A resource load balancing scheduling method based on optimal control output regulation

Abstract

The application discloses a resource load balancing scheduling method based on output regulation of optimal control, and belongs to the technical field of mobile edge computing resource scheduling and task offloading. Resource states such as CPU, memory, task queue and network bandwidth of an edge server are acquired, a system state vector is constructed, and a resource load output vector is obtained through an output matrix; an output regulation error and a performance index are constructed according to the resource load output and a target output, an error weighted term is converted into an equivalent state operation weighted form, a feedback law is obtained by solving an output regulation type Riccati equation; task offloading or task migration scheduling instructions are generated according to the feedback law, and the system state is fed back and updated in the next sampling period to form a closed-loop load balancing scheduling. The method can reduce resource load output deviation, relieve server overload and queue backlog, and improve closed-loop scheduling stability and resource utilization efficiency.

Description

Technical Field

[0001] This invention belongs to the field of mobile edge computing and is a method for edge computing resource scheduling and task offloading, applicable to distributed computing environments containing multiple heterogeneous edge servers. It is particularly suitable for server load balancing in low-latency scenarios such as industrial IoT, autonomous driving assisted computing, and augmented reality. Background Technology

[0002] With the commercial deployment of 5G mobile communication technology and the explosive growth in the number of IoT terminals, the amount of data generated at the network edge is growing exponentially. Traditional cloud computing models require all data to be transmitted to remote data centers for processing, which not only consumes a large amount of core network bandwidth but also results in end-to-end latency of tens to hundreds of milliseconds, making it difficult to meet the stringent low-latency requirements of emerging applications such as autonomous driving, industrial real-time control, and augmented reality. Mobile edge computing, by pushing computing, storage, and network resources to the network edge, allows computing tasks to be processed closer to the data source, thereby reducing end-to-end latency and core network traffic.

[0003] However, unlike cloud data centers, the computing power, storage capacity, and network bandwidth of a single edge server are relatively limited. In actual operation, due to user mobility and the spatiotemporal imbalance of business requests, some edge servers may become overloaded due to a sudden surge in computing requests in a local area, while other nearby servers are under low load or even idle. This load imbalance leads to a decline in service quality for some users and results in low overall system resource utilization. Therefore, how to achieve efficient load balancing and scheduling among multiple edge servers is one of the core technical problems that urgently need to be solved in the field of edge computing.

[0004] Existing edge computing load balancing methods mainly include static threshold scheduling, heuristic optimization scheduling, and distributed negotiation scheduling. These methods typically perform task migration and resource allocation based on individual states or empirical rules such as server CPU utilization, memory utilization, task queue length, and network bandwidth utilization. While they can alleviate local overload problems in some scenarios, they lack a unified resource load output modeling mechanism, making it difficult to simultaneously transform multi-dimensional resource states into adjustment objects that can directly reflect the degree of load balancing.

[0005] In existing methods, resource utilization or state variables are often directly used as judgment criteria or optimization variables, with few instances of setting an output matrix to map multidimensional system states to resource load output vectors. Therefore, a clear correspondence between scheduling objectives and actual load balancing results is easily lacking. For example, when scheduling is based solely on the CPU threshold or queue length changes of a single server, it is difficult to simultaneously consider the comprehensive output performance of memory, bandwidth, task queues, and load distribution among servers.

[0006] Some methods have attempted to apply optimal control or feedback control concepts to resource scheduling, but their performance metrics typically focus more on state variables or control costs themselves, failing to fully reflect the closed-loop adjustment objective of resource load output deviation. When task arrival rates, network bandwidth, server computing power, or resource occupancy change dynamically, traditional static threshold methods, heuristic search methods, or single-state-variable feedback methods struggle to maintain continuous multi-server load balancing, and also find it difficult to generate scheduling instructions consistent with the load balancing objective in a timely manner based on changes in resource load output.

[0007] Therefore, a resource load balancing scheduling method for edge computing needs to be proposed. This method can construct the state vector of the edge computing system, map the system state to the resource load output vector through the output matrix, and construct the output adjustment performance index with the resource load output deviation as the direct adjustment object. On this basis, the output adjustment type Riccati equation is solved to generate task offloading scheduling instructions or task migration scheduling instructions that achieve load transfer through task migration, thereby forming a closed-loop resource load balancing scheduling mechanism oriented towards resource load output. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a resource load balancing scheduling method based on optimal control output adjustment to solve the problem of unbalanced load on edge computing servers. This invention acquires the resource status information of the edge computing server cluster and constructs an edge computing system state vector. And based on the output matrix or Map the state vector to the resource load output vector. The output regulation error is constructed based on the resource load output vector and the load balancing objective, and then the output regulation performance index is constructed around the resource load output rather than the deviation of ordinary state variables.

[0009] Furthermore, this invention transforms the output deviation weighted index into one containing... The equivalent state-weighted form of the output deviation domain is used to solve the weighted adjustment objective in the state space. Solving the output-adjustment-type Riccati equation aims to obtain the feedback gain and output adjustment feedback law for resource load balancing scheduling of edge servers. Based on this feedback law, task offloading scheduling instructions or task migration scheduling instructions that achieve load transfer through task migration are generated. Thus, resource load output mapping, output deviation domain weighting, state space solution, and scheduling execution form a continuous technology chain, enabling the optimal control solution to directly serve the closed-loop resource load balancing scheduling of the edge computing server cluster.

[0010] It should be noted that the output-adjusted Riccati solution mentioned above does not redefine the mathematical type of the Riccati equation, but rather refers to output based on resource load. and its relative to the expected output Output adjustment error To directly adjust the object, the output domain weights are used. The output deviation weighting term is formed and then... This is transformed into an equivalent weighting term in the state space; based on this equivalent weighting, the feedback gain can still be obtained by solving the Riccati equation in standard optimal control, and the continuous control quantity is then converted into an equivalent weighting term in the state space. This is further converted into the number of non-negative integer tasks on a preset directed link and executable task unloading or task migration scheduling instructions.

[0011] A resource load balancing scheduling method based on output regulation of optimal control includes the following steps: Step 1: Real-time data collection from the edge computing network The operational status data of the edge server, including the first i Each edge server at the time of data collection CPU utilization Memory usage 、 Network bandwidth utilization and task queue length A sliding window averaging filter is applied to the acquired raw data to eliminate instantaneous measurement noise. A CPU utilization threshold is set. and memory usage threshold When any server satisfies or When an overloaded server is identified, its adjacent servers on the topology graph are added to the candidate receiving server set.

[0012] Specifically, lightweight monitoring agents are deployed on each edge server, following a fixed sampling period. Collect the operating status parameters of each edge server, among which, The value ranges from 0.5 seconds to 5 seconds and is configured according to network size and control precision requirements. The collected raw data undergoes a sliding window averaging filter, with a filter window width of 3 to 10 sampling periods to eliminate the interference of instantaneous jitter on scheduling decisions. The filtered CPU utilization is recorded as... The filtered CPU utilization rate is compared with a preset threshold, which is set based on the statistical mean and standard deviation of historical operating data. ,in, The value range is from 1.5 to 2.5. When or When an overloaded server is identified, it is marked as such, and its neighboring servers with a hop count of 1 in the edge computing network topology are added to the candidate receiving server set. Simultaneously, the server with the lowest CPU utilization is selected. One server is added to the candidate receiving server set, among which... , This represents the total number of servers in the edge computing network.

[0013] Step 2, establish the linear time-varying dynamic equations of the edge computing system:

[0014] in, for A state vector, or deviation form, of a 3D edge computing system is composed of various state components required to construct the resource load output deviation. For systems containing... For a system with multiple servers, the CPU utilization deviation, memory usage deviation, and task queue length deviation of each server can be selected as basic state components. When network bandwidth usage is included in the output adjustment target, the network bandwidth usage deviation corresponding to the link or server is further added to the state vector, or a measurable bandwidth usage deviation component is added to the extended state vector. for The dimensional scheduling control input vector, where each component corresponds to a pre-defined directed task offloading link or task migration link in the control link set. i → j And indicate that along this directed link from the server i to server j The amount of task number to be adjusted for unloading, or the amount of task number to be adjusted for load balancing through task migration; if i → j and j → i If both are allowed, then they are encoded as two different directed links and respectively correspond to... Different control components in; It equals the total number of directed links in the edge computing network that allow task offloading or task migration. for 3D system state matrix, for Dimensional control input matrix.

[0015] Based on the above state equation, this invention further defines the resource load output equation:

[0016] in, for The resource load output vector represents the resource load output that needs to be adjusted, obtained from the system state mapping, rather than the output adjustment error itself. The resource load output may include CPU utilization output, memory utilization output, task queue length output, network bandwidth usage output, and output related to server load balancing. for A dimensional output matrix is ​​used to map the system state to a resource load output vector. The output adjustment error is obtained by the difference between the resource load output vector and the resource load balancing target output, and is used to characterize the deviation of resource load outputs such as CPU utilization, memory utilization, task queue length, network bandwidth utilization, and server load balancing degree from the balancing target. After converting the resource load balancing target output to a zero target according to the equivalent coordinate translation step described later, the output adjustment error... This indicates that the resource load has reached the preset balancing target; when system parameters change slowly or remain approximately constant within the sampling period. It can be taken as a constant matrix ,Right now .

[0017] Furthermore, the output matrix This is used to characterize the resource load output mapping relationship between the system state vector and the resource load output vector. System state vector This can include status information such as CPU utilization, memory utilization, task queue length, network bandwidth utilization, task arrival rate, service processing capacity, and remaining resources of multiple edge servers, or the deviation of the above status information from the load balancing target. Resource load output vector This indicates the resource load output that needs to be directly adjusted by the output regulator. It may include server load balancing deviation, resource utilization output, queue backlog output, bandwidth usage output, or a weighted combination output formed by multiple resource indicators.

[0018] Output matrix or It can be constructed based on the correspondence between the selected resource indicators and state variables. When it is necessary to directly select the state vector... When some state variables in the data are output as resource load, or It can be configured as a selection matrix; for example, the corresponding state components such as CPU utilization deviation, task queue length deviation, or bandwidth occupancy deviation can be directly used as output components. In this case, the positions of the selected state components in the output matrix can be set to non-zero elements, and the remaining positions can be set to zero, so as to achieve direct selection from the state vector to the resource load output vector. This construction transforms the scheduling object from a single resource threshold judgment to a unified adjustment of the selected resource load output and its deviation from the relative balance target.

[0019] When it is necessary to combine multiple resource status variables to form a load-balanced output or It can be configured as a weighted mapping matrix. The elements of the weighted mapping matrix can be set according to factors such as CPU resource importance, memory resource importance, task queue backlog, link bandwidth usage, task priority, service latency requirements, or server heterogeneity capabilities, so that an output component can be obtained by combining multiple state components according to preset weights. For example, the server comprehensive load output can be formed by combining CPU utilization deviation, memory utilization deviation, and task queue length deviation according to weights; the link load output can be formed by combining bandwidth usage deviation and task migration traffic status.

[0020] When the importance of various resource indicators changes under different business types, task priorities, or scheduling cycles... It can be configured as a time-varying output matrix and dynamically updated based on changes in business priority, task type, service quality requirements, or resource consumption; when the resource indicator weights remain unchanged within a scheduling cycle or fixed business scenario, It can be taken as a constant output matrix Regardless of the time-varying output matrix used. Still a constant output matrix Their dimensions should all be the same as those of the state vector. and resource load output vector Matching, that is, when for dimensional vector, for When using a 3D vector, or for The choice of a matrix, a weighted mapping matrix, or a time-varying output matrix enables the scheduler to directly select resource indicators, comprehensively represent multiple resource states, and update output mapping relationships under different business scenarios, respectively. The output matrix differs from the control input matrix. and output adjustment feedback gain matrix Its function is to determine the resource load output mapping and provide a basis for constructing subsequent output regulation performance indicators and output regulation errors, so that subsequent control solutions always revolve around the selected or combined resource load output.

[0021] The control objective of this invention is to adjust the resource load output deviation, rather than to track the output of a given time trajectory. Before constructing the output adjustment performance index, the state and output are translated using the resource load balancing target output as a reference; this equivalence processing transforms the resource load balancing target into a zero target, and the state vector is taken as a deviation form of the state vector relative to the balanced state, i.e.: ,

[0022] in, This represents the target output for resource load balancing; following the equivalent coordinate translation steps described above, the target output is converted into a zero vector, and the output adjustment error after translation is used as the subsequent output adjustment object. This represents the adjustment error of the resource load output relative to the balancing target. Within the output adjustment framework of this invention, the performance index is always directly related to... Perform quadratic weighting; because , and Numerically equal, but in terms of terminology, the former still represents the output regulation error and the latter still represents the resource load output vector; it does not require the introduction of an external given output that varies with time, nor does it formulate the problem as a typical output tracking control problem.

[0023] The system matrix adopts the following block diagonal structure: Where blkdiag represents the block diagonal operator composed of the sub-state matrices of each server, and each diagonal sub-block Description server i The state self-evolution characteristic without external control input; if the bandwidth occupancy deviation is added to the state vector, a state sub-block describing the bandwidth occupancy recovery or link load release process is added accordingly.

[0024] When considering three basic states—CPU utilization deviation, memory usage deviation, and task queue length deviation—the following can be taken: When bandwidth usage deviation is included, it can be expanded to .

[0025] in, , , Servers i CPU processing time constant, memory release time constant, and task processing time constant; This is the bandwidth occupancy recovery time constant. This represents the time scale by which the CPU utilization deviation decreases as the server processes tasks in the queue when no new tasks arrive. This indicates the time scale at which the server releases memory occupied by processed tasks; This represents the decay timescale of the task queue under the server's processing speed. This represents the time scale required for link occupancy to return to a balanced level without any new offloading traffic. These parameters can be obtained online by identifying the least squares parameters from historical operational data.

[0026] Control Matrix elements Description from server i to server j When unloading a unit of computing task, or achieving load balancing by migrating that unit of computing task, the server state components are affected. p The resulting instantaneous impact. The mechanism of this impact is as follows: for the source server... i Unloading or migrating a task reduces its CPU usage and queue length, corresponding to... For the target server j Receiving a task increases its CPU utilization and queue length, corresponding to... For state components unaffected by the task transition path, take... The ratio of increase to decrease can be determined by the heterogeneous computing power ratio of the two servers. It means that among them and Servers i and server j CPU computing power This is a conversion factor for computing power. Each column corresponds to the impact of a pre-defined unit task unloading or unit task migration on the state vector on the directed link; if i → j and j → i If both are allowed, then they correspond to control vectors respectively. Different control components and Different columns, not through i → j Negating the link control component enables reverse task transfer.

[0027] Step 3: Construct a linear quadratic performance index functional for output regulation as a quantitative criterion for evaluating the performance of the scheduling strategy. This performance index is weighted by the output regulation error; after the equivalent coordinate translation described above, the deviation-form state vector and the output regulation error satisfy... The performance metrics are still expressed as a quadratic form of the output regulation error. Zero output corresponds to a resource load balancing state, rather than requiring the original resource utilization to approach zero. The performance metric functional takes the following form:

[0028] The first term corresponds to the terminal output adjustment error cost term, which is used to measure the optimization of the time-domain terminal moment. The degree of penalty for terminal output adjustment error; for 3D positive semidefinite symmetric terminal output weighting matrix; the integral term in The corresponding process output adjustment error cost term is used to measure the cumulative magnitude of output adjustment errors such as CPU utilization deviation, memory utilization deviation, task queue length deviation, network bandwidth usage deviation, and load balancing deviation throughout the entire control process. for 3D positive semidefinite symmetric output weighting matrix; The energy cost term is used to measure the cost of task unloading or task migration scheduling control. for A positive definite symmetric control weighting matrix. The integration time interval is... , The optimized time domain length is fixed, typically set to 5 to 30 seconds.

[0029] After completing the equivalent coordinate translation described above, the output adjustment error mapping relationship is used. The above output regulation performance indicators can be equivalently transformed into a computational form in the state space:

[0030] make , ,but for 3D equivalent state terminal weighted matrix, for The equivalent state-run weighting matrix is ​​used. Subsequent solutions to the Riccati equation and output adjustment feedback gain are based on... and Instead of using a weighted matrix based directly on the state vector, it operates on the direct state. Therefore, It is not a simple matrix transformation independent of resource scenarios, but rather... The weighted relationships of the output deviations of resource loads such as CPU, memory, task queue, and network bandwidth are preserved in the state space solution so that the feedback gain obtained from the Riccati equation corresponds to the adjusted resource load output deviations. Accordingly, the terminal output deviation constraint is transformed into the terminal condition of the Riccati equation.

[0031] The construction of the output bias weighting matrix and the control weighting matrix in the performance index functional is one of the key innovative features that distinguishes this invention from existing technologies. This invention proposes a method for constructing an adaptive output bias weighting matrix for heterogeneous sensing: Step 3.1, output the weighting matrix Construction of diagonal elements: For servers i The diagonal elements corresponding to the CPU utilization deviation output, memory utilization deviation output, and queue length deviation output are calculated using the following formulas:

[0032]

[0033]

[0034] in, For server i CPU computing power For memory capacity, This represents the average task processing rate (tasks / second). , , The normalization coefficients ensure that the weights of various output deviations remain on the same order of magnitude across the server set. The physical meaning of this construction is as follows: servers with strong computing power and ample memory have smaller resource load output deviation weights, meaning the system allows these servers to withstand larger recoverable load fluctuations; conversely, servers with weaker resource capacity or lower task processing speeds have larger output deviation weights, causing the same magnitude of output adjustment error to incur a higher cost in performance metrics. Therefore, the output regulator prioritizes suppressing these high-risk resource output deviations.

[0035] Step 3.2, control the weighting matrix Construction of diagonal elements from the server: i To the server j The diagonal element corresponding to the task unloading or task migration control component is calculated using the following formula:

[0036] in, For server i With server j The network topology distance between them (in hops). This represents the average transmission rate of the link. This represents the current load rate of the link. and The weighting coefficient is used for the following: The physical meaning of this construction method is that the control cost of paths with long topological distances, low transmission rates and high current loads is high. The output regulation feedback law will tend to reduce or avoid using such paths for task offloading or migration, thereby suppressing unreasonable task migration on high transmission cost links and guiding the system to prioritize task offloading or migration paths with higher transmission efficiency and lighter loads. The off-diagonal elements are typically set to zero, indicating that the control energies between different task offloading or migration paths are independent. The diagonal elements of the control weighting matrix are formed by the network topology distance, the average link transmission rate, and the current link load rate, and are used to comprehensively characterize the transmission cost of the task transfer link. Communication delay is only used as an optional compensation or path cost correction factor as described in Example 2, without changing the main formula of the control weighting matrix, and no independent task migration cost variable is set.

[0037] Step 3.3, Terminal output bias weighting matrix Take it as a diagonal matrix, where the output sub-block related to each server is constructed as follows:

[0038] in, For server i The highest historical CPU utilization This is the terminal penalty amplification factor (typically ranging from 0.5 to 2.0). This is the identity matrix. The meaning of this construction is: for servers that have historically reached high load levels, a stronger regression constraint is imposed on their resource load output deviation at the end of the optimization time domain, making... The high-risk resource load output deviation that still exists at the end of the scheduling time domain results in a higher terminal cost, thereby reducing the possibility that the optimization process only improves the load at intermediate moments while still retaining a large imbalance at the end of the time domain.

[0039] In the embodiment defined in claim 10, the performance index functional in step 3 is extended to a multi-objective weighted form based on the existing output regulation performance index: .in, The target term for output adjustment error is determined by performing equivalent coordinate translation to ensure that the deviation-form state vector and the output adjustment error satisfy a mapping relationship. , or take . For system energy consumption target items, ,in For server Power consumption. For network latency target item, ,in For link transmission delay, For a moment Corresponding to the pre-defined directed link Number of tasks on the platform. Weight , and satisfy And dynamically adjust according to the service quality requirements of the application scenario. The multi-objective weighted form is used to extend the performance index construction in step 3, wherein Maintain the basic adjustment target that focuses on output adjustment error. and The system energy consumption and network latency targets are respectively characterized; this extension does not change the requirements of claim 1. , , , , , And the existing meaning of the task unloading or task migration scheduling main line.

[0040] Step 4: Solve the output regulation-type optimal control problem to obtain the output regulation feedback gain matrix. This step is used to generate the output regulation feedback law based on the output regulation performance index, and specifically includes the following sub-steps: Step 4.1, construct the Hamiltonian function:

[0041] in, for dimensional costate vector, This matrix is ​​derived from the output regulation error weighting term in the output regulation performance index. .

[0042] Step 4.2: Based on the minimum principle, the costate equation, terminal conditions, and governing equations are obtained:

[0043]

[0044]

[0045] Step 4.3, let Substituting these into the state equation and costate equation, we obtain the output-adjusted Riccati matrix differential equation and the terminal condition:

[0046]

[0047] Step 4.4, according to Calculate the output regulation feedback gain matrix and the output regulation feedback law: , .in, for The solution matrix of the Riccati equation. for The output-adjusted feedback gain matrix. The purpose of solving the Riccati equation is not merely to obtain... Instead, it is based on containing and The output regulation performance metrics are used for edge server resource load balancing scheduling. , and by Continuous scheduling control quantities are generated on each preset directed link to determine the number of tasks unloaded along the corresponding directed link, or to achieve load transfer through task migration; the scheduling direction is determined by the preset directed link corresponding to the control component. The output regulation feedback law takes the system state vector as input; in the scenario of this invention, the state quantities can be directly collected by the monitoring agent, and some states that cannot be directly collected can be estimated from the resource load output when the system meets the observability conditions.

[0048] Step 4.5: Solve the Riccati equation using inverse numerical integration. In the finite optimization time domain... Internally, from the terminal conditions We begin recursively calculating back to the initial time, obtaining the discrete time values. and It is used for online scheduling and control.

[0049] The above steps result in an output matrix. Determine the mapping relationship between resource load output and output regulation error, and weight the output deviation matrix accordingly. Weighted matrix of terminal output deviation through and Enter the solution process of the Riccati equation, and from and The output adjustment target is then passed to the scheduling control quantity. Therefore, the Riccati equation, feedback gain, and feedback law are not mathematical solutions independent of the resource scheduling scenario, but rather computational steps connecting resource load output deviation, heterogeneous weights, and scheduling instruction generation.

[0050] In other words, the Riccati equation in this invention is responsible for transforming the weighted objective of output deviation into a solution that can calculate feedback gain. Its mathematical form can follow the Riccati solution structure of linear quadratic optimal control. The difference in this method is that the source of weights, the object being adjusted, and the subsequent scheduling implementation of this solution structure are all set around the resource load output deviation and task unloading / task migration execution chain, rather than just performing ordinary state adjustment on the original state variables.

[0051] Step 5: Using the output adjustment feedback gain matrix obtained in Step 4, calculate the real-time scheduling control quantity and generate task scheduling instructions. At the beginning of each control cycle, the scheduling controller performs the following operations: Step 5.1: Collect the current operating status of all servers across the network and calculate the state vector. And by the output equation The resource load output vector is obtained, and then the output adjustment error is calculated based on the difference between the resource load output vector and the resource load balancing target output. When using coordinate translation processing, the resource load balancing target output can be expressed as... At this time, the output adjustment error is related to They are numerically equal, but in terms of terminology they still represent output regulation error and resource load output vector, respectively. Step 5.2, from pre-stored sequence sum The output adjustment feedback gain matrix corresponding to the current time step is read from the sequence, where From containing The output-regulated Riccati equation is obtained; Step 5.3: Perform matrix-vector multiplication to calculate the optimal output regulation control quantity for output regulation: ; Step 5.4: Interpret each component in the control vector as its corresponding preset directed link. i → j The continuous task offloading volume or the continuous load transfer volume achieved through task migration; the scheduling direction is determined by the directed link, if i → j and j → i If both are allowed, they are represented by different control components. For continuous control components less than zero, they are treated as 0 in non-negative integer projection or constraint correction, and no directed link task transfer is generated, nor is a reverse instruction generated through this component. The number of task offloading or task migration is corrected by combining the number of tasks to be processed on the source server, the available computing resources, available storage resources, queue receiving capacity, and the link transmission conditions characterized by the average link transmission rate, the current link load rate, and the available bandwidth.

[0052] Due to the output regulation feedback law For each preset directed link, there are continuous control components, and task unloading or task migration is performed according to the number of tasks. In step 5.4, the continuous control components have been connected to the preset directed links. After corresponding, less than zero continuous control components are processed to 0 or no directed link task transfer is generated, global non-negative integer projection and rounding are performed on the continuous control components of each directed link. Let... Assign vectors to tasks. To preset the number of directed links, Each component is a non-negative integer, and the control vector is continuous. For the projected quantity, we obtain . The corresponding pre-defined directed link The component is denoted as This represents the number of tasks unloaded or migrated along the directed link, but does not represent the number of signed tasks. The projection result satisfies the supply-demand balance constraint. ,in For server The number of excess tasks; the summation symbol Indicates the relationship with the source server The corresponding target server index on the preset directed link. (Projected) The number of non-negative integer tasks to be used as the initial number for subsequent hard constraint checks; if Then take ,in For server The number of tasks to be processed; if the integer number of tasks exceeds the queue capacity of the target server, then the number is truncated according to the queue capacity; if the receiving server The available computing or storage resources are insufficient to process For each additional task, the reduction is based on the available resources, of which The amount of available resources, The amount of resources required to process a single task.

[0053] Step 5.5 performs a constraint check on the non-negative integer task unloading or migration counts after projection and rounding: If the total number of task migrations corresponding to a source server exceeds its current number of pending tasks, the count is reduced proportionally; the target server's queue receiving capacity is determined by the difference between the maximum task queue capacity and the current queue length; if the rounded number of tasks exceeds the queue's receiving capacity, the count is truncated according to the queue's receiving capacity; if the target server's available computing or storage resources are insufficient to handle new tasks, the count is further reduced proportionally to available resources; if the link transmission conditions, characterized by the average link transmission rate, current link load rate, and available bandwidth, are insufficient, the number of tasks is further reduced or the corresponding task migration is not executed. The remaining tasks remain on the source server awaiting subsequent scheduling. By checking the number of tasks on the source server, the resource capacity of the target server, the queue receiving capacity of the target server, and the link transmission conditions, it is possible to avoid generating unexecutable task unloading or migration instructions that exceed the source server's task volume, the target server's resource capacity, the queue receiving capacity of the target server, or the link transmission conditions.

[0054] Step 5.6: Generate a machine-executable task unloading or task migration instruction tuple: (Source Server ID, Target Server ID, Number of Unloading or Migrating Tasks, Task Priority, Timestamp), where the Source Server ID and Target Server ID are determined by the corresponding preset directed link, and the Number of Unloading or Migrating Tasks field is a non-negative integer representing the number of tasks. The instruction tuple does not contain a negative task quantity field, nor does it have a negative component commutation mechanism, and is sent to the data plane of the edge computing network for execution.

[0055] Step 6: Perform task unloading or task migration and status update to form a closed-loop feedback.

[0056] Step 6.1: The source server sends the specified computation task (including task data and execution context) to the target server via the network link. For stateful tasks that require continuous execution, a snapshot of the task context is synchronously transmitted to ensure correct resumption of execution on the target server.

[0057] Step 6.2: After receiving the unloading or migration task, the target server inserts the task into its local task queue according to priority. The default scheduling strategy is first-come, first-served; for real-time tasks with a clear deadline, the earliest deadline priority strategy can be used.

[0058] Step 6.3: When the next status collection cycle arrives, repeat Step 1 to collect the updated status data of each server. If the CPU utilization and memory usage of all servers have fallen back below their respective thresholds at this time, then this round of scheduling ends, and the system continues at the normal cycle. Perform status monitoring; if any server is still overloaded (possibly due to new tasks arriving or insufficient unloading or migration of tasks from the previous round), return to steps 2-5 for the next round of optimization scheduling.

[0059] Furthermore, this invention employs a rolling time-domain optimization strategy to address the long-term dynamic changes in the edge computing environment. Its working principle is as follows: an infinitely long scheduling timeline is divided into a series of continuous finite optimization time-domain windows, where the length of each window is fixed. In the... k Within each window, using the current actual state as the initial condition, and combining it with the current output matrix... and output weighted matrix The output regulation optimal control problem is solved to obtain the output regulation control sequence for this window, but only the output regulation control quantity of the first control cycle in the sequence is actually executed. When the next window starts, the actual measured state and resource load output deviation after the previous window are used as new inputs to resolve the output regulation problem. This "optimization-execution-reoptimization" rolling mechanism enables the controller to continuously use the latest resource load imbalance output information for decision updates, effectively dealing with the randomness of task arrival and the slow time-varying characteristics of system parameters.

[0060] Furthermore, when the edge computing system satisfies the following condition: (a) System matrix and (b) Output matrix (not changing with time) Output weighted matrix and control weighting matrix (c) As the optimization time domain approaches infinity, the steady-state solution of the output-regulated Riccati differential equation (17) degenerates into the algebraic Riccati equation:

[0061] This is a nonlinear matrix algebraic equation. The output adjustment feedback gain matrix is ​​now... Since it is a constant matrix, the output regulation feedback law simplifies to: .in The equivalent state operation weighting matrix in steady-state output regulation The system meets the requirements of stability, detectability, and... semi-positive definite Under positive definite conditions, this closed-loop resource load balancing scheduling structure helps improve the stability of the resource load output deviation adjustment process, so that the load of each server gradually tends to the desired balanced state.

[0062] The beneficial effects of this invention are: (1) Multiple resource states and scheduling objectives form a direct correspondence: This invention constructs an output matrix based on resource indicators such as CPU, memory, task queue and network bandwidth. or Mapping multidimensional resource states to resource load output vectors This allows the adjusted objects to directly correspond to resource load performance such as CPU utilization, memory utilization, task queue length, network bandwidth usage, and server load balancing. The selection matrix can directly select the resources to be adjusted, the weighted mapping matrix can comprehensively represent the states of multiple types of resources, and the time-varying output matrix can adjust the output mapping relationship according to the business scenario. Thus, the scheduling objective is transformed from judging a single resource threshold to the unified adjustment of the load output of multiple resources and its deviation from the balance target.

[0063] (2) Server heterogeneity and link transmission cost are incorporated into the adjustment target: output weighting matrix The control weighting matrix is ​​constructed based on server computing power, memory capacity, task processing rate, historical load, or current resource status, giving higher adjustment weights to resources with weaker resource carrying capacity or higher deviation risk; Based on network topology distance, average link transmission rate, and current link load rate, the scheduling control quantity corresponding to high transmission cost links is more strongly constrained; communication latency can be used as an optional latency compensation or path cost correction factor in specific implementations. Terminal output weighting matrix. This is used to constrain the resource load output deviation at the end of the scheduling time domain. The three factors work together to ensure that the scheduling process simultaneously considers resource output deviation, server heterogeneity capabilities, and link transmission costs.

[0064] (3) The state-space solution results correspond to the resource load output adjustment target: This invention achieves this through... The weighted adjustment objective in the output deviation domain is transformed into an equivalent operational weighted objective in the state space, and then utilizes the... Solving the output-modulated Riccati equation This leads to the output adjustment feedback gain used for edge server resource load balancing scheduling. This enables the output regulation feedback law. The weighted index is derived from the output adjustment error, rather than the direct state operation weighted index which is unrelated to the resource scenario; the continuous control component of the feedback law is used to form task unloading scheduling instructions or task migration scheduling instructions, so that the Riccati solution can be passed to the actual scheduling execution.

[0065] (4) Continuous control quantities can be converted into executable scheduling instructions: The continuous control quantities generated by the output regulation feedback law are rounded and combined with the number of tasks to be processed on the source server, the available computing resources and storage resources of the target server, the receiving capacity of the target server queue, and the link transmission conditions characterized by the average transmission rate of the link, the current load rate of the link, and the available bandwidth for constraint correction. The number of executable task offloading can be determined on the preset directed link, or the amount of load transfer achieved through task migration can be determined. The scheduling direction is determined by the corresponding directed link, so as to avoid the scheduling amount exceeding the task quantity of the source server, the resource capacity of the target server, the receiving capacity of the target server queue, or the link transmission conditions, thereby improving the feasibility of scheduling actions.

[0066] (5) Closed-loop feedback helps to continuously correct resource load distribution: After the scheduling command is executed, the server status is re-acquired and the resource load output and output adjustment error are updated, so that subsequent control cycles can continue to adjust task unloading or task migration based on the actual execution results. Under the conditions that the system meets the requirements of stability, detectability, and positive semi-definite output weighting matrix and positive definite control weighting matrix, this closed-loop structure helps to reduce resource load output deviation and improve the comprehensive utilization efficiency of edge computing system for resources such as CPU, memory, queue and bandwidth. When adding server nodes or adding resource load output dimensions, the output matrix and corresponding weighting matrix can be extended to continue to adapt to the scheduling scenario. Attached Figure Description

[0067] Figure 1 This is the overall flowchart of a resource load balancing scheduling method based on output adjustment of optimal control.

[0068] Figure 2 This is a schematic diagram of the edge computing network topology and task offloading.

[0069] Figure 3 This is a block diagram of a closed-loop resource load balancing scheduling system based on optimal control output regulation.

[0070] Figure 4 This is a flowchart of the inverse numerical integration solution for the output-adjusted Riccati equation.

[0071] Figure 5 This is a flowchart of the task unloading and resource scheduling execution process.

[0072] Figure 6 This is a timing diagram of the rolling time-domain optimization strategy.

[0073] Figure 7 This is a schematic diagram of a resource load balancing scheduling method based on output adjustment of optimal control. Detailed Implementation

[0074] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. All equivalent implementations or modifications made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0075] Example 1: This embodiment uses an edge computing network comprising multiple heterogeneous edge servers as an example to illustrate the resource load balancing scheduling process based on output adjustment of the present invention. Each edge server is deployed at different access points or small base station locations. Task migration between servers occurs via wired or wireless links. The scheduling controller generates task offloading scheduling instructions based on changes in resource load output, or generates task migration scheduling instructions that achieve load transfer through task migration.

[0076] Step 1: Resource Status Information Collection and Preprocessing. Monitoring agents deployed on each edge server periodically collect resource status information such as CPU utilization, memory utilization, task queue length, network bandwidth utilization, task arrival rate, server processing capacity, and remaining resources. The collected data is then processed using sliding window filtering to reduce the impact of instantaneous measurement jitter on output adjustment and scheduling decisions.

[0077] Step 2: Construct the system state vector and resource load output vector. Construct the edge computing system state vector based on the collected data. The It can consist of multiple servers, various resource states, or their deviations from the equilibrium target. An output matrix is ​​constructed based on the adjusted resource load metrics. or , and by or Obtain the resource load output vector .in, This represents the adjusted resource load output, not the ordinary state vector itself; when When the deviation from the equilibrium state has been used in the construction, It still represents the resource load output vector, and the output adjustment error is obtained from the difference between this output vector and the resource load balancing target output.

[0078] Step 3: Construct output regulation error and output regulation performance indicators. Define output regulation error. ,in This represents the target output for load balancing; before constructing performance metrics, an equivalent coordinate translation is performed based on the target output for load balancing, and... The output adjustment error after translation is used as the subsequent output adjustment object; the output adjustment error after translation and Numerically equal, but the resource load output vector is still based on... This indicates that an output regulation performance index is constructed based on the output regulation error and the scheduling control input. It is used to suppress resource load output deviation and scheduling control costs.

[0079] Step 4: Form an equivalent state-weighted operating equation and solve the output-adjusted Riccati equation. After completing the above equivalent coordinate translation, according to... or The output adjustment error weighted index is converted into an equivalent state operation weighted form. When a terminal output weighting term exists, further construction is performed. Then solve for the contents or The output-regulated Riccati equation is obtained, and the solution matrix of the Riccati equation is obtained. This step uses the output-regulated Riccati equations, rather than the uninterpreted direct-state weighted Riccati equations.

[0080] Step 5: Generate the output regulation feedback law and scheduling control quantity. Based on... Calculate the output regulation feedback gain obtained based on the output regulation performance index. And form an output regulation feedback law. ,or Although this feedback law is based on the system state vector As input, but its output regulation feedback gain is derived from the output regulation performance index and includes a weighted transformation obtained from the output regulation error. The Riccati equation.

[0081] Step 6: Generate task scheduling instructions and provide feedback updates. The scheduling controller, based on the output adjustment control quantity obtained from the output adjustment feedback law, generates scheduling instructions such as the number of tasks to be unloaded on a preset directed link, or the number of tasks to be migrated to achieve load transfer through task migration. The source server and target server of the scheduling instructions are determined by the corresponding preset directed link, and the number of tasks is a non-negative integer. The source server unloads or migrates tasks to the target server according to the scheduling instructions, and the target server adds the tasks to its local task queue upon receiving them. In the next sampling period, status data is re-acquired, and the resource load output vector and output adjustment error are calculated. If load imbalance still exists, the next round of output adjustment scheduling is entered, thus forming a closed-loop resource load balancing scheduling.

[0082] Example 2 (Scenario with communication latency): When communication delays exist in the task unloading or task migration link, a delay compensation mechanism can be introduced based on Example 1. The controller determines the average network delay based on the timestamp in the monitoring data or the task unloading / migration feedback, and uses this delay to predict and correct the state vector or resource load output vector, thereby reducing the impact of communication delay on the control effect. The average network delay is... ,in To fix the propagation delay, This represents the average amount of data per task. For server i With server j Average transmission rate of the link between them This is the estimated link queuing delay. The communication delay can be used for communication delay compensation, task transfer path cost correction, or state / output estimation correction; as an optional correction factor, it does not change the main formula of the control weighting matrix described in step 3.2, nor does it constitute an independent control action. Based on the corrected state vector... and resource load output vector After calculating the output regulation error, the process continues with the steps of constructing the output regulation performance index, solving the output regulation type Riccati equation, and generating the feedback law.

[0083] Example 3 (Approximate Steady-State Scenario): When the operating state of an edge computing system changes slowly over a period of time, and the system matrix, output matrix, and weighting matrix can be approximated as constant matrices, the algebraic Riccati equation can be used to solve for the output regulation feedback gain in the approximate steady state. This approach reduces the number of repeated solutions to differential equations, facilitates faster solution of the output regulation model, and reduces computational overhead during long-term operation, without altering the core technology of regulating resource load output deviation.

[0084] The above embodiments show that the present invention can be implemented according to the following steps: "resource status acquisition, state vector construction, output matrix mapping, output adjustment error construction, output adjustment performance index construction, including the weighted transformation obtained from the output adjustment error". The sequential implementation of "solving the Riccati equation, generating the output regulation feedback law, executing resource load balancing scheduling, and updating feedback" ensures that scheduling instructions correspond to changes in resource load output, which helps reduce resource load output deviations among multiple servers and improves load balancing.

[0085] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A resource load balancing scheduling method based on output regulation of optimal control, characterized in that, Includes the following steps: Step 1: Collect the resource status of multiple edge servers in the edge computing network in real time. The resource status includes at least one of CPU utilization, memory usage, task queue length, and network bandwidth utilization. Identify the server to be adjusted or the overloaded server according to the preset resource threshold or resource load output status, and determine the triggering time and candidate scheduling objects for output adjustment resource load balancing scheduling. Step 2: Construct the state vector of the edge computing system based on the resource status. Establish linear time-varying dynamic equations And through the output matrix Establish resource load output mapping ;in, Output vector for the resource load to be adjusted. The input vector is the scheduling control vector; the output or desired output is based on the resource load balancing target. Calculate output adjustment error For non-zero expected output, Equivalent coordinate translation is performed based on the benchmark, and the subsequent performance indicators and control solutions are unified into the form of output deviation adjustment. Step 3: Construct the output-adjustable linear quadratic performance index functional. ;in, To output the bias domain weight matrix, The weight matrix represents the control cost of task unloading or task migration. The terminal output deviation weighting matrix, To optimize the time domain; by Obtain the equivalent state weighting matrix and This allows the subsequent solution of the Riccati equation to be based on a weighted objective of the output adjustment error; Step 4: Solve the output-adjusted Riccati matrix differential equation. Boundary conditions are The Riccati matrix is ​​obtained. And calculate the output adjustment feedback gain matrix. ; Step 5: Calculate the continuous scheduling control quantity based on the output adjustment feedback gain matrix. Each control component in the control vector corresponds to a pre-defined directed task unloading link or task migration link i→j in the control link set; project the continuous control components on each directed link and round them to the nearest non-negative integer number of tasks. The This indicates the number of tasks unloaded along the preset directed link i→j, or the amount of load transfer achieved through task migration; it is categorized according to the number of tasks pending on the source server, the available computing resources and storage resources on the target server, the target server's queue receiving capacity, and the link transmission conditions. Perform constraint corrections and generate task unloading scheduling instructions or task migration scheduling instructions; Step 6: Execute the task unloading scheduling instruction or task migration scheduling instruction to send the corresponding task to the target server for execution; after the scheduling cycle or optimization time domain ends, re-collect the status and resource load output of each server, and enter the next round of optimization scheduling or continue status monitoring based on the feedback results.

2. The resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, The specific process of step 1 includes: Step 1.1, with a fixed sampling period Operational status parameters are collected by lightweight monitoring agents deployed on various edge servers. The value ranges from 0.5 seconds to 5 seconds, and can be configured according to the network size and control precision requirements. Step 1.2: Perform sliding window averaging filtering on the collected raw data. The filter window width is set to 3 to 10 sampling periods to eliminate the interference of instantaneous jitter on scheduling decisions. The filtered CPU utilization is recorded as... ; Step 1.3 compares the filtered CPU utilization rate with a preset threshold. The CPU utilization rate threshold is set based on the statistical mean and standard deviation of historical operating data. ,in The value range is from 1.5 to 2.5; Step 1.4, if or If the server is overloaded, it will be marked as an overloaded server, and its neighboring servers with a hop count of 1 in the edge computing network topology will be added to the candidate receiving set; at the same time, the server with the lowest CPU utilization will be added to the candidate receiving set. Several servers were also added to the candidate receiving set, among which , This represents the total number of servers in the network.

3. The resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, In step 2, the system dynamic equations The specific construction method is as follows: Step 2.1, define the deviation component in the state vector as... , , ,in , and Servers The CPU utilization balance value, memory usage balance value, and queue length balance value; can be taken as follows: , , When bandwidth usage is included in the output conditioning target, further definition is required. ; Step 2.2, Construct the system matrix Each diagonal sub-block Description server State decay characteristics without external control; when including three types of states: CPU, memory, and queue length, the following can be considered: , , and , Servers The CPU processing time constant, memory release time constant, and task processing time constant are obtained through online parameter identification. Step 2.3, Construct the control matrix , elements Indicates from server to server When unloading a unit task, or transferring load by migrating that unit task, the state components are... The instantaneous effect produced: if Belongs to the source server ,but This indicates that the source server is being less overloaded; if Belongs to the target server ,but This indicates that the target server is under increased load. Other unrelated state components are taken The increase in burden can be determined by... express, and These represent the CPU computing power of server i and server j, respectively. This represents the conversion factor for transforming the computing load of a single source server into the computing load of the target server. Each column corresponds to a predefined directed link in the control link set. If both i→j and j→i are allowed, then they correspond to control vectors respectively. Different control components and Different columns.

4. The resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, In step 3, the performance index functional The weighted matrix is ​​constructed using a heterogeneous sensing adaptive method: Step 3.1, output the weighting matrix The diagonal elements are configured differently based on the server's computing and storage resources: for CPU utilization deviations, the corresponding elements are output. Output the corresponding element for the memory usage deviation. Output the corresponding element for the queue length deviation. ,in For CPU computing power, For memory capacity, This represents the average task processing rate (tasks / second). , , The normalization coefficient ensures that the weights of various output biases remain on the same order of magnitude. Step 3.2, control the weighting matrix The diagonal elements are constructed based on the network link transmission cost: ,in For server With server The network topology distance (hop count) between them. This represents the average transmission rate of the link. This represents the current load rate of the link. and These are the weighting coefficients; Step 3.3, the terminal outputs the weighting matrix. Take it as a diagonal matrix, where the output sub-blocks related to each server can be taken as... , For server The highest historical CPU utilization This is the terminal penalty amplification factor. It is the identity matrix, and is composed of Transform into the terminal conditions of the Riccati equation.

5. The resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, Step 4, the derivation and solution process of the Riccati differential equation includes: Step 4.1: Construct the Hamiltonian function from the output regulation performance index functional and the system dynamic equations: ,in for dimensional costate vector, ; Step 4.2, according to Pontryagin's minimum principle, the costate equation and boundary conditions are as follows: , ;Depend on Obtain the governing equations ,because Since it is positive definite and symmetric, it is invertible, thus yielding optimal control. ; Step 4.3, will Substituting into the canonical equation yields the two-point boundary value problem: , , ; Assumption For any Establish, substitute and utilize The arbitrariness of the output leads to the output-adjusted Riccati matrix differential equation. , ; Step 4.4, solve the Riccati equation using inverse numerical integration: with a step size interval Discrete, from the terminal time The result is obtained by recursively calculating towards the initial time 0. , then calculate ,in For the first Each discrete time point.

6. The resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, In step 5, the process of converting the output regulation feedback law into task scheduling instructions includes: Step 5.1, in each control cycle At the initial moment, collect the current state of the entire network and calculate the state vector or deviation form of the state vector. From the output equation Obtain the resource load output vector and calculate the output adjustment error of the resource load output vector relative to the resource load balancing target output; then, compare the current state vector with the pre-stored output adjustment feedback gain matrix. Multiply to obtain the control vector ; Step 5.2, transfer the control vector The continuous control components on each preset directed link are projected and rounded to a non-negative integer number of tasks for task unloading or task migration. ,in This represents the number of tasks executed along the preset directed link i→j, not the number of signed tasks; continuous control components less than zero are treated as 0 or no task transition is generated for that directed link in the non-negative integer projection; a global rounding strategy with small rounding error is adopted: assuming the non-negative integer task allocation vector satisfies At the same time, it satisfies the supply and demand balance constraint. ,in Let be the number of excess tasks for server i. The number of tasks in each directed link is a non-negative integer, where, The number of directed links is preset; Step 5.3, perform a hard constraint check on the number of non-negative integer tasks after rounding: if ,in For server The number of tasks to be processed is then taken as... If the number of integer tasks exceeds the queue's receiving capacity, then the number is truncated according to the queue's receiving capacity; if the receiving server... The available computing or storage resources are insufficient to process For each additional task, the reduction is based on the available resources, of which The amount of available resources, The amount of resources required to process a single task; Step 5.4: Generate a task unloading scheduling instruction or a task migration scheduling instruction based on the number of non-negative integer tasks after constraint correction. The scheduling instruction is represented by an instruction tuple, which includes at least the source server ID, the target server ID, and the number of non-negative integer tasks on the corresponding preset directed link, and may further include task priority and timestamp.

7. The resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, The method employs a rolling time-domain optimization strategy: dividing the infinitely long scheduling time into continuous finite optimization time-domain windows. ,in To optimize window numbering, The prediction time domain length is fixed; the output regulation type optimal control problem is solved in each window to obtain the output regulation control sequence of the current window, but only the output regulation control quantity of the first control cycle is executed; the next window is solved again with the actual state after execution and the resource load output deviation as the initial information to form a rolling output regulation closed loop.

8. A resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, When optimizing the time domain The system matrix tends to infinity. , Output matrix Output weighted matrix and control weighting matrix When the matrix is ​​time-invariant, the Riccati differential equation degenerates into the algebraic Riccati equation. The steady-state output regulation feedback gain matrix obtained by solving It is a constant matrix; the system satisfies stability, detectability, and semi-positive definite Under positive definite conditions, this closed-loop resource load balancing scheduling structure helps improve the stability of the resource load output deviation adjustment process.

9. A resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, As an optional implementation, when communication delays exist during task unloading, task migration, or state acquisition between servers, the system dynamic equations in step 2 can further consider the impact of network transmission delays on state evolution, and incorporate control input terms. Replace with ,in For the server to server Average network latency for performing task unloading or task migration transfer. , To fix the propagation delay, This represents the average amount of data per task. This represents the average transmission rate of the link. For queuing delay estimation; The communication delay compensation is an auxiliary correction in the presence of delay, used to correct the state estimation or resource load output estimation, and to convert the corrected value into a more accurate and accurate result. and Used for subsequent output regulation error calculation, output regulation feedback law application, and scheduling control input generation.

10. A resource load balancing scheduling method based on output regulation of optimal control according to claim 1, characterized in that, The performance index functional described in step 3 can be extended into a multi-objective weighted form. ,in To output the adjustment error target term; after performing the equivalent coordinate translation process described in step 2, ensuring that the deviation-form state vector and the output adjustment error satisfy the mapping relationship, take... or ; For the system energy consumption target item, , For server Power consumption; For network latency target item, , For link transmission delay; weight , and satisfy And dynamically adjust according to the service quality requirements of the application scenario.

Images