A hybrid cloud resource scheduling method and system

Abstract

This invention provides a hybrid cloud resource scheduling method and system, which relates to the field of cloud resource scheduling. The method includes: combining the Markov decision process (MDP) and Q-learning reinforcement learning to optimize task scheduling, dynamically adjust resource allocation, and utilize the M / M / c queuing model to quantify QoS constraints, thereby improving hybrid cloud resource utilization efficiency. Furthermore, the system incorporates the enhanced whale predator algorithm (EMPA) and the adaptive simulated annealing algorithm (ASA) through the EMPA-ASA model to enhance search capabilities, introduces the Lévy flight strategy into scheduling, and utilizes an adaptive update mechanism to increase global convergence speed, thereby improving scheduling stability. This method minimizes total costs while ensuring QoS.

Description

Technical Field

[0001] The present invention relates to the field of cloud resource scheduling, and in particular to a hybrid cloud resource scheduling method and system. Background Art

[0002] With the rapid development of cloud computing and big data technologies, hybrid cloud architectures have become a key approach to modern computing resource management. By combining the strengths of public and private clouds, hybrid clouds can provide enterprises with flexible and reliable resource scheduling and management capabilities. In hybrid cloud environments, the primary goal of resource scheduling is to efficiently allocate computing, storage, and network resources to various tasks based on varying Quality of Service (QoS) requirements and resource constraints. However, hybrid cloud resource scheduling faces complex challenges, including balancing multiple objectives (such as latency, energy consumption, and cost), coping with dynamic changes in task requirements, and achieving load balancing and efficient resource utilization in diverse network environments. To improve the efficiency of hybrid cloud resource scheduling, researchers have recently proposed a variety of methods based on heuristic algorithms, metaheuristic algorithms, and reinforcement learning to optimize resource allocation. While these methods have achieved some success in certain scenarios, existing scheduling algorithms still have limitations due to the complexity of hybrid cloud environments. In particular, the computational efficiency and adaptability of existing solutions remain to be improved when handling large-scale, highly dynamic tasks. During the implementation of the present invention, the applicant discovered that while many prior art studies have focused on optimizing task offloading and resource scheduling, most have focused solely on a single optimization objective (such as task completion time, energy consumption, or resource utilization), while neglecting the comprehensive optimization of user experience and cost. Particularly in high-load and dynamically changing environments, prior art approaches fail to balance the real-time changes in factors such as task urgency, network bandwidth, and node load, resulting in unstable system QoS and a poor user experience. Furthermore, prior art approaches often fail to effectively control the total cost of the system, leading to wasted resources and excessive energy consumption, thus impacting the overall performance and affordability of the system. Summary of the Invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related technology, and discloses a hybrid cloud resource scheduling method and system, which improves the user experience while improving the overall performance and economy of the hybrid cloud resource scheduling system.

[0004] Terminology Explanation: Markov Decision Process (MDP): A Markov Decision Process (MDP) is a mathematical model that describes how actions are selected across different states to maximize a certain objective (usually a reward). An MDP consists of a state space, an action space, transition probabilities, and a reward function. It is often used to solve optimization problems in uncertain and dynamic environments.

[0005] Q-value update method: The Q-value update method is a strategy in reinforcement learning that estimates the expected reward of taking an action in a specific state through the learning process. The Q-value represents the cumulative reward obtained by taking an action in a given state. It is iteratively updated using the Bellman equation to find the optimal strategy during the learning process.

[0006] The M / M / C model in queuing theory is a classic queuing model used to describe queuing systems with multiple service channels (C service windows). The M / M / C model assumes that customer arrivals follow a Poisson process (M), service times follow an exponential distribution (M), and the system has C service channels. This model is used to analyze system performance metrics such as throughput, response time, and queue length, and to optimize the system's quality of service (QoS).

[0007] Quality of Service (QoS): Quality of Service (QoS) is a measure of network service performance in computer networks. QoS typically includes parameters such as latency, bandwidth, packet loss rate, and jitter. It is used to ensure that user requests are met, especially in dynamic, multi-tasking or multi-user network environments, ensuring service reliability and stability.

[0008] Enhanced Marine Predators Algorithm (EMPA): The Enhanced Marine Predators Algorithm (EMPA) is an optimization algorithm that simulates the hunting behavior of marine predators. It uses the predator's movement patterns to conduct global and local searches to find solutions to optimization problems. The enhanced algorithm enhances search capabilities and improves optimization efficiency by introducing adaptive factors and improved search strategies, making it particularly suitable for dynamic and complex optimization problems.

[0009] Adaptive Factor: An adaptive factor is a mechanism used to dynamically adjust optimization process parameters. In task scheduling and resource allocation optimization, the adaptive factor adjusts task allocation and resource scheduling strategies based on the system's real-time status (such as load and network bandwidth), thereby achieving a balance between global and local optimization, improving optimization efficiency and system performance.

[0010] A first aspect of the present invention discloses a hybrid cloud resource scheduling method, comprising: task scheduling: in a hybrid cloud resource scheduling task, constructing a task scheduling model based on a Markov decision process to generate a task scheduling decision; iteratively updating a Q-value function of the Markov decision process using a Bellman equation;

[0011] Optimizing QoS: The reward function of the Markov decision process is defined as:

[0012] R(s,a)=Cost(s,a)+QoS(s,a)

[0013] Where R(s,a) represents the reward function of the Markov decision process, Cost(s,a) represents the total cost under the current state and action, and QoS(s,a) represents the quality of service score under the current state and action. The QoS(s,a) is optimized based on the M / M / C model of queuing theory.

[0014] Resource allocation: Based on the task scheduling strategy, an EMPA-ASA model is constructed for the resource allocation problem of each task in the scheduling strategy to generate an optimal solution for resource allocation. The calculation process of the EMPA-ASA model includes: generating an initial solution based on the enhanced marine predator algorithm; updating the adaptive factor based on the adaptive simulated annealing algorithm to generate a new solution of the enhanced marine predator algorithm; comparing the total cost of the new solution with the total cost of the initial solution, and taking the new solution as the current solution if the total cost of the new solution is less than the total cost of the initial solution; and calculating the acceptance rate of the new solution by the adaptive simulated annealing algorithm if the total cost of the new solution is greater than or equal to the total cost of the initial solution; and updating the adaptive factor of the adaptive simulated annealing algorithm based on the current solution or the acceptance rate.

[0015] In this technical solution, the hybrid cloud resource scheduling optimization problem is modeled as a Markov decision process (MDP). The optimal task scheduling policy is learned through Q-learning to achieve the coordinated optimization of quality of service (QoS) and cost under dynamic resource conditions. Based on this, the proposed EMPA-ASA algorithm conducts global resource search and scheduling solution optimization under policy guidance. The Enhanced Marine Predator Algorithm (EMPA) uses Lévy flights to achieve large-scale search, while the Adaptive Simulated Annealing Algorithm (ASA) further improves the convergence stability of the solution. The two algorithms work together to form a "policy-search" linkage mechanism, optimizing both the task decision-making and scheduling execution levels. This allows the system to minimize resource costs and improve scheduling efficiency while maintaining QoS.

[0016] According to the hybrid cloud resource scheduling method disclosed in the present invention, preferably, the total cost includes: communication cost, computing cost and storage cost; the service quality score includes: response time score, bandwidth score and task completion time score.

[0017] According to the hybrid cloud resource scheduling method disclosed in the present invention, preferably, the calculation process of the adaptive simulated annealing algorithm includes:

[0018]

[0019]

[0020] Among them, g k represents the adaptive factor, Δf is the rate of change of the objective function, and the weight coefficient 、 and It is used to control the influence of each item in the adaptive factor on temperature change, R target represents the target resource utilization, R k Indicates the resource utilization rate in round k, ΔN target Indicates the target load change rate, ΔN k represents the load change rate in the kth round, is the objective function value of the current solution, is the mean of the objective function over the first m iterations.

[0021] The hybrid cloud resource scheduling method disclosed in the present invention preferably further includes:

[0022] The process of adaptive temperature update is as follows:

[0023]

[0024] in, 、 、 Represent the objective function change rate, acceptance rate difference and neighborhood change respectively, 、 、 is the weight coefficient, which is used to control the impact of various factors on temperature update.

[0025] According to the hybrid cloud resource scheduling method disclosed in the present invention, preferably, the calculation process of the enhanced ocean predator algorithm includes:

[0026] Generate new solutions through Levy flight strategy to expand the search range of solutions and generate new solutions with stronger global search capabilities. assignments The generation formula is as follows:

[0027]

[0028]

[0029] Among them, current assignmentsis the current solution, which represents the initial solution in the initialization stage, Bias weights for predation behavior, The random values ​​generated for the Levy flight strategy, u and v are normal distribution random numbers, is the scaling parameter of the Levy distribution, is a scaling factor that controls the jump step size.

[0030] According to the hybrid cloud resource scheduling method disclosed in the present invention, preferably, the step of optimizing the QoS (s, a) based on the queuing theory M / M / C model specifically includes: optimizing the task allocation of each server by adjusting the number of servers and service rate, calculating the queue length and blocking probability, and controlling the queue length and arrival rate of each controller.

[0031] According to the hybrid cloud resource scheduling method disclosed in the present invention, preferably, in the Markov decision process, the transition of the system state follows the state transition probability distribution P(s′|s,a):

[0032]

[0033] Among them, P(s′|s,a) represents the probability of the system transitioning to the next state s′ given the current state s and action a, S t represents the system state at time t, A t represents the action taken at time t.

[0034] A second aspect of the present invention discloses a hybrid cloud resource scheduling system, comprising: a memory for storing program instructions; a processor for calling the program instructions stored in the memory to implement a hybrid cloud resource scheduling method as described in any of the above technical solutions.

[0035] Compared with existing technologies, the present invention offers at least the following advantages: it flexibly balances task priorities and system resources, achieving dual optimization of user experience and cost. This not only improves task offloading success rate and response speed, but also significantly reduces system energy consumption and communication costs, enhancing the system's economic viability and sustainability. The present invention's optimization strategy incorporates multiple layers of integration and adjustment. First, by combining the Markov decision process (MDP) with the Q-value update method from reinforcement learning, the system's state space is constructed to represent the resource usage status and task requirements. The Bellman equation is used to iteratively update the Q-value, and an epsilon-greedy strategy is employed to balance exploration and utilization, thereby optimizing resource allocation and reducing the system's total cost. Second, based on the queuing theory M / M / C model, the system's quality of service (QoS) is optimized. By adjusting the number of servers and service rates, a load balancing algorithm is employed to optimize task allocation across servers, calculate queue lengths and blocking probabilities, and ensure reliable QoS in dynamic environments. Finally, an improved enhanced marine predator algorithm (EMPA) is employed to generate an optimized solution under conditions of dynamically changing task loads. This algorithm enhances global search capabilities by introducing a fitness function and a dynamic update mechanism. It also incorporates an adaptive factor to adjust the predator's movement, optimizing task offloading decisions and resource scheduling. The adaptive factor dynamically adjusts task allocation and resource scheduling strategies based on system load and network changes, ensuring stable system operation and efficient task completion. It also balances global search with local optimization, improving user experience and reducing overall system costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of a hybrid cloud resource scheduling method according to an embodiment of the present invention is shown.

[0037] Figure 2 A schematic diagram of the process of generating a new EMPA solution and updating an ASA adaptive factor according to an embodiment of the present invention is shown.

[0038] Figure 3 A schematic block diagram of a hybrid cloud resource scheduling system according to an embodiment of the present invention is shown.

[0039] Figure 4 A schematic diagram of cloud resource scheduling according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0040] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0041] like Figure 1 According to one embodiment of the present invention, a hybrid cloud resource scheduling method is disclosed, including:

[0042] Step S1, task scheduling: In the hybrid cloud resource scheduling task, a task scheduling model is constructed based on the Markov decision process to generate task scheduling decisions; the Q value function of the Markov decision process is iteratively updated using the Bellman equation;

[0043] Step S2, optimize QoS: The reward function of the Markov decision process is defined as:

[0044] R(s,a)=Cost(s,a)+QoS(s,a)

[0045] Where R(s,a) represents the reward function of the Markov decision process, Cost(s,a) represents the total cost under the current state and action, and QoS(s,a) represents the quality of service score under the current state and action. QoS(s,a) is optimized based on the M / M / C model of queuing theory.

[0046] Step S3, generating an initial solution: generating an initial solution to the resource allocation problem based on the enhanced marine predator algorithm;

[0047] Step S4, generating a new EMPA solution: updating the adaptive factor based on the adaptive simulated annealing algorithm to generate a new EMPA solution;

[0048] Step S5, iteratively generate the optimal solution;

[0049] Step S6: record the optimal solution.

[0050] In this embodiment, the core of the Markov Decision Process (MDP) is to optimize cost and user experience in hybrid cloud resource scheduling by learning the optimal policy π to minimize total cost while meeting QoS requirements. In the MDP, system state transitions follow the state transition probability distribution P(s′|s,a), which describes the probability of the system transitioning to the next state s′ given the current state s and action a. This probability is defined in the scheduling process as:

[0051]

[0052] Among them, St represents the system state at time t, A t is the action taken in this state. The state transition probability is determined by the environment dynamics and the current load of the resource. The reward function R(s,a) is used to evaluate the immediate benefit of taking action a in state s, and the formula is defined as:

[0053] R(s,a)=Cost(s,a)+QoS(s,a)

[0054] Cost(s,a) represents the total cost of the current state and action, including communication, computing, storage, and other resource consumption; QoS(s,a) is the quality of service score for the current state and action. By optimizing R(s,a), the system can reduce the total cost of resource scheduling while ensuring QoS.

[0055] The goal of MDP is to find the optimal strategy that maximizes the cumulative reward under a given strategy π. The expected cumulative reward of this strategy is expressed as:

[0056]

[0057] in, Indicates the mathematical expectation of the subsequent formula, γ t represents the discount factor raised to the power of t;

[0058] On this basis, the Bellman equation can be used to further transform the value function V π (s) is expressed as a recursive summation of the next state:

[0059]

[0060] Among them, V π (s) represents the expected cumulative return (value) in the future under the strategy π, starting from state s; s ' represents the next state, γ represents the discount factor;

[0061] In the optimization process, the Q-value function Q(s,a) is a key element in the MDP. It represents the expected cumulative reward when taking action a in state s and satisfies the following Bellman equation:

[0062]

[0063] This equation illustrates the recursive relationship of the Q-value function: the value of the current state-action pair depends not only on the immediate reward R(s, a) but also on the weighted accumulation of the optimal Q-values ​​in the next state s'. By repeatedly updating the Q(s, a) value, the system gradually approaches the optimal policy. By learning the optimal policy π of the MDP above, the total cost is minimized.

[0064] Optimization based on user experience:

[0065] The queuing theory M / M / C model is used to optimize the QoS parameters in the system. The specific methods are as follows:

[0066] The response time in the hybrid cloud scheduling system consists of the task's queue time and service time, and is calculated using the M / M / c queuing model. The response time is calculated as follows: The system's load rate ρ represents the load of each server and is defined as:

[0067]

[0068] in, represents the request arrival rate, is the service rate of a single server, and c represents the number of servers (i.e., the number of cloud controllers). This formula helps estimate the average load of the system in a multitasking environment. Furthermore, the probability of no requests, P0, refers to the probability that there are no requests in the system, that is, the probability that the system is idle. P0 is calculated as follows:

[0069]

[0070]

[0071] P0 can effectively estimate the probability of the system being idle under different load conditions, providing data support for load management and resource allocation. In a multi-controller hybrid cloud environment, the differences in resource processing capabilities and task loads require that the average response time of each controller be calculated separately. In order to accurately evaluate the system performance under multiple controllers, we introduce the optimization objective of average response time D(t). Average response time D(t) is used to quantify the response time of user requests within time slot t, and takes into account the resource usage of private cloud and public cloud controllers. The specific definitions are as follows:

[0072]

[0073] Among them, Q j (t) and Q k (t) represents the average queue length of the private cloud and public cloud controllers in time slot t, λ j (t) and λ k (t) represents the average arrival rate (requests / second) assigned to the private cloud and public cloud controllers in time slot t. The response time of the private cloud and public cloud is determined by their average queue length Q j (t) and Q k (t) is determined by the difference between the arrival rate and the service rate. During resource scheduling, the queue length and arrival rate of each controller are controlled through reasonable allocation to minimize the value of D(t), optimize the average response time of the system, and thus optimize the user experience value.

[0074] Throughput refers to the number of tasks processed by the system per unit time, which is related to the task arrival rate and service capacity. Based on the response time calculation parameters, the throughput formula is:

[0075]

[0076] When the system load is low (ρ < 1), the throughput is equal to the task arrival rate ,When the system load is high (ρ ≥ 1), the throughput is determined by the system service capacity.,Where,c·μ,represents the total system service capacity.

[0077] Latency is the sum of the waiting time and processing time of a task in the system and is an important metric for QoS. Based on the parameters of the M / M / C model, latency is estimated by the average service time and queue time:

[0078]

[0079] Among them, B total It represents the total bandwidth demand of the system. Its ratio to the total service capacity of the system is used to calculate the average task execution time on a single server, and then the sum of the total waiting time and processing time of the task is calculated.

[0080] In addition, jitter J reflects the fluctuation of latency and response time and is calculated by multiplying the difference between response time and latency by the adjustment factor:

[0081] J=|LW|·β

[0082] L represents latency, which is the sum of the waiting time and processing time of a task. It is calculated based on the M / M / c model and reflects the total time it takes for the system to actually process tasks. W represents response time, which is determined by a combination of indicators such as the average queue length, arrival rate, and service rate of tasks in private and public clouds. β represents the jitter adjustment factor, which is used to adjust the weight of jitter in the overall QoS evaluation and control the sensitivity of latency fluctuations to the final experience indicator.

[0083] Through this approach, key performance indicators such as response time, throughput, latency, and jitter can be derived and quantified based on the system's task arrival rate, service rate, and bandwidth requirements. The calculation of these indicators provides a quantitative basis for evaluating resource scheduling performance in hybrid cloud environments and lays the foundation for comparing and optimizing different scheduling strategies.

[0084] EMPA-ASA resource allocation strategy:

[0085] (1) New solution for EMPA generation

[0086] In the simulated annealing algorithm used in this embodiment, the parameter temperature (T) refers to the key variable of the solution space search range in the system, and its role is to gradually approach the optimal resource scheduling solution by iterating different resource scheduling solutions. In the initial stage, T is set relatively high to allow the algorithm to explore different resource allocation strategies in a larger search space, improve the global search capability, and avoid falling into the local optimum too early. As the scheduling process continues to iterate, T gradually decreases, which is equivalent to gradually converging the resource scheduling solution, reducing the acceptance of poor scheduling solutions, and shifting the resource allocation strategy from extensive search to local fine optimization, and finally finding the optimal resource scheduling solution that meets the QoS constraints. To this end, this embodiment introduces an adaptive factor g in the traditional simulated annealing algorithm. k , which is used to adjust the descent rate in real time according to the solution status during the search process.

[0087] Adaptive factor g k It is used to flexibly adjust the rate of temperature change in different search stages. The calculation formula is as follows:

[0088]

[0089] in, is the rate of change of the objective function, which is used to reflect the volatility of the solution;

[0090]

[0091] in, is the objective function value of the current solution, It is before The mean of the objective function of the iterations, the difference in acceptance rate Used to determine whether to accept the suboptimal solution to prevent falling into the local optimum; neighborhood change Used to determine whether to accept a suboptimal solution to prevent falling into a local optimum:

[0092]

[0093] in, and Represent the status of the current solution and the neighborhood solution respectively, is the number of neighborhood solutions. In addition, the weight coefficient 、 and It is used to control the influence of each item in the adaptive factor on the temperature change. In general, by introducing the adaptive factor g k ,The ASA algorithm can dynamically adjust the temperature update rate according to different ,solution states, and thus balance the relationship between global ,exploration and local search during the optimization process, effectively ,enhancing the ability to explore the global optimal solution.

[0094] The formula for adaptive temperature update (adaptive dynamic step size update) is as follows:

[0095]

[0096] in, 、 、 Represent the objective function change rate, acceptance rate difference and neighborhood change respectively, 、 、 is the weight coefficient, which is used to control the impact of various factors on temperature update.

[0097] (2) Adaptive update mechanism:

[0098] The brief process of generating new solutions and updating adaptive factors is as follows: Figure 2 As shown. Among them, By dynamically adjusting the step size L to control the Lévy flight search range, a balance between global search and local optimization is achieved. After initializing the task allocation, the range is dynamically adjusted according to the search phase to adapt to different optimization requirements. The new solution is generated by the Lévy flight strategy, and the step size L is controlled by Influence, the initial large step size enhances global exploration, and the later small step size improves the accuracy of the solution. If the new solution is better than the current solution, it is directly accepted, otherwise it is decided whether to accept it according to the simulated annealing criterion to prevent falling into the local optimum. Dynamically adapt to changes in the search space, and ultimately record the optimal solution to achieve efficient and stable resource scheduling optimization.

[0099] The Enhanced Marine Predator Algorithm (EMPA) is used to generate new solutions through the Levy flight strategy to expand the search range of solutions and generate new solutions with stronger global search capabilities. The generation formula of the new solution is as follows:

[0100]

[0101] Among them, new assignments Indicates a new solution, current assignments represents the current solution or initial solution, Bias weights for predation behavior to control the balance between local and global search, is a random value generated by the Levy flight strategy, combined with the predator's behavior to achieve randomness in long and short jumps. It is usually generated by the following formula:

[0102]

[0103] in, and is a normally distributed random number, is the scaling parameter of the Levy distribution (1< ≤3), A scaling factor used to control the jump step length to ensure the distribution of long and short jumps. The calculation is done to ensure the randomness and breadth of the solution.

[0104] The goal of the EMPA-ASA algorithm is to minimize the total cost while meeting the QoS constraints. , including computation, storage, and communication costs. The optimization objective function is defined as follows:

[0105]

[0106] Among them, the calculation cost , storage costs and communication costs Represents the cost of resources in computing, storage, and communication respectively. The specific formula is as follows:

[0107]

[0108] Among them, private means private cloud, public means public cloud, 、 is the computing cost coefficient of private cloud and public cloud, 、 are the CPU and memory usage at time t respectively.

[0109] like Figure 2 As shown above, the EMPA generates a new solution combined with the updated adaptive factor g k The brief process includes: k By dynamically adjusting the step size L to control the Lévy flight search range, a balance between global search and local optimization is achieved. After initializing the task allocation, the range is dynamically adjusted according to the search phase to adapt to different optimization requirements. The new solution is generated by the Lévy flight strategy, and the step size L is controlled by g k Influence, the initial large step size enhances global exploration, and the later small step size improves the accuracy of the solution. If the new solution is better than the current solution, it is directly accepted, otherwise it is decided whether to accept it according to the simulated annealing criterion to prevent falling into the local optimum. Subsequently, g k Dynamically adapt to changes in the search space, and ultimately record the optimal solution to achieve efficient and stable resource scheduling optimization.

[0110] like Figure 3 As shown, according to another embodiment of the present invention, a hybrid cloud resource scheduling system 300 is also disclosed, including: a memory 301 for storing program instructions; a processor 302 for calling the program instructions stored in the memory to implement the hybrid cloud resource scheduling method as described in the above embodiment.

[0111] like Figure 4 According to another embodiment of the present invention, a schematic diagram of a cloud resource allocation process based on the above hybrid cloud resource scheduling system in an actual application scenario is also disclosed: Figure 4 This paper demonstrates the EMPA-ASA resource scheduling optimization process, which is divided into two parts: the task module and the resource module. First, in the task module, tasks are assigned based on their requirements. MDP reinforcement learning is used to optimize task scheduling, ensuring efficient execution and achieving expected results. Tasks are then classified according to different load levels (low, medium, and high) and submitted to the cloud resource pool for processing. Second, in the resource module, the resource constraint module assesses the QoS requirements of tasks and monitors the status of individual resources. The EMPA-ASA policy optimizes resource allocation based on the scheduling requirements of tasks and resources, minimizing computational, storage, and communication costs while satisfying QoS constraints. Finally, tasks are analyzed and rationally assigned, and the final task results are output, achieving efficient resource scheduling and task execution.

[0112] In summary, the optimization strategy of the present invention involves multiple levels of integration and adjustment. First, by combining the Markov decision process (MDP) with the Q-value update method in reinforcement learning, the system's state space is constructed to represent the usage status of each resource and task requirements. The Bellman equation is used to iteratively update the Q-value, and an epsilon-greedy strategy is employed to balance exploration and utilization, thereby optimizing resource allocation and reducing the system's total cost. Second, based on the queuing theory M / M / C model, the system's quality of service (QoS) is optimized. By adjusting the number of servers and service rates, a load balancing algorithm is used to optimize task allocation across servers. Queue lengths and blocking probabilities are calculated to ensure reliable service quality in dynamic environments. Finally, under the condition of dynamically changing task loads, an improved enhanced marine predator algorithm (EMPA) is used to generate an optimal solution. This algorithm enhances global search capabilities by introducing a fitness function and a dynamic update mechanism. It also incorporates an adaptive factor to adjust the predator's movement pattern, optimizing task offloading decisions and resource scheduling. The adaptive factor dynamically adjusts task allocation and resource scheduling strategies based on system load and network changes, ensuring stable system operation and efficient task completion, while balancing global search and local optimization to improve user experience and reduce total system costs.

[0113] All or part of the steps in the various methods of the above embodiments can be completed by controlling related hardware through a program. The program can be stored in a readable storage medium, and the storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electronically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0114] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A hybrid cloud resource scheduling method, characterized in that: include: Task Scheduling: In hybrid cloud resource scheduling tasks, a task scheduling model is built based on the Markov decision process to generate task scheduling decisions. The Q-value function of the Markov decision process is iteratively updated using the Bellman equation. Optimizing QoS: The reward function of the Markov decision process is defined as: R(s,a)=Cost(s,a)+QoS(s,a) Where R(s,a) represents the reward function of the Markov decision process, Cost(s,a) represents the total cost under the current state and action, and QoS(s,a) represents the quality of service score under the current state and action. The QoS(s,a) is optimized based on the M / M / C model of queuing theory. Resource allocation: Based on the task scheduling strategy, the EMPA-ASA model is constructed for the resource allocation problem of each task in the scheduling strategy to generate the optimal resource allocation solution. The calculation process of the EMPA-ASA model includes: Generate initial solutions based on the enhanced marine predator algorithm; The adaptive factor is updated based on the adaptive simulated annealing algorithm to generate new solutions for the enhanced marine predator algorithm; The calculation process of the adaptive simulated annealing algorithm includes: g k =h1·Δf+h2·(R target -R k )+h3·(ΔN target -ΔN k ) Among them, g k represents the adaptive factor, Δf is the rate of change of the objective function, h1, h2 and h3 are weight coefficients, C current is the objective function value of the current solution, is the mean of the objective function of the first m iterations; R target represents the target resource utilization, R k Indicates the resource utilization rate in round k, ΔN target Indicates the target load change rate, ΔN k represents the load change rate of the kth round; Comparing the total cost of the new solution with the total cost of the initial solution, taking the new solution as the current solution if the total cost of the new solution is less than the total cost of the initial solution, and calculating the acceptance rate of the adaptive simulated annealing algorithm for the new solution if the total cost of the new solution is greater than or equal to the total cost of the initial solution; An adaptive factor of the adaptive simulated annealing algorithm is updated according to the current solution or the acceptance rate.

2. The hybrid cloud resource scheduling method according to claim 1, characterized in that: The total cost includes: communication cost, computing cost and storage cost; the service quality score includes: response time score, bandwidth score and task completion time score.

3. The hybrid cloud resource scheduling method according to claim 1, characterized in that: Also includes: The process of adaptive temperature update is as follows: T k+1 =T k ×(1+β·Δf+γ·Δa+δ·Δn) Among them, Δf, Δa, and Δn represent the rate of change of the objective function, the difference in acceptance rate, and the amount of neighborhood change, respectively, and β, γ, and δ are weight coefficients.

4. The hybrid cloud resource scheduling method according to claim 1, characterized in that: The calculation process of the enhanced marine predator algorithm includes: Generate new solutions through Levy flight strategy to expand the search range of solutions and generate new solutions with stronger global search capabilities. assignments The generation formula is as follows: Among them, current assignments is the current solution, which represents the initial solution in the initialization phase, λ is the predation behavior offset weight, Levy (μ) is the random value generated by the Levy flight strategy, u and v are normal distribution random numbers, and μ is the scaling parameter of the Levy distribution. is a scaling factor that controls the jump step size.

5. The hybrid cloud resource scheduling method according to claim 1, characterized in that: The step of optimizing the QoS (s, a) based on the queuing theory M / M / C model specifically includes: By adjusting the number of servers and service rates, a load balancing algorithm is used to optimize the task allocation of each server, and the queue length and blocking probability are calculated to control the queue length and arrival rate of each controller.

6. The hybrid cloud resource scheduling method according to claim 1, characterized in that: In the Markov decision process, the transition of the system state follows the state transition probability distribution P(s′|s,a): P(s′|s,a)=Pr(S t+1 =s′|S t =s,A t =a) Among them, P(s′|s,a) represents the probability of the system transitioning to the next state s′ given the current state s and action a, S t represents the system state at time t, A t represents the action taken at time t.

7. A hybrid cloud resource scheduling system, characterized in that: include: a memory for storing program instructions; A processor, configured to call the program instructions stored in the memory to implement the hybrid cloud resource scheduling method according to any one of claims 1 to 6.

Images