not less than a predetermined proportion, x passes; otherwise, it fails. This invention sets the predetermined proportion at 50% (only as an example, and can be adjusted as needed). The higher the predetermined proportion, the lower the tolerance of the tenant group for abnormally high resource consumption. If the voting result fails, the data is judged as abnormal data, and the expansion operation is directly canceled. The reason for the expansion cancellation will be sent to the tenants to encourage improvement of user skills.

[0082] For example, suppose the current tenant is tenant A (planned expansion, estimated capacity nodes x=15), and the target tenants are tenants B, C, and D (the remaining tenants excluding A). The preset proportion threshold is 50%, and the specific voting process is as follows: The first step is to generate a reference number of computing power nodes. Based on the identifiers of tenants B, C, and D, samples are taken from their respective experience databases. In tenant B's experience database, a reference number of computing power nodes is matched to those in scenarios similar to A's current load. =18; The number of reference computing nodes for similar scenarios is sampled from tenant C's experience database. =14; Tenant D's experience database has no matching historical data (sampling is empty), therefore it does not participate in the voting; The second step is target tenant voting, specifically tenant A's vote. =18>x=15, vote in favor; Tenant B's =14 < x = 15, vote against; tenant D does not participate due to lack of experience data, total votes are 2 (only B and C participated); The third step is to calculate the approval ratio and verification result. The approval vote is 1, and the approval ratio = 1 ÷ 2 = 50%, which is equal to the preset 50% threshold. Therefore, the verification is deemed successful, and tenant A's estimated capacity node number of 15 can proceed to the subsequent expansion process.

[0083] In this embodiment of the invention, a voting mechanism is used to validate the resource requirements of a single tenant within a framework of collective experience. This leverages the resource allocation patterns of similar tenants in similar scenarios to determine whether the current estimated value exceeds a reasonable range, preventing resource allocation imbalances caused by individual demands deviating from collective consensus. Through the collaborative constraints of collective experience, unreasonable expansion is effectively curbed, maintaining the fairness and stability of resource allocation in multi-tenant scenarios and providing a mechanism guarantee for the efficient sharing of cloud platform resources. Simultaneously, by comparing the approval ratio with a preset threshold, the degree of collective acceptance of the current resource requirements is quantified, preventing abnormally high demands from a few tenants from consuming excessive resources. Furthermore, tenants without experience data are filtered out from voting, ensuring that the verification basis comes from entities with practical operational experience, thus improving the reliability of the verification.

[0084] S3. Send the estimated number of computing nodes to the cloud platform and obtain the reference load utilization rate after the cloud platform adjusts the computing resources of the tenant based on the estimated number of computing nodes.

[0085] For example, assuming the preliminary process determines that the current tenant's estimated computing power nodes are 12 (initially 10 nodes, requiring an expansion of 2), a resource adjustment instruction containing the estimated number of 12 nodes, tenant identifier, and serial number is sent to the cloud platform. After receiving the instruction, the cloud platform locates the corresponding tenant's resource pool based on the tenant identifier and automatically executes the computing power node expansion operation, allocating 2 nodes from the platform's idle resources to the tenant, increasing the tenant's current available computing power nodes from 10 to 12. After the expansion operation is completed and the system is running stably, the cloud platform's controller collects the tenant's adjusted real-time load data (such as average CPU utilization), i.e., the reference load utilization. Finally, the cloud platform sends back the reference load utilization and the corresponding serial number, request time, and other information through a feedback interface, providing actual data for the discriminator to evaluate the accuracy of the estimated number of computing power nodes in subsequent steps.

[0086] S4. Evaluate the reference load utilization rate to obtain the evaluation result.

[0087] Further, step S4 specifically includes: when the current computing power resource adjustment operation is an expansion operation, if the reference load utilization rate is less than or equal to the expansion threshold, the evaluation result is that the estimated number of computing power nodes is accurate; if the reference load utilization rate is greater than the expansion threshold, the evaluation result is that the estimated number of computing power nodes is inaccurate. When the current computing power resource adjustment operation is a reduction operation, if the reference load utilization rate is greater than or equal to the reduction threshold, the evaluation result is that the estimated number of computing power nodes is accurate; if the reference load utilization rate is less than the reduction threshold, the evaluation result is that the estimated number of computing power nodes is inaccurate.

[0088] For example, if this is a capacity expansion operation, and the expansion has already been completed based on the estimated number of capacity nodes, the discriminator judges the accuracy by comparing the reference load utilization rate with the expansion threshold H. If the reference load utilization rate is ≤ H (e.g., the actual load drops to 75% ≤ 80%), it means that the expanded resources can effectively control the load within a reasonable upper limit, the estimated value matches the actual demand, and the evaluation result is that the estimated number of capacity nodes is accurate; if the reference load utilization rate is > H (e.g., the actual load is still 83% > 80%), it means that the expanded resources still cannot meet the load demand (the estimated value is too low), the expansion target has not been achieved, and the evaluation result is that the estimated number of capacity nodes is inaccurate.

[0089] For example, if this is a scaling-down operation, and scaling down has already been completed based on the estimated number of capacity nodes, then the accuracy is judged by comparing the reference load utilization rate with L. If the reference load utilization rate is ≥ L (e.g., the actual load rises to 35% ≥ 30%), it indicates that the resources after scaling down have not caused an excessive reduction in load, there is no significant idle resources, the estimated value is reasonable, and the evaluation result is that the estimated number of capacity nodes is accurate. If the reference load utilization rate is < L (e.g., the actual load drops to 25% < 30%), it indicates that there are still a lot of idle resources after scaling down, the estimated value is too low, exceeding the necessity of scaling down, and the evaluation result is that the estimated number of capacity nodes is inaccurate.

[0090] In this embodiment of the invention, focusing on two core scenarios—expansion and contraction—a precise and target-oriented evaluation logic is formed using a preset load threshold as a clear judgment standard. By uniformly using the load threshold as the evaluation benchmark, deviations caused by inconsistent evaluation dimensions are avoided. This ensures consistency in evaluation rules across different tenants and adjustment scenarios, guaranteeing the effectiveness of computing resource adjustments. It avoids repeated adjustments due to inaccurate estimations and provides reliable decision support for the dynamic optimization of cloud platform resources, further improving the stability and efficiency of resource scheduling.

[0091] S5. When the evaluation result indicates that the estimated number of computing power nodes is inaccurate, the estimated number of computing power nodes is adjusted according to the evaluation result to obtain the target number of computing power nodes, and the target number of computing power nodes is sent to the cloud platform.

[0092] Furthermore, when the current computing power resource adjustment operation is an expansion operation, a preset expansion step size is added to the estimated number of computing power nodes to obtain the target number of computing power nodes. When the current computing power resource adjustment operation is a reduction operation, a preset callback step size is reduced to the estimated number of computing power nodes to obtain the target number of computing power nodes.

[0093] For example, if this is an expansion operation and the initial assessment indicates that the estimated number of computing nodes is inaccurate, such as if the reference load utilization is still greater than the expansion threshold H, then the expansion step size is increased based on the original estimated number of computing nodes to obtain the target number of computing nodes, thereby further increasing resource availability and reducing load. Assuming the preset expansion step size is 2, if the estimated number of computing nodes is 12, then the target number of computing nodes is 14. If this is a reduction operation and the assessment indicates that the estimated number of computing nodes is inaccurate, such as if the reference load utilization is still less than the reduction threshold L, then the callback step size is decreased based on the original estimated number of computing nodes to obtain the target number of computing nodes, thereby further releasing redundant resources and improving load utilization. Assuming the preset callback step size is 1, if the estimated number of computing nodes is 12, then the target number of computing nodes is 11.

[0094] In this embodiment of the invention, for scenarios where the estimated value is inaccurate, the target number of computing power nodes is generated by fine-tuning with a fixed step size. This avoids resource fluctuations caused by blindly making large adjustments, such as adding too many nodes at once during expansion, which would cause waste, or reducing nodes excessively during shrinkage, which would cause a sudden increase in load. This ensures the gradualness and stability of resource adjustments.

[0095] Compared to existing technologies, the computing power resource adjustment method disclosed in this invention dynamically acquires tenant load and node information, accurately determines the estimated number of computing power nodes by combining tenant identifiers, and performs evaluation and secondary optimization based on the actual load feedback after adjustment, ultimately outputting the appropriate target number of computing power nodes. This method can accurately adapt to the regular or irregular fluctuations of business operations, avoiding business congestion and reducing the deployment and maintenance costs of the cloud platform. Furthermore, the closed-loop mechanism of initial estimation and feedback adjustment ensures that computing power resource allocation closely matches the actual load of tenants, avoiding resource surplus or shortage and improving resource utilization efficiency. Evaluation and adjustment based on real-time load utilization can quickly respond to fluctuations in tenant workload, ensuring business operation stability. Combining historical characteristics or configuration rules associated with tenant identifiers enables differentiated resource allocation, balancing the needs of individual tenants with the overall resource balance of the cloud platform. Through fully automated evaluation and adjustment, manual intervention is reduced, improving the efficiency and intelligence of cloud platform computing power resource scheduling and supporting fair and efficient resource sharing in multi-tenant scenarios.

[0096] Furthermore, embodiments of the present invention also provide a computing power resource adjustment device, the computing power resource adjustment device comprising: A generator is used to obtain the tenant identifier and computing resource information of any tenant on the cloud platform; wherein, the computing resource information includes the initial load utilization rate and the initial number of computing nodes; based on the tenant identifier and the computing resource information, the generator determines the estimated number of computing nodes for the tenant; the generator sends the estimated number of computing nodes to the cloud platform, and obtains the reference load utilization rate after the cloud platform adjusts the computing resources of the tenant based on the estimated number of computing nodes; A discriminator is used to evaluate the reference load utilization, obtain an evaluation result, and when the evaluation result indicates that the estimated number of computing power nodes is inaccurate, adjust the estimated number of computing power nodes according to the evaluation result to obtain the target number of computing power nodes, and send the target number of computing power nodes to the cloud platform.

[0097] It should be noted that the computing resource adjustment device, acting as a DSCS (Dynamic Scaling Cloud Services), is responsible for dynamically scaling resource computation and providing interface services to the outside world. Its scaling engine logic includes the generative adversarial network model described in this case, as well as the Vote mechanism. Of course, the computing engine can also be implemented using other schemes and algorithms, but it must provide standard services to external platforms by exposing an interface. The working process of the generator and the discriminator can be referred to the working process of the generator and discriminator in the computing resource adjustment method described in the above embodiments, and will not be repeated here.

[0098] Furthermore, the computing power resource adjustment device also includes an interface module, which provides a data interaction channel between the computing power resource adjustment device and the cloud platform.

[0099] For example, see Figure 7 , Figure 7 This is a structural block diagram of the dynamically scalable DSCS and cloud platform provided in this embodiment of the invention. The cloud platform side includes computing resources and a resource controller. The computing resources are the set of computing nodes actually used by the tenant. The resource controller is responsible for receiving resource adjustment instructions from the DSCS and performing expansion or contraction operations on the computing resources, while also feeding back the adjusted load data. The DSCS side includes an interface service and a scaling engine. The interface service serves as a data interaction channel, receiving load data from the cloud platform (such as initial load utilization and reference load utilization) and sending resource adjustment instructions, such as the estimated number of computing nodes, to the cloud platform. The scaling engine embeds core logic such as a generative adversarial network model (including a generator and a discriminator) and a Vote mechanism, and is responsible for performing the calculation, verification, and secondary adjustment decisions of the estimated number of computing nodes. Finally, the decision results are transmitted to the cloud platform through the interface service to realize the dynamic scaling of computing resources.

[0100] In this embodiment of the invention, the architecture design clarifies the functional boundaries and collaboration process of both parties. The cloud platform focuses on the physical management and data collection of computing resources, while DSCS focuses on intelligent decision-making and service encapsulation for resource adjustment. Through standardized interaction of interface modules, the efficiency and flexibility of dynamic scaling of computing resources are ensured, which not only meets the needs of precise resource scheduling in multi-tenant scenarios, but also reserves expansion space for subsequent algorithm iterations (such as replacing computing engines).

[0101] Furthermore, the interface module includes: A first interface is used to transmit a first request message and a first response message; wherein the first request message carries the tenant identifier and the computing power resource information, and the first response message carries the estimated number of computing power nodes; The second interface is used to transmit a second request message and a second response message; wherein the second request message carries the reference load utilization rate, and the second response message carries the target number of computing power nodes; Both the first and second interfaces use HTTPS (Secure Hypertext Transfer Protocol) and employ a signature mechanism for data verification.

[0102] For example, see Figure 8 , Figure 8 This is a schematic diagram of information interaction between the dynamically scalable DSCS and the cloud platform provided in an embodiment of the present invention. The information interaction process between the dynamically scalable DSCS and the cloud platform is implemented through a first interface and a second interface, and the specific process is as follows: 1) First interface interaction (GAN generation + Vote stage): The cloud platform controller collects load data (such as load metric z, including initial load utilization, initial number of computing nodes, and other computing resource information) from computing resources, and encapsulates it into a first request message carrying the tenant identifier. This message is then sent to the DSCS via HTTPS and verified through a signature mechanism. Upon receiving the message, the DSCS uses the GAN model within its scaling engine to generate a preliminary estimate of the number of computing nodes. If the scaling operation is for a non-high-priority whitelisted tenant, a Vote mechanism will be initiated for verification. After successful verification, the preliminary estimate of the number of computing nodes is encapsulated into a first response message, which is also sent back to the controller via HTTPS and a signature mechanism. Based on the preliminary estimate of the number of computing nodes in the first response message, the controller performs resource scaling operations to expand or shrink the computing resources.

[0103] 2) Second interface interaction (GAN adversarial phase): After the resource adjustment is completed, the controller collects the adjusted reference load utilization rate from the computing power resources again, encapsulates it into a second request message, and sends it to the DSCS via HTTPS protocol and signature mechanism. After receiving the message, the DSCS uses the GAN model (discriminator) in the scaling engine to evaluate the accuracy of the estimated number of computing power nodes based on the reference load utilization rate. If it is inaccurate, it calculates the target number of computing power nodes (such as increasing the step size for expansion and decreasing the step size for shrinkage), and encapsulates the target number of computing power nodes into a second response message, which is sent back to the controller via HTTPS protocol and signature mechanism to provide a basis for possible secondary resource adjustments in the future.

[0104] It should be noted that the message formats of the first request message, the first response message, the second request message, and the second response message can be set as needed, and this invention does not impose specific limitations on them. Regarding the dimension of the load metric z, the interface generally requires a fixed dimension. This is because most models and deep learning frameworks require input data to have a fixed dimension, and the advantage of a fixed dimension is improved computational efficiency. Vector data with variable dimensions is also feasible in some scenarios; some advanced model architectures with adaptive capabilities can handle inputs of different dimensions. However, using vector data with variable dimensions increases the complexity and difficulty of model training, requiring more complex preprocessing and post-processing steps, as well as more refined model design and tuning strategies. DSCS issues signing public and private keys to the platform. The peer's public key is used to sign messages in the interface, and the peer possesses the private key, thus enabling designing.

[0105] In this embodiment of the invention, this dual-interface-based interactive design ensures the security of data transmission through the HTTPS protocol, ensures data integrity through the signature mechanism, and decouples the resource estimation and verification process into two stages. This not only meets the efficiency requirements of real-time resource adjustment, but also achieves closed-loop optimization of the estimation results through GAN adversarial mechanisms, ultimately improving the intelligence and reliability of the dynamic scaling of cloud platform computing resources.

[0106] See Figure 9 , Figure 9 This is a structural block diagram of a computing resource adjustment device 200 provided in an embodiment of the present invention. The computing resource adjustment device 200 includes a processor 11, a memory 12, and a computer program stored in the memory 12 and executable on the processor 11. When the processor 11 executes the computer program, it implements the steps in the above-described embodiments of the computing resource adjustment methods.

[0107] For example, the computer program can be divided into one or more modules / units, which are stored in the memory 12 and executed by the processor 11 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the computing resource adjustment device 200.

[0108] The computing resource adjustment device 200 may include, but is not limited to, a processor 11 and a memory 12. Those skilled in the art will understand that the schematic diagram is merely an example of the computing resource adjustment device 200 and does not constitute a limitation on the computing resource adjustment device 200. It may include more or fewer components than shown in the diagram, or combine certain components, or different components. For example, the computing resource adjustment device 200 may also include input / output devices, network access devices, buses, etc.

[0109] The processor 11 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 11 is the control center of the computing power resource adjustment device 200, connecting all parts of the computing power resource adjustment device 200 via various interfaces and lines.

[0110] The memory 12 can be used to store the computer programs and / or modules. The processor 11 implements various functions of the computing power resource adjustment device 200 by running or executing the computer programs and / or modules stored in the memory 12 and calling the data stored in the memory 12. The memory 12 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 12 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0111] Wherein, if the modules / units integrated by the computing power resource adjustment device 200 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor 11, it can implement the steps of the above method embodiments. Wherein, the computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0112] Furthermore, the present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the computing resource adjustment method as described in any of the above embodiments.

[0113] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for adjusting computing resources, characterized in that, include: Obtain the tenant identifier and computing resource information of any tenant on the cloud platform; wherein, the computing resource information includes the initial load utilization rate and the initial number of computing nodes; Based on the tenant identifier and the computing power resource information, determine the estimated number of computing power nodes for the tenant; The estimated number of computing nodes is sent to the cloud platform, and the reference load utilization rate after the cloud platform adjusts the computing resources of the tenant based on the estimated number of computing nodes is obtained. The reference load utilization rate is evaluated to obtain the evaluation results; When the evaluation result indicates that the estimated number of computing power nodes is inaccurate, the estimated number of computing power nodes is adjusted according to the evaluation result to obtain the target number of computing power nodes, and the target number of computing power nodes is sent to the cloud platform.

2. The computing resource adjustment method as described in claim 1, characterized in that, The step of determining the estimated number of computing power nodes for a tenant based on the tenant identifier and the computing power resource information includes: Query the experience database corresponding to the tenant identifier; When the historical computing resource information corresponding to the computing resource information is found in the experience database, the estimated number of computing nodes for the tenant is determined based on the historical computing resource information. When no historical computing resource information corresponding to the computing resource information is found in the experience database, the estimated number of computing nodes for the tenant is calculated using the computing resource information and a preset load threshold.

3. The computing resource adjustment method as described in claim 2, characterized in that, The step of determining the estimated number of computing power nodes for the tenant based on the historical computing power resource information includes: Extract the corresponding historical final values ​​and corresponding probability distribution features from the historical computing power resource information; Based on the probability distribution characteristics of the historical final values, the estimated number of power nodes for the tenant is determined.

4. The computing resource adjustment method as described in claim 2, characterized in that, The load threshold includes a scaling-up threshold and a scaling-down threshold. The step of calculating the tenant's estimated number of computing power nodes using the computing power resource information and the preset load threshold includes: When the initial load utilization rate is greater than the expansion threshold, the estimated number of computing nodes for the tenant is calculated based on the initial load utilization rate, the initial number of computing nodes, and the expansion threshold. When the initial load utilization is less than the scaling-down threshold, the estimated number of computing nodes for the tenant is calculated based on the initial load utilization, the initial number of computing nodes, and the scaling-down threshold.

5. The computing resource adjustment method as described in claim 4, characterized in that, The step of calculating the estimated number of computing nodes for the tenant based on the initial load utilization, the initial number of computing nodes, and the expansion threshold includes: The product of the initial load utilization rate and the initial number of computing nodes is calculated as the computing load reference value; Calculate the first ratio between the computing power load reference value and the expansion threshold; The first ratio is rounded up to obtain the estimated number of power nodes for the tenant.

6. The computing resource adjustment method as described in claim 4, characterized in that, The step of calculating the estimated number of computing nodes for the tenant based on the initial load utilization, the initial number of computing nodes, and the scaling-down threshold includes: The product of the initial load utilization rate and the initial number of computing nodes is calculated as the computing load reference value; Calculate a second ratio between the computing power load reference value and the scaling-down threshold; The second ratio is rounded down to obtain the estimated number of power nodes for the tenant.

7. The computing resource adjustment method as described in claim 1, characterized in that, Before sending the estimated number of power nodes to the cloud platform, the method further includes: When the current computing power resource adjustment operation is an expansion operation, the total number of computing power nodes currently allocated to all tenants is counted, and the reference number of nodes to be expanded in this operation is summarized. Add the sum of the number of computing power nodes to the number of reference nodes to obtain the estimated total occupancy; If the estimated total occupancy is greater than or equal to the total resources of the cloud platform, the current computing power resource adjustment operation is stopped; if the estimated total occupancy is less than the total resources of the cloud platform, the step of sending the estimated number of computing power nodes to the cloud platform is executed.

8. The computing resource adjustment method as described in claim 1, characterized in that, The evaluation of the reference load utilization rate to obtain the evaluation result includes: When the current computing power resource adjustment operation is an expansion operation, if the reference load utilization rate is less than or equal to the expansion threshold, the evaluation result is that the estimated number of computing power nodes is accurate; if the reference load utilization rate is greater than the expansion threshold, the evaluation result is that the estimated number of computing power nodes is inaccurate.

9. The computing resource adjustment method as described in claim 8, characterized in that, The step of adjusting the estimated number of computing power nodes based on the evaluation results to obtain the target number of computing power nodes includes: The target number of computing power nodes is obtained by adding a preset expansion step size to the estimated number of computing power nodes.

10. The computing resource adjustment method as described in claim 1, characterized in that, The evaluation of the reference load utilization rate to obtain the evaluation result includes: When the current computing power resource adjustment operation is a scaling-down operation, if the reference load utilization rate is greater than or equal to the scaling-down threshold, the evaluation result is that the estimated number of computing power nodes is accurate; if the reference load utilization rate is less than the scaling-down threshold, the evaluation result is that the estimated number of computing power nodes is inaccurate.

11. The computing resource adjustment method as described in claim 10, characterized in that, The step of adjusting the estimated number of computing power nodes based on the evaluation results to obtain the target number of computing power nodes includes: The target number of computing power nodes is obtained by reducing the preset callback step size based on the estimated number of computing power nodes.

12. The computing resource adjustment method as described in claim 1, characterized in that, Before sending the estimated number of power nodes to the cloud platform, the method further includes: When the current computing power resource adjustment operation is an expansion operation, if the tenant does not belong to the preset high-priority whitelist, the tenant voting mechanism is activated to verify the preset number of computing power nodes. If the verification fails, stop the current computing power resource adjustment operation; if the verification passes, execute the step of sending the estimated number of computing power nodes to the cloud platform.

13. The computing resource adjustment method as described in claim 12, characterized in that, The tenant voting mechanism includes: Based on the tenant identifier of the target tenant, the reference computing power node number for each is generated by sampling from the experience database corresponding to the target tenant; wherein, the target tenant is the tenant other than the current tenant; If any reference computing power node number is greater than the estimated computing power node number, then vote in favor; otherwise, vote against. After all target tenants have completed their voting, calculate the percentage of affirmative votes out of the total votes. If the approval rate is greater than or equal to a preset threshold, the verification is deemed successful; if the approval rate is less than the threshold, the verification is deemed unsuccessful.

14. A computing power resource adjustment device, characterized in that, include: A generator is used to obtain the tenant identifier and computing resource information of any tenant on the cloud platform; wherein, the computing resource information includes the initial load utilization rate and the initial number of computing nodes; based on the tenant identifier and the computing resource information, the generator determines the estimated number of computing nodes for the tenant; the generator sends the estimated number of computing nodes to the cloud platform, and obtains the reference load utilization rate after the cloud platform adjusts the computing resources of the tenant based on the estimated number of computing nodes; A discriminator is used to evaluate the reference load utilization, obtain an evaluation result, and when the evaluation result indicates that the estimated number of computing power nodes is inaccurate, adjust the estimated number of computing power nodes according to the evaluation result to obtain the target number of computing power nodes, and send the target number of computing power nodes to the cloud platform.

15. The computing power resource adjustment device as described in claim 14, characterized in that, The computing power resource adjustment device also includes an interface module, which provides a data interaction channel between the computing power resource adjustment device and the cloud platform.

16. The computing power resource adjustment device as described in claim 15, characterized in that, The interface module includes: A first interface is used to transmit a first request message and a first response message; wherein the first request message carries the tenant identifier and the computing power resource information, and the first response message carries the estimated number of computing power nodes; The second interface is used to transmit a second request message and a second response message; wherein the second request message carries the reference load utilization rate, and the second response message carries the target number of computing power nodes; The first interface and the second interface verify data through a signature mechanism.

17. A computing power resource adjustment device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the computing resource adjustment method as described in any one of claims 1 to 13.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the computing resource adjustment method as described in any one of claims 1 to 13.

19. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, implements the computing resource adjustment method as described in any one of claims 1 to 13.

Images