Intelligent computing task scheduling method based on heterogeneous computing power

Abstract

The invention discloses an intelligent computing task scheduling method based on heterogeneous computing power, and belongs to the technical field of task scheduling. According to the method, the system state is sensed according to a scheduler, initial parameters are acquired, and an optimal scheduling decision which meets time delay and accuracy constraints and enables the average energy consumption to be minimum is made for an intelligent calculation task from a time slot t to a client through a Lyapunov optimization framework and a Markov approximation algorithm. The scheduler uses a Lyapunov optimization framework to convert long-term constraints of system time delay and accuracy into instantaneous decisions; the scheduler solves an instantaneous decision by using a dynamic parameter Markov approximation algorithm; and after solving by using the dynamic parameter Markov approximation algorithm, the scheduler returns an instantaneous decision making the objective function to be minimum, and the scheduler performs task execution and energy consumption calculation according to the instantaneous decision and updates the system state in time. On the premise of ensuring long-term constraint of time delay and accuracy, the average energy consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical fields of heterogeneous computing power, intelligent computing and task scheduling, and specifically to an intelligent computing task scheduling method based on heterogeneous computing power. Background Art

[0002] The execution of intelligent computing tasks relies on a heterogeneous computing network consisting of end devices, edge nodes, and cloud nodes. End devices have relatively weak computing power and often lack powerful computing devices such as GPUs. Edge nodes offer proximity to end devices and low transmission latency, but their computing resources and storage capacity are relatively limited, supporting only real-time inference for lightweight models. Cloud nodes, on the other hand, possess powerful computing clusters and massive storage resources, enabling complex model training and large-scale data processing. However, data transmission to and from the cloud introduces high latency.

[0003] Intelligent computing task scheduling refers to the process of dynamically allocating task execution locations based on task characteristics (such as computational complexity, real-time requirements, data sensitivity, etc.) and node status (such as computing load, network bandwidth, storage capacity, etc.). Taking the intelligent video analysis scenario as an example, terminal devices such as cameras are responsible for collecting video streams. The target detection task of the real-time video stream must be completed immediately at the edge node to meet the millisecond-level response requirements of security monitoring; while the feature extraction and model training tasks of historical video data can be scheduled to the cloud for batch processing. An effective scheduling strategy requires a dynamic mapping between task requirements and node capabilities. The core challenge lies in balancing computing efficiency, transmission latency, and energy costs.

[0004] Traditional task scheduling methods are mostly based on static rules (such as local execution priority, load balancing strategies, etc.), lack the ability to perceive real-time network status and dynamic changes in tasks, and it is difficult to achieve optimal resource allocation in complex scenarios.

[0005] Traditional cloud-edge-end scheduling strategies primarily address single-dimensional resource allocation issues, such as avoiding cloud node overload through load balancing algorithms or reducing edge node network dependence through local computing. However, in intelligent computing task scheduling scenarios, these strategies exhibit significant limitations: First, multi-objective optimization conflicts are prominent, with a natural trade-off between latency, accuracy, and energy consumption. While edge computing can reduce latency, limited computing power can lead to decreased model accuracy; cloud computing can improve accuracy, but data transmission increases latency and energy consumption. Second, network fluctuations significantly impact scheduling, and in large-scale task concurrency scenarios, data congestion and scheduling delays are prone to occur. Furthermore, task dynamics and node heterogeneity are difficult to adapt to. The computing power requirements of intelligent computing tasks may change dynamically with model iterations, while the computing power of edge nodes varies significantly (e.g., between embedded devices and edge servers). Traditional static strategies are unable to perceive and respond to these changes in real time.

[0006] In summary, the existing edge-cloud scheduling strategies face multiple challenges in intelligent computing task scenarios, including multi-objective optimization, network asymmetric constraints, and dynamic heterogeneous adaptation. There is an urgent need for an intelligent scheduling method that combines task characteristics, node status, and network conditions to achieve efficient utilization of edge-cloud resources and comprehensive improvement of computing performance. Summary of the Invention

[0007] The purpose of the present invention is to provide an intelligent computing task scheduling method based on heterogeneous computing power. By constructing a multi-dimensional performance evaluation model of latency, accuracy, and energy consumption, combining the Lyapunov optimization framework and the dynamic parameter Markov approximation algorithm, according to the characteristics of the intelligent computing task, the edge cloud computing power status and network conditions, the distribution strategy of tasks at the edge and the cloud is dynamically optimized to solve the problem that the traditional edge cloud scheduling strategy cannot effectively balance the latency, accuracy, energy consumption and edge cloud node computing power in multi-objective optimization and computing power heterogeneous scenarios, and the computing power resource utilization rate is low.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] The intelligent computing task scheduling method based on heterogeneous computing power includes the following steps:

[0010] After the new task arrives at the user terminal k, the scheduler senses the system state and obtains the initial parameters and determines the average accuracy of the system in time slot t and the total energy consumption in time slot t based on the sensed system state and the obtained initial parameters;

[0011] Determine the long-term constraints on system delay and accuracy based on the system's average accuracy and total energy consumption in time slot t;

[0012] The scheduler uses the Lyapunov optimization framework to transform long-term constraints into instantaneous decisions. Specifically, by constructing a Lyapunov virtual queue, the long-term constraints on system latency and accuracy are transformed into a queue stability control problem. The cumulative effect of constraint deviations is quantified by defining Lyapunov and drift functions. A trade-off parameter is introduced to achieve a linear combination of the drift function and the energy optimization objective. This decouples the long-term optimization problem across time slots into a single-time slot instantaneous decision problem, which can be solved solely based on the queue state and system parameters of the current time slot.

[0013] The scheduler uses a dynamic parameter Markov approximation algorithm to solve instantaneous decisions;

[0014] After solving the problem using the dynamic parameter Markov approximation algorithm, the scheduler returns an instantaneous decision (α * ,β * ), the scheduler makes instantaneous decisions (α * ,β * ) to perform task execution and energy consumption calculation, and update the system status in time, where α* Indicates whether the task is offloaded to the edge server or cloud server, β * Indicates the server to which the download is to be made.

[0015] The scheduler is deployed on the cloud server and is used to schedule each request task, scheduling n request tasks that arrive at the same time each time;

[0016] According to the above technical solution, the system status includes:

[0017] User-side device status, including the remaining capacity S of the task buffer k (t)-C k (t), CPU computing power MIPS C (t), local model accuracy A UC ;

[0018] Among them, the maximum capacity of the client is Task buffer, S k (t) represents the original task size of the buffer of user k in time slot t, N k (t) and C k (t) represents the new tasks and completed tasks in the buffer of user k in time slot t, where N k (t)≤S max , so the buffer update formula is

[0019] The client's energy consumption is negligible, and only the computational delay is considered. The average accuracy of the model deployed on the user side is a constant A. UE , using MIPS C (t) represents the computing power of the CPU of user k in time slot t (measured by the number of tasks completed per unit time τ), let Indicates that user k wants to complete S in time slot t k (t) Task calculation time:

[0020] The actual computation time of user k in time slot t is: The total delay of the user end in time slot t: τ is the delay constraint.

[0021] Edge server status, including edge server buffer capacity The current amount of unloaded tasks S MEC (t), GPU computing power MIPS G (t), power p EC (t), marginal model accuracy A EC ;

[0022] Among them, the edge server buffer capacity S MEC (t), compared with the user side, the edge side or cloud side has more powerful processing capabilities. Therefore, once the offloading decision is made, all tasks S in the buffer k (t) will be offloaded, and the capacity of the edge server is

[0023] The average accuracy of the model deployed at the edge is constant A EC , using MIPS G (t) represents the computing power of the GPU of the edge server at time slot t, using p EC (t) represents the power of the edge server at time slot t, let represents the computation time of the edge server to which the task of user device k is offloaded in time slot t:

[0024] p EC (t) represents the power of the edge server at time slot t, represents the energy consumption of the edge server to which the task of user device k is offloaded in time slot t:

[0025] Cloud server status, including cloud service transmission delay L CC (t), cloud service model accuracy A CC , single task energy consumption e CC ;

[0026] Among them, the task transmission delay from edge server to cloud server is mainly determined by the propagation delay, which is determined by the geographical distance and is regarded as a constant l CC The energy consumption generated by each task in the cloud server is basically fixed, so the energy consumption can be regarded as a constant e CC .

[0027] Transmitter status, including the number of currently unloaded users, channel bandwidth, and modulation level.

[0028] The channel is a noiseless channel, and the modulation level L of the channel is a constant. Therefore, according to the Nyquist theorem, the upload rate r of user k in time slot t is k (t) is: Therefore, the transmission delay of user k in time slot t is for: M(t) represents the modulation signal, and B represents the bandwidth.

[0029] In summary, the total latency and total energy consumption at the edge and cloud are as follows:

[0030] The total delay and total energy consumption of the edge end in time slot t:

[0031] Total edge delay LEC (t):

[0032] Total energy consumption at the edge Eng EC (t):

[0033] The total delay and total energy consumption of the cloud server in time slot t:

[0034] Total cloud latency L CC (t):

[0035] Cloud total energy consumption Eng CC (t):

[0036] The total system delay is composed of the user device computing delay, edge server computing delay, edge transmission delay, and cloud-edge transmission delay. Therefore, we can get the total system delay at time slot t as L(t) = L UE (t)+L EC (t)+L CC (t), and in a long-term operation environment, the total average delay of each time slot must meet the constraint conditions:

[0037] According to the above technical solution, the average accuracy of the system in time slot t is defined as:

[0038]

[0039] The average accuracy of the model deployed on the user side is a constant A UE , the average accuracy of the model deployed on the edge is constant A EC , the average accuracy of the model deployed in the cloud is a constant A CC . S k (t) indicates that C k (t) represents, α k (t) indicates whether the intelligent computing task k is unloaded at time slot t, β k (t) indicates to which server the intelligent computing task k is offloaded at time slot t.

[0040] It is required that in a long-term operation environment, the average accuracy of each time slot must be greater than the accuracy constraint limit value A required by the system. Limit ,Right now

[0041] The system ignores the energy consumption generated by user devices and only focuses on the energy consumption generated by edge servers and cloud servers. Therefore, the total energy consumption of the system in time slot t is: Eng(t) = Eng EC (t)+Eng CC (t).

[0042] In the formula, Eng EC (t) represents the total energy consumption at the edge, Eng CC (t) represents the total energy consumption of the cloud.

[0043] According to the above technical solution, the steps of converting the long-term constraint into the instantaneous decision-making include:

[0044] Constructing a Lyapunov virtual queue; the Lyapunov virtual queue includes:

[0045] Delay constraint queue q1(t), reflecting the current delay and constraint value L Limit The cumulative deviation of

[0046] Accuracy constraint queue q2(t), reflecting the current delay and constraint value A Limit The cumulative deviation.

[0047] Determine the Lyapunov virtual queue update rule:

[0048]

[0049] Determine the Lyapunov function and the drift function;

[0050] The drift function is combined with the energy consumption E(t), and a linear combination of the trade-off parameter V optimization objectives is introduced to decouple the long-term optimization problem across time slots into a single time slot instantaneous decision problem.

[0051] Where q k (t) represents the kth Lyapunov virtual queue in time slot t, q k (t+1) represents the kth Lyapunov virtual queue at time slot t+1, δ k (t) represents the cumulative deviation of the delay and accuracy in time slot t from their respective constraints, k∈{1,2}, A Limit Indicates the system accuracy constraint limit value, L Limit represents the limit value of the total average delay of each time slot, L(t) represents the total delay of the system in time slot t, A(t) represents the average accuracy in the time slot, and δ1(t) represents the delay in time slot t and the constraint value L Limit The cumulative deviation of δ2(t) represents the time delay in time slot t and the constraint value A Limit The cumulative deviation.

[0052] According to the above technical solution, the Lyapunov function:

[0053] Where, Indicates that the parameter is The Lyapunov function of Represents the virtual queue of time slot t, which is a vector containing two components q1(t) and q2(t). Represents the virtual queue at time slot t+1.

[0054] The drift function:

[0055] Where, Indicates that the parameter is The Lyapunov function.

[0056] According to the above technical solution, the single-slot instantaneous decision-making problem is:

[0057] min{q1(t)L(t)-q2(t)A(t)+V·Eng(t);

[0058] Where V represents the trade-off parameter.

[0059] This formula transforms the original long-term optimization problem into a single-time-slot instantaneous decision problem P2, which can be solved by only using the queue states q1(t), q2(t) of the current time slot and the system parameters.

[0060] According to the above technical solution, the solution steps of the Markov approximation algorithm include:

[0061] Initialization decision: All tasks are offloaded to the edge server first, i.e. α k =1(unload), β k =0 (edge ​​end).

[0062] Dynamic parameter adjustment: Define the dynamic search parameter γ(t) to control the search range and probability transfer strength:

[0063] Markov chain iterative search: Perform R max Iterates, and each iteration performs the following operations: adjusts the exploration scope in real time based on the Lyapunov queue status and energy consumption changes - when the queue deviation increases, the search is narrowed to prioritize the constraints; when the energy consumption fluctuates, the search is expanded to generate candidate decisions in a random perturbation manner, and uses an exponential probability function to accept a better solution or tolerate a suboptimal solution;

[0064] Output the optimal decision: After the iteration, select the decision that minimizes the objective function (α * ,β * ), which determines whether each task is to be uninstalled and where the uninstallation is to be performed.

[0065] Based on Markov Chain Monte Carlo theory, the algorithm introduces dynamically adjusted search parameters and adjusts the exploration scope in real time based on the Lyapunov queue status and energy consumption fluctuations. When the queue deviation increases, the search is narrowed to prioritize constraint satisfaction, and when energy consumption fluctuates violently, the search is expanded to seek a stable solution. The algorithm generates candidate decisions using random perturbations and utilizes an exponential probability function to accept a superior solution or tolerate a suboptimal solution, avoiding falling into a local optimum. Through multiple rounds of iterative searches in the feasible solution space, the algorithm efficiently approaches the instantaneous optimal scheduling decision within a single time slot, providing real-time decision support for the Lyapunov optimization framework and ensuring a dynamic trade-off between latency, accuracy, and energy consumption under complex constraints, achieving both efficient solutions and theoretical performance guarantees.

[0066] According to the above technical solution, the dynamic search parameter γ(t)

[0067]

[0068] In the formula, γ0 is the benchmark control parameter, which controls the initial exploration ability; q 1,max ,q 2,max are the upper tolerance limits of the latency and accuracy of the Lyapunov queue respectively; Indicates the average energy consumption; ΔEng(t) = |Eng(t) - Eng(t-1)| indicates the variation of energy consumption.

[0069] Including another embodiment, an electronic device includes a memory and a processor, the memory stores a computer program, and when the computer program runs on the processor, it executes the intelligent computing task scheduling method based on heterogeneous computing power described in the above technical solution.

[0070] Another embodiment is also included, a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the intelligent computing task scheduling method based on heterogeneous computing power described in the above technical solution.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] 1. This paper constructs a long-term constraint framework based on Lyapunov optimization, transforming the long-term constraints on latency and accuracy into a virtual queue stability control problem. By dynamically adjusting trade-off parameters, it achieves decoupled control between energy optimization and performance assurance. This framework, without relying on future state predictions and solely based on real-time queue deviation adjustment strategies, significantly improves the system's adaptability to dynamic, heterogeneous environments. While maintaining long-term constraints on latency and accuracy, it achieves reduced average energy consumption and avoids the resource waste associated with traditional static strategies that excessively pursue a single objective.

[0073] 2. This paper proposes a dynamic parameter Markov approximation instantaneous decision algorithm. This algorithm adjusts the search range by real-time sensing of queue status and energy consumption fluctuations, and leverages a probabilistic acceptance mechanism to balance exploration and exploitation. This algorithm generates and evaluates candidate solutions within a single time slot, significantly improving solution efficiency compared to traditional heuristic algorithms. Furthermore, theoretical proof is provided to ensure long-term system performance convergence, avoiding local optima. This algorithm is particularly suitable for complex scenarios with heterogeneous edge node computing power and time-varying network bandwidth.

[0074] 3. Through hierarchical task allocation and multi-dimensional performance modeling, the present invention retains lightweight tasks on the user side, offloads medium tasks to the edge, and schedules complex tasks on the cloud, matching the computing power-energy consumption characteristics of different nodes. The edge optimizes computing time based on a dynamic power model, and the cloud uses fixed energy consumption characteristics to process high-throughput tasks, optimizing the overall energy consumption structure of the system. Under the same task load, the overall energy consumption is lower than that of traditional cloud-edge collaborative solutions, while meeting real-time and accuracy requirements, providing an efficient solution for green intelligent computing. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0076] Figure 1 Schematic diagram of a "cloud-edge-end" framework for intelligent computing task scheduling using the intelligent computing task scheduling method based on heterogeneous computing power of the present invention;

[0077] Figure 2 This is a schematic diagram of intelligent computing task scheduling. DETAILED DESCRIPTION

[0078] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0079] The present invention provides a technical solution, a method for scheduling intelligent computing tasks based on heterogeneous computing power, the steps of which include:

[0080] S1: The scheduler perceives the system status and obtains initial parameters. These system status and initialization parameter information provide the basis for the next decision. When a new task arrives at the user end k, the scheduler first needs to perceive the system's current time slot t, including but not limited to the following status information:

[0081] User-side status information: task buffer remaining capacity S k(t)-C k (t), CPU computing power MIPS C (t), local model accuracy A UC .

[0082] Edge status information: Maximum capacity of the edge server buffer The current amount of tasks to be unloaded S MEC (t), GPU computing power MIPS on edge servers G (t), power p EC (t), marginal model accuracy A EC .

[0083] Cloud status information: Cloud service transmission delay L CC (t), cloud service model accuracy A CC , single task energy consumption e CC .

[0084] Transmission status: the number of users that need to be offloaded M(t), the channel bandwidth B Hz, and the number of modulation levels L.

[0085] Decision variable modeling: Let variable α k (t) indicates whether the intelligent computing task k is unloaded at time slot t, β k (t) indicates which server the intelligent computing task k is offloaded to at time slot t. k (t),β k (t)∈{0,1},k∈N,β k The value of (t) is only k Valid when (t)=1.

[0086]

[0087] The average accuracy of the system in time slot t is defined as:

[0088]

[0089] The average accuracy of the model deployed on the user side is a constant A UE , the average accuracy of the model deployed on the edge is constant A EC , the average accuracy of the model deployed in the cloud is a constant A CC , S k (t) indicates that C k (t) represents, α k (t) indicates whether the intelligent computing task k is unloaded at time slot t, β k (t) indicates which server the intelligent computing task k is offloaded to at time slot t. It is required that in a long-term operation environment, the average accuracy of each time slot must be greater than the accuracy constraint limit value A required by the system.Limit ,Right now

[0090] The system ignores the energy consumption generated by user devices and only focuses on the energy consumption generated by edge servers and cloud servers. Therefore, the total energy consumption of the system in time slot t is: Eng(t) = Eng EC (t)+Eng CC (t);

[0091] In the formula, Eng EC (t) represents the total energy consumption at the edge, Eng CC (t) represents the total energy consumption of the cloud.

[0092] S2: The scheduler uses the Lyapunov optimization framework to transform long-term constraints into instantaneous decisions. Specifically:

[0093] S201. Construct a Lyapunov virtual queue. The Lyapunov virtual queue is represented by a vector: Among them, the delay constraint queue: q1(t), reflects the current delay and constraint value L Limit Accuracy constraint queue: q2(t), reflecting the current delay and constraint value A Limit The cumulative deviation.

[0094] S202. Determine the Lyapunov virtual queue update rule:

[0095]

[0096] Where q k (t) represents the kth Lyapunov virtual queue in time slot t, q k (t+1) represents the kth Lyapunov virtual queue at time slot t+1, δ k (t) represents the cumulative deviation of the delay and accuracy in time slot t from their respective constraints, k∈{1,2}, A Limit Indicates the system accuracy constraint limit value, L Limit represents the limit value of the total average delay of each time slot, L(t) represents the total delay of the system in time slot t, A(t) represents the average accuracy in the time slot, and δ1(t) represents the delay in time slot t and the constraint value L Limit The cumulative deviation of δ2(t) represents the time delay in time slot t and the constraint value A Limit The cumulative deviation.

[0097] S203. Determine the Lyapunov function With drift function

[0098] Where, Indicates that the parameter is The Lyapunov function of represents the virtual queue at time slot t, represents the virtual queue at time slot t+1; Indicates that the parameter is The Lyapunov function.

[0099] S204, combining the drift function with the energy consumption E(t), and introducing a linear combination of the trade-off parameter V optimization objectives, decoupling the long-term optimization problem across time slots into a single time slot instantaneous decision problem;

[0100] Among them, the single-slot instantaneous decision problem is:

[0101] min{q1(t)L(t)-q2(t)A(t)+V·Eng(t);

[0102] Where V represents the trade-off parameter.

[0103] S3: The scheduler uses the dynamic parameter Markov approximation algorithm to solve the instantaneous decision. The specific steps include:

[0104] Initialization decision: All tasks are offloaded to edge servers first;

[0105] Dynamic parameter adjustment: Define the dynamic search parameter γ(t) to control the search range and probability transfer strength:

[0106]

[0107] Among them, γ0 is the benchmark control parameter, which controls the initial exploration ability; q 1,max ,q 2,max are the upper tolerance limits of the latency and accuracy of the Lyapunov queue respectively; Indicates the average energy consumption; ΔEng(t) = |Eng(t) - Eng(t-1)| indicates the variation of energy consumption.

[0108] Markov chain iterative search: Perform R max Iterates, and each iteration performs the following operations: adjusts the exploration scope in real time based on the Lyapunov queue status and energy consumption changes - when the queue deviation increases, the search is narrowed to prioritize the constraints; when the energy consumption fluctuates, the search is expanded to generate candidate decisions in a random perturbation manner, and uses an exponential probability function to accept a better solution or tolerate a suboptimal solution;

[0109] Output the optimal decision: After the iteration, select the decision that minimizes the objective function (α * ,β * ), which determines whether each task is to be uninstalled and where the uninstallation is to be performed.

[0110] Through multiple rounds of iterative searches in the feasible solution space, the algorithm efficiently approaches the instantaneous optimal scheduling decision within a single time slot, providing real-time decision support for the Lyapunov optimization framework, ensuring a dynamic trade-off between latency, accuracy, and energy consumption under complex constraints, and achieving both solution efficiency and theoretical performance guarantees.

[0111] S4: After solving the problem using the dynamic parameter Markov approximation algorithm, the scheduler returns an instantaneous decision (α * ,β * ), the scheduler makes instantaneous decisions (α * ,β * ) to execute tasks, update the buffer and constraint queue, calculate the energy consumption of local devices, edge servers, and cloud servers, and update the device status. * Indicates whether the task is offloaded to the edge server or cloud server, β * Indicates the server to which the download is to be made.

[0112] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0113] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. Intelligent computing task scheduling method based on heterogeneous computing power, characterized by: The steps include: After the new task arrives at user terminal k, the scheduler senses the system state and obtains the initial parameters and determines the average accuracy of the system in time slot t and the total energy consumption in time slot t based on the sensed system state and the obtained initial parameters; Determine the long-term constraints on system delay and accuracy based on the system's average accuracy and total energy consumption in time slot t; The scheduler uses the Lyapunov optimization framework to transform long-term constraints on system latency and accuracy into instantaneous decisions; The scheduler uses a dynamic parameter Markov approximation algorithm to solve instantaneous decisions; After solving the problem using the dynamic parameter Markov approximation algorithm, the scheduler returns an instantaneous decision that minimizes the objective function. The scheduler executes tasks and calculates energy consumption according to the instantaneous decision, and updates the system status in a timely manner.

2. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 1 is characterized in that: The system status includes: User-side device status, including remaining task buffer capacity, CPU computing power, and local model accuracy; Edge server status, including edge server buffer capacity, current offloaded tasks, GPU computing power, power, and edge model accuracy; Cloud server status, including cloud service transmission delay, cloud service model accuracy, and single-task energy consumption; Transmitter status, including the number of currently unloaded users, channel bandwidth, and modulation level.

3. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 1 is characterized in that: The average accuracy A(t) of the system at time slot t: Where A UE Indicates that the average accuracy of the model deployed on the user side is constant, A EC A represents the average accuracy of the model deployed at the edge, which is a constant. CC The average accuracy of the model deployed on the cloud is constant, S k (t) indicates that C k (t) represents, α k (t) indicates whether the intelligent computing task k is unloaded at time slot t, β k (t) indicates to which server the intelligent computing task k is offloaded at time slot t; The total energy consumption of the system at time slot t is Eng(t): Eng(t)=Eng EC (t)+Eng CC (t); In the formula, Eng EC (t) represents the total energy consumption at the edge, Eng CC (t) represents the total energy consumption of the cloud.

4. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 1 is characterized in that: The steps of converting the long-term constraints into the instantaneous decision-making include: Construct a Lyapunov virtual queue, which is represented by a vector: Among them, the delay-constrained queue q1(t) and the accuracy-constrained queue q2(t); Determine the Lyapunov virtual queue update rule: Determine the Lyapunov function and the drift function; The drift function is combined with the total energy consumption Eng(t), and a linear combination of the trade-off parameter V optimization objectives is introduced to decouple the long-term optimization problem across time slots into a single time slot instantaneous decision problem. Where q k (t) represents the kth Lyapunov virtual queue in time slot t, q k (t+1) represents the kth Lyapunov virtual queue at time slot t+1, δ k (t) represents the cumulative deviation of the delay and accuracy in time slot t from their respective constraints, k∈{1,2}, A Limit Indicates the system accuracy constraint limit value, L Limit represents the limit value of the total average delay of each time slot, L(t) represents the total delay of the system in time slot t, A(t) represents the average accuracy in the time slot, and δ1(t) represents the delay in time slot t and the constraint value L Limit The cumulative deviation of δ2(t) represents the time delay in time slot t and the constraint value A Limit The cumulative deviation.

5. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 4 is characterized in that: The Lyapunov function: Where, Indicates that the parameter is The Lyapunov function of represents the virtual queue at time slot t, represents the virtual queue at time slot t+1; The drift function: Where, Indicates that the parameter is The Lyapunov function.

6. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 4 is characterized in that: The single-slot instantaneous decision problem: min{q1(t)L(t)-q2(t)A(t)+V·Eng(t); Where V represents the trade-off parameter.

7. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 1 is characterized in that: The solution steps of the Markov approximation algorithm include: Initialization decision: All tasks are offloaded to edge servers first; Dynamic parameter adjustment: Define the dynamic search parameter γ(t) to control the search range and probability transfer intensity: Markov chain iterative search: Perform R max Iterates, and each iteration performs the following operations: adjusts the exploration scope in real time based on the Lyapunov queue status and energy consumption changes - when the queue deviation increases, the search is narrowed to prioritize the constraints; when the energy consumption fluctuates, the search is expanded to generate candidate decisions in a random perturbation manner, and uses an exponential probability function to accept a better solution or tolerate a suboptimal solution; Output the optimal decision: After the iteration, select the decision that minimizes the objective function (α * ,β * ), that is, determine whether each task is to be unloaded and where to unload; where α * Indicates whether the task is offloaded to the edge server or cloud server, β * Indicates the server to which the download is to be made.

8. The intelligent computing task scheduling method based on heterogeneous computing power according to claim 7 is characterized in that: The dynamic search parameter γ(t) Among them, γ0 is the benchmark control parameter, which controls the initial exploration ability; q 1,max represents the upper limit of the delay tolerance of the Lyapunov queue, q 2,max represents the upper tolerance limit of the accuracy Lyapunov queue; E avg It represents the average energy consumption; ΔEng(t) represents the variation of energy consumption.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program runs on the processor, the method for scheduling intelligent computing tasks based on heterogeneous computing power according to any one of claims 1 to 8 is executed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for scheduling intelligent computing tasks based on heterogeneous computing power according to any one of claims 1 to 8 is implemented.

Images