Heterogeneous computing transaction dynamic pricing and cross-domain resource scheduling method

Abstract

The application relates to the technical field of artificial intelligence, and discloses a heterogeneous computing power transaction dynamic pricing and cross-domain resource scheduling method, which comprises the following steps: collecting user historical task data, resource running state data and computing power transaction record data to construct a training data set; predicting future CPU demand, GPU demand and AI inference demand by using an AI prediction model; calculating the dynamic value weight of heterogeneous resources according to the prediction result, resource utilization rate and energy consumption cost, and mapping the three kinds of heterogeneous resources to a unified computing power value space; establishing a Stackelberg game model to generate a dynamic resource price; performing transaction matching with the maximum user utility and the minimum time delay as targets; triggering cross-domain scheduling when the regional load exceeds a threshold value, selecting an optimal path to migrate tasks according to a migration cost function; and collecting feedback data to update model parameters. The application can improve GPU utilization, reduce user time delay and resource idle rate, and improve platform revenue and social welfare.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for dynamic pricing of heterogeneous computing power transactions and cross-domain resource scheduling. Background Technology

[0002] With the rapid development of artificial intelligence technology, new intelligent applications such as large-scale model training, AI inference, autonomous driving, AIGC, and cloud-edge collaborative computing are placing increasingly higher demands on the scale, real-time performance, and elasticity of computing resources. Computing power has gradually become a crucial basic production factor in the digital economy era. Against this backdrop, edge computing systems, as key infrastructure supporting low-latency, high-bandwidth intelligent services, have been widely deployed and applied. However, traditional mobile edge computing systems typically manage computing resources using fixed pricing, static resource allocation, and reactive scheduling. Their resource allocation models mostly assume that user needs are known or can be obtained by the system in real time, and pricing mechanisms are primarily designed for single-type CPU computing resources. Resource scheduling is also often limited to a single edge node or a single region. Simultaneously, with the continuous advancement of computing network construction, the differences in computing load between different regions are becoming increasingly significant, placing higher demands on cross-regional resource collaborative scheduling capabilities. Therefore, how to achieve efficient configuration and intelligent scheduling of multiple types of computing resources in complex and dynamic computing network environments has become an important research topic urgently needing to be addressed in the current field of edge computing and computing network.

[0003] However, the aforementioned existing technical solutions still face numerous bottlenecks in practical applications, making it difficult to meet the development needs of future intelligent computing networks. Specifically: First, traditional resource allocation models lack the ability to anticipate future demand changes. Due to the randomness of user task arrivals, the suddenness of AI tasks, and the drastic fluctuations in GPU demand, coupled with the fact that user information is often not fully public, resource allocation decisions are delayed, easily leading to a series of problems such as GPU resource contention, computing power congestion, server idleness, and increased task latency. Second, existing pricing models are only applicable to single resource types, while future computing networks contain various heterogeneous resources such as CPU computing power, GPU computing power, NPU computing power, and AI inference computing power. These resources differ fundamentally in architecture, performance, and application scenarios, and currently lack a unified value measurement mechanism, making it difficult to achieve unified quantification and trading of heterogeneous computing power. Third, traditional dynamic pricing methods are "passive pricing" mechanisms, meaning that price increases are only implemented after demand increases, failing to guide resource flow in advance, especially under scenarios such as peak AI inference periods, surges in large model training loads, and regional computing power shortages, resulting in insufficient response speed. Fourth, most traditional edge computing models are designed for single regions or single servers and lack cross-regional computing power collaboration mechanisms. They are difficult to adapt to the actual needs of cross-regional computing power interconnection and collaborative scheduling, resulting in low overall resource utilization and the coexistence of local congestion and idleness.

[0004] In view of this, how to provide a computing power trading and scheduling scheme that can integrate future computing power demand perception, unified value assessment of heterogeneous computing power, proactive dynamic pricing mechanism and cross-regional resource collaborative scheduling, so as to achieve dynamic matching of computing power supply and demand, effective improvement of resource utilization and optimization of overall system benefits, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions, which solves the problems of low resource utilization, insufficient dynamic response capability, difficulty in unified trading of heterogeneous computing power, and lack of cross-regional collaborative scheduling in the prior art.

[0006] To achieve the above objectives, this invention provides a method for dynamic pricing and cross-domain resource scheduling in heterogeneous computing power transactions, comprising the following steps: Collect user historical task data, resource operation status data, and computing power transaction record data. Perform noise reduction, standardization, and missing data imputation preprocessing on the collected data, and slice it according to time windows to form a training dataset. The training dataset is input into the AI ​​prediction model to predict the heterogeneous computing power demand within a future time window, and the predicted values ​​of CPU demand, GPU demand, and AI inference demand are output. Based on the heterogeneous computing power demand forecast, combined with the current resource utilization and energy consumption costs, the dynamic value weights of three heterogeneous resources—CPU, GPU, and AI inference—are calculated, and the three heterogeneous resources are mapped to a unified computing power value space to obtain the unified computing power value assessment results. Based on the heterogeneous computing power demand forecast results and the unified computing power value assessment results, a Stackelberg game model with the computing power platform as the leader and users as followers is established. The equilibrium solution of the game model is solved to generate dynamic resource prices. Based on the dynamic resource price, the system receives resource requests submitted by users, which include task type, budget, latency constraints, and resource requirement parameters. With the optimization goal of maximizing user utility and minimizing task latency, the system performs transaction matching between user tasks and computing resources and outputs the transaction matching results. The resource load status of computing nodes in each region is periodically monitored. When the load in the first region exceeds a preset threshold and there are idle resources in the second region, cross-regional resource scheduling is triggered. A migration path is selected according to the migration cost function, and some tasks in the first region are migrated to the second region for execution. Collect system operation feedback data after transaction matching and cross-regional scheduling execution, and use the feedback data to retrain the AI ​​prediction model and update the model parameters.

[0007] Specifically, the training dataset is input into the AI ​​prediction model to predict heterogeneous computing power demand within a future time window, and the model outputs predicted CPU demand, predicted GPU demand, and predicted AI inference demand, including: Constructing the sliding time window input sequence: ; in, The time window length, for The status of resource demand at any given moment; The input sequence is fed into an LSTM network for prediction. The LSTM network extracts temporal features through a forget gate, an input gate, and a unit state update mechanism. Output prediction results: ; in, To predict CPU demand, To predict GPU demand, To predict AI inference needs.

[0008] Specifically, based on the heterogeneous computing power demand forecast results, combined with the current resource utilization rate and energy consumption cost, the dynamic value weights of three heterogeneous resources—CPU, GPU, and AI inference—are calculated. These three heterogeneous resources are then mapped to a unified computing power value space to obtain a unified computing power value assessment result, which includes: Building a unified computing power value model: ; in, CPU resource quantity For GPU resources, For AI inference resources, , , These are CPU dynamic weights, GPU dynamic weights, and AI inference dynamic weights, respectively. The dynamic weights are dynamically updated based on changes in market demand. The formula for updating CPU dynamic weights is: ; The formula for GPU dynamic weight update is: ; The formula for updating dynamic weights in AI inference is: ; in, For CPU demand growth rate, For GPU demand growth rate, For the growth rate of AI inference demand, For CPU utilization, For GPU utilization, For GPU power consumption costs; Calculate the overall dynamic value of the GPU based on the updated dynamic weights: ; in, For the dynamic value of GPU, For GPU dynamic weights, For GPU utilization, For GPU demand growth rate, This refers to the power consumption cost of the GPU.

[0009] Specifically, based on the heterogeneous computing power demand forecast results and the unified computing power value assessment results, a Stackelberg game model is established with the computing power platform as the leader and users as followers. The equilibrium solution of the game model is solved to generate dynamic resource prices, including: Constructing user utility functions: ; in, Apply for computing power for users, Price per unit For time delay sensitivity coefficient, For local execution latency, Delay for task unloading; Define the total latency for task unloading: ; in, To increase the amount of data uploaded, For wireless transmission rate, For task computation, To allocate computing power; Regarding the user utility function Taking the derivative and setting it to zero, we obtain the optimal resource request amount for the user: ; Constructing the platform profit function: ; in, Price per unit Unit operating cost To meet user resource needs; Based on the ratio of AI-predicted demand to current supply, dynamic resource prices are updated according to the following formula: ; in, To predict demand, For current supply, To adjust the step size.

[0010] Specifically, based on the dynamic resource price, the system receives resource requests submitted by users, including task type, budget, latency constraints, and resource requirement parameters. With the optimization objective of maximizing user utility and minimizing task latency, it performs transaction matching between user tasks and computing resources, and outputs the transaction matching results, including: Receive resource requests submitted by users, including task type, budget limit, maximum tolerable latency, CPU requirements, and GPU requirements; With the optimization objectives of maximizing user utility and minimizing task latency, the optimization function is defined as follows:

[0011] in, For the benefit of users, For time delay, This is the delay penalty coefficient; Set total resource constraints: ; in, This refers to the platform's total computing power capacity. Under the condition of satisfying the total resource constraint, perform bilateral matching or priority queue matching on multiple submitted user resource requests, and allocate the successfully matched user tasks to the corresponding computing resources for execution.

[0012] Specifically, triggering cross-regional resource scheduling involves selecting a migration path based on a migration cost function, migrating some tasks from the first region to the second region for execution. The migration cost function is: ; in, For regional distance, For bandwidth usage, For migration delay, , , These are the distance cost weight, bandwidth usage cost weight, and migration delay weight, respectively. The optimization objective of cross-regional resource scheduling is:

[0013] Under the condition of meeting regional capacity constraints Under the given conditions, the region with the lowest migration cost is selected as the target migration region for task migration. The current load of the target area, This represents the maximum capacity of the target area.

[0014] This includes collecting system operation feedback data after transaction matching and cross-regional scheduling execution, and using the feedback data to retrain the AI ​​prediction model and update the model parameters, specifically including: The system collects operational feedback data, including actual transaction prices, user satisfaction, GPU utilization, and average task latency. Construct the model loss function using the feedback data: ; in, For real needs, To predict demand; Update according to the gradient descent formula: ; Update the parameters of the AI ​​prediction model. ,in, For model parameters, For learning rate, This is the gradient of the loss function with respect to the model parameters.

[0015] The preprocessing of the collected data includes denoising, standardization, and missing data imputation, specifically: The standardization process uses the Min-Max normalization method, and the normalization formula is: ; in, The original data, For the normalized data, The maximum value of the data in the current dimension. This is the minimum value of the data in the current dimension.

[0016] This includes collecting user historical task data, resource operation status data, and computing power transaction record data, specifically including: The user's historical task data includes task submission time, task type, CPU demand, GPU demand, inference request volume, latency requirements, and historical payment price, forming a user feature vector: ; in, For the user's task submission time, CPU demand, For GPU demand, For AI reasoning needs, For time delay constraints, Paying prices based on history; The resource operation status data includes CPU utilization, GPU utilization, node idle rate, network bandwidth, power cost, and heat dissipation status, forming a resource status vector: ; in, For CPU utilization, For GPU utilization, For network bandwidth, For electricity costs, This is a heat dissipation load.

[0017] This invention discloses a method for dynamic pricing and cross-domain resource scheduling in heterogeneous computing power transactions. It constructs a training dataset by collecting historical user task data, resource operation status data, and computing power transaction records. Using an AI prediction model, it performs structured predictions of CPU, GPU, and AI inference demands within a future time window. Based on these predictions, it calculates the dynamic value weights of the three heterogeneous resources (CPU, GPU, and AI inference) according to the prediction results, resource utilization, and energy costs. These resources are then mapped to a unified computing power value space. A Stackelberg game model, with the computing power platform as the leader and users as followers, generates dynamic resource prices. Subsequently, it matches user tasks with computing power resources with the goal of maximizing user utility and minimizing task latency. When regional load exceeds a threshold, cross-regional scheduling is triggered. The optimal migration path is selected based on the migration cost function to execute task migration. Finally, system operation feedback data is collected to retrain the AI ​​prediction model and update its parameters, thereby achieving prediction-driven proactive pricing and cross-domain resource collaborative scheduling. This invention can anticipate future trends in computing power demand and proactively adjust pricing strategies accordingly. Through a dynamic unified value assessment mechanism, it achieves unified quantification and trading of heterogeneous resources such as CPUs, GPUs, and AI inference. By leveraging cross-regional resource scheduling, it effectively alleviates local congestion and improves overall resource utilization. Ultimately, it achieves the technical effects of increasing GPU utilization, reducing average user latency, reducing resource idle rate, and improving platform revenue and overall social welfare. It effectively solves the problems of low resource utilization, poor trading efficiency, and unsatisfactory system revenue caused by the lack of future demand perception, difficulty in unified trading of heterogeneous computing power, delayed response of passive pricing mechanisms, and lack of cross-regional collaboration in existing technologies. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0019] Figure 1 This is a flowchart of the steps of the heterogeneous computing power transaction dynamic pricing and cross-domain resource scheduling method of the present invention. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0021] Please refer to Figure 1 This invention provides a method for dynamic pricing and cross-domain resource scheduling in heterogeneous computing power transactions, comprising the following steps: S101: Collect user historical task data, resource operation status data, and computing power transaction record data. Perform noise reduction, standardization, and missing data imputation preprocessing on the collected data, and slice it according to time windows to form a training dataset. Specifically, the system collects historical user behavior, resource operation status, and computing power transaction records to provide a data foundation for subsequent AI demand prediction, dynamic pricing, and resource scheduling. The system primarily collects three types of data: The first category is user-side data, including task submission time, task type, CPU demand, GPU demand, inference request volume, latency requirements, and historical payment prices, forming a user feature vector: ; in, For the user's task submission time, CPU demand, For GPU demand, For AI reasoning needs, For time delay constraints, Historically paid prices.

[0022] The second category is resource-side data, including CPU utilization, GPU utilization, node idle rate, network bandwidth, power cost, and heat dissipation status, forming a resource status vector: ; in, For CPU utilization, For GPU utilization, For network bandwidth, For electricity costs, This is a heat dissipation load.

[0023] The third category is market transaction data, including actual transaction prices, resource supply, resource demand, rejection rate, and user waiting time.

[0024] The collected data underwent preprocessing, specifically including denoising, standardization, missing data imputation, and time window slicing. The standardization process employed Min-Max normalization to apply a uniform scale mapping to data across different dimensions. The normalization formula is as follows: ; in, The original data, For the normalized data, The maximum value of the data in the current dimension. This is the minimum value of the data in the current dimension.

[0025] By mapping data of different dimensions to the [0,1] interval, the impact of different scales of data such as CPU, GPU, price, and latency on model training can be avoided, thereby improving the convergence speed and stability of AI prediction models.

[0026] The final training dataset includes user-side data, resource-side data, and market transaction data.

[0027] S102: Input the training dataset into the AI ​​prediction model to predict the heterogeneous computing power demand within the future time window, and output the predicted values ​​of CPU demand, GPU demand, and AI inference demand. Specifically, this step is used to predict changes in different types of computing power demand within a future time window, including CPU demand, GPU demand, and AI inference demand.

[0028] The input uses a sliding time window and includes historical GPU utilization, CPU request time series, AI inference request time series, number of online users, and current time period features. The input sequence is defined as follows: ; in, The time window length, for The status of resource demand at any given moment; The prediction model employs an LSTM network, which extracts temporal features through a forget gate, an input gate, and a unit state update mechanism. The LSTM state update formula is as follows: Forgotten Gate: Input Gate: Candidate state: Cell status update: Hidden layer output: in, For the Gate of Oblivion For input gate, For the unit state, In hidden state, For the Sigmoid function, Hadamard product Output prediction results: ; in, To predict CPU demand, To predict GPU demand, To predict AI inference needs.

[0029] Traditional methods predict total demand, while this invention predicts a heterogeneous demand structure consisting of CPU, GPU, and AI inference demands, thereby enabling finer-grained resource control.

[0030] S103: Based on the heterogeneous computing power demand prediction results, combined with the current resource utilization rate and energy consumption cost, calculate the dynamic value weights of three heterogeneous resources: CPU, GPU and AI inference, and map the three heterogeneous resources to a unified computing power value space to obtain the unified computing power value assessment results. Specifically, this step addresses the problem of the inability to unify the trading of CPU, GPU, and AI inference computing power. This invention proposes a Dynamic Unified Computing Power (DUCP) model, which maps CPU, GPU, and AI inference resources to a unified value space.

[0031] Building a unified computing power value model: ; in, CPU resource quantity For GPU resources, For AI inference resources, , , These are CPU dynamic weights, GPU dynamic weights, and AI inference dynamic weights, respectively. Unlike traditional fixed weights, in this invention... , , The value of resources changes dynamically with market conditions; that is, it is not a fixed constant but a dynamic parameter based on supply and demand. The dynamic weights are updated dynamically according to changes in market demand. The formula for updating CPU dynamic weights is: ; The formula for GPU dynamic weight update is: ; The formula for updating dynamic weights in AI inference is: ; in, For CPU demand growth rate, For GPU demand growth rate, For the growth rate of AI inference demand, For CPU utilization, For GPU utilization, For GPU power consumption cost; Since GPUs are a core resource in the AI ​​computing power market, this invention focuses on constructing a dynamic value model for GPUs and calculating the comprehensive dynamic value of GPUs based on updated dynamic weights: ; in, For the dynamic value of GPU, For GPU dynamic weights, For GPU utilization, For GPU demand growth rate, This refers to the power consumption cost of the GPU.

[0032] The formula for calculating the GPU demand growth rate is:

[0033] in: Current GPU request volume. : GPU request volume at the previous moment. This formula reflects the changing trend of GPU market demand.

[0034] GPU utilization:

[0035] in: GPU resources are already in use. Total GPU resources.

[0036] GPU power consumption cost:

[0037] in: Average GPU power consumption; Real-time electricity price.

[0038] The higher the GPU utilization rate, the higher the real-time value of the GPU; the faster the GPU demand grows, the stronger the market scarcity; the higher the GPU energy consumption cost, the higher the resource price. Therefore, this model can achieve a joint dynamic value assessment of "GPU supply and demand status, scarcity, and energy consumption cost".

[0039] S104: Based on the heterogeneous computing power demand forecast results and the unified computing power value assessment results, establish a Stackelberg game model with the computing power platform as the leader and users as followers, solve the equilibrium solution of the game model, and generate dynamic resource prices. Specifically, this step uses game theory to generate the optimal resource price. In the game, the computing platform acts as the leader, responsible for publishing resource prices and controlling resource supply; users act as followers, responsible for deciding the amount to purchase and whether to uninstall tasks.

[0040] Constructing user utility functions: ; in, Apply for computing power for users, Price per unit For time delay sensitivity coefficient, For local execution latency, Delay for task unloading; Define the total latency for task unloading: ; in, To increase the amount of data uploaded, For wireless transmission rate, For task computation, To allocate computing power; Wireless transmission rates are calculated using Shannon's formula:

[0041] in: Channel bandwidth Transmission power, Channel gain Noise power spectral density; Regarding the user utility function Differentiate:

[0042] make:

[0043] To obtain the optimal resource request volume for the user: ; The higher the user's latency sensitivity and the greater the task's computational load, the more resources they request; the higher the price, the fewer resources they request.

[0044] Constructing the platform profit function: ; in, Price per unit Unit operating cost To meet user resource needs; Based on the ratio of AI-predicted demand to current supply, dynamic resource prices are updated according to the following formula: ; in, To predict demand, For current supply, To adjust the step size.

[0045] Prices rise when demand exceeds supply and fall when supply is plentiful. The system dynamically adjusts prices based on AI-predicted demand, current resource utilization, and users' willingness to pay.

[0046] S105: Based on the dynamic resource price, receive the resource request submitted by the user, which includes task type, budget, latency constraints and resource requirement parameters. With the optimization goal of maximizing user utility and minimizing task latency, perform transaction matching between user tasks and computing resources, and output the transaction matching result. Specifically, this step is used to achieve automatic matching of supply and demand.

[0047] Receive resource requests submitted by users, including task type, budget limit, maximum tolerable latency, CPU requirements, and GPU requirements.

[0048] With the optimization objectives of maximizing user utility and minimizing task latency, the optimization function is defined as follows:

[0049] in, For user benefit, For time delay, This is the delay penalty coefficient; Set total resource constraints: ; in, This refers to the platform's total computing power capacity. Under the condition of satisfying the total resource constraint, perform bilateral matching or priority queue matching on multiple submitted user resource requests, and allocate the successfully matched user tasks to the corresponding computing resources for execution.

[0050] When a user task includes GPU requirements, the task matching priority is calculated using the following formula:

[0051] in, Pay according to user budget. The maximum tolerable latency for the task. and The priorities are budget weight and latency weight, respectively, and resources are matched to tasks that require GPUs in descending order of priority.

[0052] S106: Periodically monitor the resource load status of computing nodes in each region. When the load of the first region exceeds the preset threshold and there are idle resources in the second region, trigger cross-regional resource scheduling, select a migration path according to the migration cost function, and migrate some tasks in the first region to the second region for execution. Specifically, this step is used to achieve cross-regional computing resource collaboration. The system maintains multiple regional resource pools, including GPU clusters, CPU clusters, and AI inference nodes.

[0053] The resource load status of computing nodes in each region is periodically monitored. When resource congestion is detected in the first region and idle resources are detected in the second region, task migration is automatically executed.

[0054] The migration cost function is: ; in, For regional distance, For bandwidth usage, For migration delay, , , These are the distance cost weight, bandwidth usage cost weight, and migration delay weight, respectively. The optimization objective of cross-regional resource scheduling is:

[0055] Under the condition of meeting regional capacity constraints Under the given conditions, the region with the lowest migration cost is selected as the target migration region for task migration. The current load of the target area, This represents the maximum capacity of the target area.

[0056] S107: Collect system operation feedback data after transaction matching and cross-regional scheduling execution, and use the feedback data to retrain the AI ​​prediction model and update the model parameters.

[0057] Specifically, this step is used to achieve adaptive optimization of the system.

[0058] The system collects operational feedback data, including actual transaction prices, user satisfaction, GPU utilization, and average task latency.

[0059] Construct the model loss function using the feedback data: ; in, For real needs, To predict demand; Update according to the gradient descent formula: ; Update the parameters of the AI ​​prediction model. ,in, For model parameters, For learning rate, This is the gradient of the loss function with respect to the model parameters.

[0060] The system continuously collects feedback data such as transaction prices, user satisfaction, GPU utilization, and average latency, and retrains the AI ​​prediction model to achieve adaptive optimization. Ultimately, it forms a dynamic closed-loop system of "prediction → pricing → transaction → scheduling → feedback → re-prediction".

[0061] Example 1:

[0062] (1) System Deployment Environment. This embodiment constructs a heterogeneous computing power network environment, including a GPU training cluster, AI inference nodes, a CPU computing cluster, and GPU inference nodes. Among them, the total number of GPU resources is 2000, and the total number of CPU resources is 5000 cores. The system adopts Kubernetes and mobile edge computing system for collaborative deployment, and the computing power nodes are interconnected through a high-speed computing power network.

[0063] (2) Historical Data Collection. The system continuously collects GPU utilization, user task request volume, AI inference request volume, market transaction prices, and user waiting time. The system constructs a time-series dataset with a 5-minute time window: The Min-Max method was used for normalization.

[0064] (3) AI Demand Forecasting Process. The system uses an LSTM network to predict GPU demand changes over the next 30 minutes. Assuming the current GPU utilization is 82%, and the number of AI inference requests continues to increase, with the number of GPU requests increasing from 1200 to 1500 in the last 10 minutes. The GPU demand growth rate is calculated as follows:

[0065] That is, the GPU demand growth rate is 25%. The LSTM model predicts that GPU demand will further increase to:

[0066] This indicates that the GPU market will face resource shortages in the future.

[0067] (4) Dynamic value assessment of GPUs: The system calculates the dynamic value of the GPU based on GPU utilization, demand growth rate, and energy consumption cost. Assumptions: GPU utilization:

[0068] GPU dynamic weights:

[0069] GPU power consumption cost:

[0070] The dynamic value of the GPU is then:

[0071] The calculation yielded:

[0072] This indicates that the current GPU resources are in a high-value state.

[0073] (5) Dynamic pricing process: The system dynamically adjusts GPU prices based on AI-predicted demand.

[0074] Assumption: Current resource supply:

[0075] AI predicts demand:

[0076] Current price:

[0077] Adjust step size:

[0078] The price will then be updated as follows:

[0079] The calculation yielded:

[0080] Because demand exceeds supply, the system automatically raises GPU prices.

[0081] (6) User resource request response: Computational complexity of a certain AI training user task:

[0082] User latency sensitivity factor:

[0083] System resource pricing:

[0084] The optimal number of resource requests for the user is:

[0085] The calculation yielded:

[0086] The system then allocates resources accordingly.

[0087] (7) Cross-regional resource allocation: When the load of the Beijing GPU cluster reaches:

[0088] Exceeding the load threshold:

[0089] The system automatically triggered cross-regional scheduling, migrating some AI inference tasks to the Shenzhen edge node.

[0090] Migration cost function:

[0091] The system selects the path with the lowest migration cost to complete the task migration.

[0092] (8) System operation results: After dynamic pricing and cross-regional scheduling, GPU utilization increased by 21%, average user latency decreased by 18%, resource idle rate decreased by 26%, platform revenue increased by 15%, and cross-regional resource utilization increased by 19%. The results show that the present invention can effectively realize AI prediction-driven dynamic pricing, GPU dynamic value assessment, cross-regional computing resource collaborative scheduling, and overall system revenue optimization.

[0093] Example 2:

[0094] To further verify the dynamic adjustment capability of the present invention under low load scenarios, this embodiment simulates the resource optimization process under the condition of decreased AI computing power demand at night.

[0095] (1) System operating status. Assume the current time is 02:00 at night. The system detects a significant decrease in the number of AI inference requests, an increase in GPU resource idle rate, and a decrease in user task arrival rate. The current GPU resource status is as follows: total GPU resources 2000, 700 GPU resources are occupied, GPU utilization rate is 35%, and the current GPU price is 10 yuan / unit.

[0096] GPU utilization is calculated as follows:

[0097] This indicates that the current system is under low load.

[0098] (2) AI demand forecast results: The system uses an LSTM model to predict changes in GPU demand over the next 30 minutes.

[0099] The prediction results show that: Future GPU request volume will be:

[0100] Descend to:

[0101] The GPU demand growth rate is:

[0102] The calculation yielded:

[0103] This means that GPU demand dropped by about 33%, indicating a significant weakening of demand in the nighttime market.

[0104] (3) Dynamic value changes of GPUs: The system calculates the dynamic value of the GPU based on GPU utilization, demand growth rate, and energy consumption cost.

[0105] Assumption: GPU dynamic weights:

[0106] GPU utilization:

[0107] GPU demand growth rate:

[0108] GPU power consumption cost:

[0109] The dynamic value of the GPU is then:

[0110] The calculation yielded:

[0111] The significantly lower dynamic value of GPUs compared to peak scenarios indicates that the system can automatically sense changes in GPU value based on market conditions.

[0112] (4) Dynamic price adjustment process: The system dynamically adjusts resource prices based on AI-predicted demand.

[0113] Assumption: Current supply:

[0114] AI predicts demand:

[0115] Current price:

[0116] Adjust step size:

[0117] The dynamic price update is as follows:

[0118] The calculation yielded:

[0119] This means the system automatically reduces the price of the GPU.

[0120] (5) Resource attraction mechanism: After the price reduction: some latency-tolerant AI training tasks began to be entered into the system; some offline training tasks that were originally waiting to be executed were activated; and the purchase volume of user resources increased significantly.

[0121] Optimal resource request volume for users:

[0122] Due to price decreases:

[0123] therefore:

[0124] The system successfully attracted more low-priority tasks to be executed during idle periods.

[0125] (6) System optimization effect: After dynamic price reduction and resource guidance, GPU utilization increased by 28%, nighttime resource idle rate decreased by 35%, average resource utilization increased by 22%, and platform long-term revenue increased by 11%.

[0126] In summary, this invention constructs a training dataset by collecting historical user task data, resource operation status data, and computing power transaction records. It then uses an AI prediction model to perform structured predictions of CPU, GPU, and AI inference demands within a future time window. Based on these predictions, the dynamic value weights of these three heterogeneous resources (CPU, GPU, and AI inference) are calculated according to the prediction results, resource utilization, and energy costs. These resources are then mapped to a unified computing power value space. A Stackelberg game model, with the computing power platform as the leader and users as followers, is established to generate dynamic resource prices. Subsequently, user tasks are matched with computing power resources with the goal of maximizing user utility and minimizing task latency. When regional load exceeds a threshold, cross-regional scheduling is triggered. The optimal migration path is selected based on the migration cost function to execute task migration. Finally, system operation feedback data is collected to retrain the AI ​​prediction model and update its parameters, thereby achieving prediction-driven proactive pricing and cross-regional resource collaborative scheduling. This invention can anticipate future trends in computing power demand and proactively adjust pricing strategies accordingly. Through a dynamic unified value assessment mechanism, it achieves unified quantification and trading of heterogeneous resources such as CPUs, GPUs, and AI inference. By leveraging cross-regional resource scheduling, it effectively alleviates local congestion and improves overall resource utilization. Ultimately, it achieves the technical effects of increasing GPU utilization, reducing average user latency, reducing resource idle rate, and improving platform revenue and overall social welfare. It effectively solves the problems of low resource utilization, poor trading efficiency, and unsatisfactory system revenue caused by the lack of future demand perception, difficulty in unified trading of heterogeneous computing power, delayed response of passive pricing mechanisms, and lack of cross-regional collaboration in existing technologies.

[0127] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that implementing all or part of the above embodiments and making equivalent changes in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for dynamic pricing and cross-domain resource scheduling in heterogeneous computing power transactions, characterized in that, Includes the following steps: Collect user historical task data, resource operation status data, and computing power transaction record data. Perform noise reduction, standardization, and missing data imputation preprocessing on the collected data, and slice it according to time windows to form a training dataset. The training dataset is input into the AI ​​prediction model to predict the heterogeneous computing power demand within a future time window, and the predicted values ​​of CPU demand, GPU demand, and AI inference demand are output. Based on the heterogeneous computing power demand forecast, combined with the current resource utilization and energy consumption costs, the dynamic value weights of three heterogeneous resources—CPU, GPU, and AI inference—are calculated, and the three heterogeneous resources are mapped to a unified computing power value space to obtain the unified computing power value assessment results. Based on the heterogeneous computing power demand forecast results and the unified computing power value assessment results, a Stackelberg game model with the computing power platform as the leader and users as followers is established. The equilibrium solution of the game model is solved to generate dynamic resource prices. Based on the dynamic resource price, the system receives resource requests submitted by users, which include task type, budget, latency constraints, and resource requirement parameters. With the optimization goal of maximizing user utility and minimizing task latency, the system performs transaction matching between user tasks and computing resources and outputs the transaction matching results. The resource load status of computing nodes in each region is periodically monitored. When the load in the first region exceeds a preset threshold and there are idle resources in the second region, cross-regional resource scheduling is triggered. A migration path is selected according to the migration cost function, and some tasks in the first region are migrated to the second region for execution. Collect system operation feedback data after transaction matching and cross-regional scheduling execution, and use the feedback data to retrain the AI ​​prediction model and update the model parameters.

2. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, The training dataset is input into the AI ​​prediction model to predict heterogeneous computing power demand within a future time window, and the model outputs predicted CPU demand, predicted GPU demand, and predicted AI inference demand, specifically including: Constructing the sliding time window input sequence: ； in, The time window length, for The status of resource demand at any given moment; The input sequence is fed into an LSTM network for prediction. The LSTM network extracts temporal features through a forget gate, an input gate, and a unit state update mechanism. Output prediction results: ； in, To predict CPU demand, To predict GPU demand, To predict AI inference needs.

3. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, Based on the heterogeneous computing power demand forecast, combined with current resource utilization and energy consumption costs, the dynamic value weights of three heterogeneous resources—CPU, GPU, and AI inference—are calculated. These three heterogeneous resources are then mapped to a unified computing power value space to obtain a unified computing power value assessment result, specifically including: Building a unified computing power value model: ； in, CPU resource quantity For GPU resources, For AI inference resources, , , These are CPU dynamic weights, GPU dynamic weights, and AI inference dynamic weights, respectively. The dynamic weights are dynamically updated based on changes in market demand. The formula for updating CPU dynamic weights is: ； The formula for GPU dynamic weight update is: ； The formula for updating dynamic weights in AI inference is: ； in, For CPU demand growth rate, For GPU demand growth rate, For the growth rate of AI inference demand, For CPU utilization, For GPU utilization, For GPU power consumption costs; Calculate the overall dynamic value of the GPU based on the updated dynamic weights: ； in, For the dynamic value of GPU, For GPU dynamic weights, For GPU utilization, For GPU demand growth rate, This refers to the power consumption cost of the GPU.

4. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, Based on the heterogeneous computing power demand forecast results and the unified computing power value assessment results, a Stackelberg game model is established with the computing power platform as the leader and users as followers. The equilibrium solution of the game model is solved to generate dynamic resource prices, specifically including: Constructing user utility functions: ； in, Apply for computing power for users, Price per unit For time delay sensitivity coefficient, For local execution latency, Delay for task unloading; Define the total latency for task unloading: ； in, To increase the amount of data uploaded, For wireless transmission rate, For task computation, To allocate computing power; Regarding the user utility function Taking the derivative and setting it to zero, we obtain the optimal resource request amount for the user: ； Constructing the platform profit function: ； in, Price per unit Unit operating cost To meet user resource needs; Based on the ratio of AI-predicted demand to current supply, dynamic resource prices are updated according to the following formula: ； in, To predict demand, For current supply, To adjust the step size.

5. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, Based on the dynamic resource price, the system receives resource requests submitted by users, including task type, budget, latency constraints, and resource requirement parameters. With the optimization objective of maximizing user utility and minimizing task latency, it performs transaction matching between user tasks and computing resources, and outputs the transaction matching results, specifically including: Receive resource requests submitted by users, including task type, budget limit, maximum tolerable latency, CPU requirements, and GPU requirements; With the optimization objectives of maximizing user utility and minimizing task latency, the optimization function is defined as follows: in, For the benefit of users, For time delay, This is the delay penalty coefficient; Set total resource constraints: ； in, This refers to the platform's total computing power capacity. Under the condition of satisfying the total resource constraint, perform bilateral matching or priority queue matching on multiple submitted user resource requests, and allocate the successfully matched user tasks to the corresponding computing resources for execution.

6. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, Trigger cross-regional resource scheduling, select a migration path based on the migration cost function, and migrate some tasks from the first region to the second region for execution, specifically including: The migration cost function is: ； in, For regional distance, For bandwidth usage, For migration delay, , , These are the distance cost weight, bandwidth usage cost weight, and migration delay weight, respectively. The optimization objective of cross-regional resource scheduling is: Under the condition of meeting regional capacity constraints Under the given conditions, the region with the lowest migration cost is selected as the target migration region for task migration. The current load of the target area, This represents the maximum capacity of the target area.

7. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, Collecting system operation feedback data after transaction matching and cross-regional scheduling execution, and using the feedback data to retrain the AI ​​prediction model and update the model parameters, specifically includes: The system collects operational feedback data, including actual transaction prices, user satisfaction, GPU utilization, and average task latency. Construct the model loss function using the feedback data: ； in, For real needs, To predict demand; Update according to the gradient descent formula: ； Update the parameters of the AI ​​prediction model. ,in, For model parameters, For learning rate, This is the gradient of the loss function with respect to the model parameters.

8. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, The collected data undergoes preprocessing including denoising, standardization, and missing data imputation, specifically including: The standardization process uses the Min-Max normalization method, and the normalization formula is: ； in, The original data, For the normalized data, This represents the maximum value of the data in the current dimension. This is the minimum value of the data in the current dimension.

9. The method for dynamic pricing and cross-domain resource scheduling of heterogeneous computing power transactions as described in claim 1, characterized in that, Collect user historical task data, resource operation status data, and computing power transaction record data, specifically including: The user's historical task data includes task submission time, task type, CPU demand, GPU demand, inference request volume, latency requirements, and historical payment price, forming a user feature vector: ； in, For the user's task submission time, CPU demand, For GPU demand, For AI reasoning needs, For time delay constraints, Paying prices based on history; The resource operation status data includes CPU utilization, GPU utilization, node idle rate, network bandwidth, power cost, and heat dissipation status, forming a resource status vector: ； in, For CPU utilization, For GPU utilization, For network bandwidth, For electricity costs, This is a heat dissipation load.

Images