Cross-layer resource scheduling method and related device

Abstract

The application discloses a cross-layer resource scheduling method and related equipment, and the method comprises the following steps: acquiring multi-source observation data of a heterogeneous edge node, and performing time alignment processing on the multi-source observation data to obtain aligned data; performing state estimation on the aligned data to obtain an uncertainty perception signal with confidence and a current state; mapping the uncertainty perception signal into a global gating state through a gated state machine; determining a resource capacity interval and a rhythm control parameter in a limited time domain rolling optimization mode based on the global gating state and the current state; selecting a target strategy generation mode among a learning strategy and a heuristic rule according to the global gating state, and generating a resource scheduling strategy for a to-be-processed task by taking the resource capacity interval, the rhythm control parameter and a safety threshold corresponding to the global gating state as constraint conditions through the target strategy generation mode.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a cross-layer resource scheduling method and related equipment. Background Technology

[0002] Although the "feedback-based closed-loop control" paradigm has formed a relatively mature technical system, in large-scale edge node deployment scenarios, resource scheduling and control heavily rely on real-time perception of node resource status. However, in actual operation, due to physical factors such as measurement noise, network latency, differences in edge node capabilities, and the unobservability of latent variables, status information often suffers from incompleteness or lag. For example, in scenarios such as emergency response, the system needs to rapidly expand to cope with a surge in user requests, but edge nodes are often in a low-power or heavily loaded state. Delayed resource perception can easily lead to containers being scheduled before they are ready, causing task failures, node overload, and consequently affecting service availability and system stability. Existing solutions generally lack quantitative characterization of the reliability of estimation results and systematic characterization of uncertainties, making it difficult for subsequent scheduling stages to effectively identify the credibility of state estimation results and failing to provide crucial confidence support for scheduling decisions.

[0003] Meanwhile, most existing resource management strategies rely on historical data or static models to generate decisions. The scheduling strategies themselves lack a universal mechanism for dynamic environments, making them ill-suited to the realities of sudden task bursts, frequent node online / offline cycles, and temporary resource failures in edge clouds. For example, in smart parks, edge nodes need to process massive amounts of camera video streams, and resource demands fluctuate drastically with time periods and events: threshold-triggered management frequently switches due to insufficient rule coverage, easily leading to resource allocation imbalances; predictive optimization management fails to predict accurately when new devices are added or scenarios change, easily causing strategy failures; while interactive learning management is adaptive, it is difficult to migrate to heterogeneous edge deployments due to inconsistent interfaces and network latency, making it difficult to form stable scheduling paths. The above scheduling strategies rely excessively on the independent triggering of local thresholds and rules within modules, resulting in insufficient adaptability between the scheduling rhythm and the overall system state.

[0004] The closed-loop feedback mechanism is responsible for adjusting the strategy or updating the model in real time based on observations during task execution to improve scheduling efficiency. However, in edge environments, due to instability factors such as high latency, large observation errors, and frequent ups and downs in the feedback chain, the feedback signal often carries strong noise and disturbance components. Without mechanisms such as adjustment rhythm control and feedback stability constraints, the system will have difficulty correctly distinguishing between "true trend changes" and "short-term anomalies," easily leading to over-adjustment, strategy jitter, frequent resource migration, and the formation of a non-convergent oscillating chain, which seriously undermines system stability.

[0005] Taking the Chinese patent "CN202411565720 Cloud Computing Storage and Training Resource Allocation Method Based on Reinforcement Learning and Dynamic Resource Scheduling" as an example, although the scheme constructs a resource state matrix and predicts resource demand, it does not introduce a compensation mechanism for measurement noise and time delay, which poses a risk of state perception lag. At the same time, although it has a real-time feedback and reward function adaptive mechanism, it still lacks frequency control and filtering design, which makes it prone to policy oscillation under high-frequency disturbances, affecting the stability and convergence of closed-loop scheduling. Summary of the Invention

[0006] In view of this, the present invention provides a cross-layer resource scheduling method to address the limitations of existing technologies in generating resource scheduling strategies. To achieve one, some, or all of the above objectives, or other objectives, the present invention proposes a cross-layer resource scheduling method, comprising:

[0007] Acquire multi-source observation data of heterogeneous edge nodes, and perform time alignment processing on the multi-source observation data to obtain aligned data; The aligned data is used to perform state estimation to obtain an uncertainty-aware signal with confidence and a global state estimation vector. The uncertainty-aware signal is then mapped to a global gated state through a gated state machine. Based on the global gating state, the resource capacity range and rhythm control parameters are determined using the current state in a finite-time-domain rolling optimization manner. Based on the global gating state, a target strategy generation method is selected from the learning-based strategy generation rules and heuristic rules. Using the target strategy generation method, a resource scheduling strategy for the task to be processed is generated with constraints such as the resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state.

[0008] Optionally, the step of performing time alignment processing on the multi-source observation data to obtain aligned data includes: Record the sending and receiving times of the state packets of each heterogeneous edge node, use the difference between the sending and receiving times as the delay observation, and calculate the delay estimate using an exponentially weighted moving average method; Align the observation time axis using the time delay estimate; If the value is not an integer, linear interpolation is used to calculate the interpolation weights, and time-aligned observations are generated based on the interpolation weights to obtain aligned data.

[0009] Optionally, the step of performing state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector includes: A weighted averaging strategy is used to construct a reference observation trajectory based on the aligned data; Calculate the Euclidean distance between each heterogeneous edge node and the reference observation trajectory, calculate the observation score signal based on the Euclidean distance, and obtain the discrete confidence signal through thresholding; When the discrete confidence signal indicates that the alignment data is reliable, the historical state estimate of the previous cycle is fused with the alignment data through Kalman filtering to obtain the global state estimation vector, the covariance matrix of the current state, and the trace value of the covariance matrix. The discrete confidence signal, the covariance matrix, and the trace of the covariance matrix are combined into a triplet, and the triplet is used as an uncertainty-aware signal with confidence.

[0010] Optionally, the gated state machine is a confidence-gated two-state finite state machine with a hysteresis mechanism, wherein the two-state finite state machine is used to generate a global gated state based on the triplet in each control cycle.

[0011] Optionally, before the step of determining the resource capacity range and rhythm control parameters using the current state based on the global gating state, the method further includes: The objective function for a finite-time rolling optimization method is constructed based on the path execution cost term, service risk penalty term, and rhythm smoothing term. The mathematical representation of the objective function is as follows:

[0012] For path execution cost, To serve risk penalty items, For rhythmic smoothness, These are the weighting coefficients. For task type In the cycle Assigned to the The scheduling volume of each node. This represents the rhythm boundary.

[0013] Optionally, the step of generating a resource scheduling strategy for the task to be processed using the target strategy generation method, with resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints, includes: When the target strategy generation method is a learning-based strategy generation rule, the set of tasks to be scheduled in the current task queue is obtained, and a task state vector for the task to be processed is constructed. Heterogeneous edge nodes that meet the activity requirements are designated as active nodes, and the node state vectors of the active nodes are constructed. The task state vector and the node state vector are concatenated to form state combination data, and the state combination data is input into the reinforcement learning policy network of the learning policy generation rule to obtain a set of candidate actions for the task to be processed. Using resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints, actions within the candidate action set are filtered to generate a resource scheduling strategy for the task to be processed.

[0014] Optionally, the method further includes: When the global gating state is 1, the operation index data during the execution of scaling up / down and resource scheduling actions are statistically analyzed, and the operation index data is weighted to obtain an immediate report. Based on the immediate feedback, the reinforcement learning policy network of the learning policy generation rules is updated using an online reinforcement learning update method based on PPO, so as to achieve continuous updating of the reinforcement learning policy network. When the global gating state is 0, the parameters of the reinforcement learning policy network in the learning policy generation rule remain unchanged.

[0015] Secondly, this application provides a cross-layer resource scheduling apparatus, comprising: The data processing module is used to acquire multi-source observation data of heterogeneous edge nodes and perform time alignment processing on the multi-source observation data to obtain aligned data. The mapping module is used to perform state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector, and to map the uncertainty-aware signal into a global gated state through a gated state machine. The limiting module is used to determine the resource capacity range and rhythm control parameters based on the global gating state and the current state, using a finite-time-domain rolling optimization method. The strategy generation module is used to select a target strategy generation method from learning-based strategy generation rules and heuristic rules based on the global gating state, and generate a resource scheduling strategy for the task to be processed by using the target strategy generation method, with resource capacity range, rhythm control parameters and the safety threshold corresponding to the global gating state as constraints.

[0016] Thirdly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the cross-layer resource scheduling method described above.

[0017] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the cross-layer resource scheduling method described above.

[0018] Implementing the embodiments of the present invention will have the following beneficial effects: This application constructs a unified resource state representation by time-aligning and fusing multi-source observation data, quantifying its sensing quality and fluctuation level, and generating a reliability signal coupled with the state. The gating system takes the "state representation + reliability signal" as input, selects a matching control mode, and ensures that the decision-making process reflects both the current system state and the stability of the sensing results, achieving effective transmission of state information to the control logic. Based on the gating signal (global gating state), resource limits are set, along with rhythm control parameters related to task migration and execution rate, serving as control inputs for lower-level scheduling. Based on the gating signal, candidate scheduling decisions are generated by switching between learning strategies and heuristic rules, and feasible region constraints are executed under resource capacity and rhythm constraints. The two-level control paths are linked in groups around the gating signal, adjusting synchronously with changes in system state to ensure the consistency of various control behaviors in terms of timing and intensity, improving the overall scheduling stability and rhythm adaptability. The gating signal serves as a pre-constraint to uniformly control updates: feedback is enabled to execute online updates only when the state is reliable and fluctuations are controllable; otherwise, parameter updates are frozen, and the variation range of key parameters is constrained to suppress the cumulative impact of abnormal disturbances and ensure the robustness and convergence of closed-loop scheduling. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] in: Figure 1 This is a flowchart illustrating a cross-layer resource scheduling method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating another cross-layer resource scheduling method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a cross-layer resource scheduling system provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this invention, it should be noted that unless otherwise explicitly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] like Figure 1 and Figure 2 As shown, this application provides a cross-layer resource scheduling method, including: S110. Obtain multi-source observation data of heterogeneous edge nodes, and perform time alignment processing on the multi-source observation data to obtain aligned data; In one possible implementation, the step of performing time alignment processing on the multi-source observation data to obtain aligned data includes: Record the sending and receiving times of the state packets of each heterogeneous edge node, use the difference between the sending and receiving times as the delay observation, and calculate the delay estimate using an exponentially weighted moving average method; Align the observation time axis using the time delay estimate; If the value is not an integer, linear interpolation is used to calculate the interpolation weights, and time-aligned observations are generated based on the interpolation weights to obtain aligned data.

[0024] For example, during the scheduling period Internally, the system from Each heterogeneous node collects operational status data for the current period to construct the multi-source observation data (original observation set). , where each observation vector Indicates the first The running status metrics of each node, such as CPU utilization, load queue length, task response latency, etc. (This represents the dimension of the state observation vector reported by each node).

[0025] Due to network jitter, asynchronous processing, and uneven node load, the reported observations from multiple nodes exhibit different latency, causing a "time misalignment" when the system aggregates observation data, which affects the accuracy of subsequent policy inference. To address this, the system introduces a unified time alignment mechanism to model and estimate the reporting latency of each node, thereby synchronizing all observation samples to a unified time benchmark.

[0026] To characterize the time difference between the observation at each node and the system reception, the true offset between the "observation occurrence time" and the "observation reception time" is modeled, and the system delay for each node is:

[0027] in, Represents a node Data transmission delay to the system scheduling center This indicates the local processing delay during node sampling, encapsulation, and enqueueing.

[0028] Because the system usually cannot obtain it directly. , The precise value is approximated using an estimation method, that is, by recording the difference between the sending and receiving times of the state packets of each node, and constructing delay observations. The system delay is estimated using an exponentially weighted moving average method.

[0029] As a smoothing factor, it controls the weight of the current observation in the estimate. This is an estimate from the previous period. This is the final calculated "time delay estimate".

[0030] The system passes Align the observation timeline. Let the current reference time of the scheduling system be... Then the node The estimated time when this observation actually occurred should be:

[0031] like For integer values ​​(observations falling on known sampling points), observations can be selected directly:

[0032] like If the value is not an integer (the observation falls between two time points), it cannot be directly assigned a value. The system uses linear interpolation to calculate the interpolation weights. And generate time-aligned observations. It simulates the actual operating state of nodes at non-integer time points, and realizes time synchronization of multi-node observation samples at a unified reference time.

[0033]

[0034]

[0035] Ultimately, the aligned observations of all nodes form the aligned set of node observations. , The nodal observation set (aligned data) will be used in subsequent steps to construct a reference trajectory, assess observation errors, and generate confidence score signals.

[0036] S120. Perform state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector, and map the uncertainty-aware signal to a global gated state through a gated state machine. In one possible implementation, the step of performing state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector includes: A weighted averaging strategy is used to construct a reference observation trajectory based on the aligned data; Calculate the Euclidean distance between each heterogeneous edge node and the reference observation trajectory, calculate the observation score signal based on the Euclidean distance, and obtain the discrete confidence signal through thresholding; When the discrete confidence signal indicates that the alignment data is reliable, the historical state estimate of the previous cycle is fused with the alignment data through Kalman filtering to obtain the global state estimation vector, the covariance matrix of the current state, and the trace value of the covariance matrix. The discrete confidence signal, the covariance matrix, and the trace of the covariance matrix are combined into a triplet, and the triplet is used as an uncertainty-aware signal with confidence.

[0037] For example, such as Figure 2 As shown, the deviations between the observations of each dimension and the unified reference trajectory are modeled and quantified. The consistency level of the current observation state is measured from two aspects: temporal consistency and numerical deviation, so as to comprehensively characterize the overall quality of the input data.

[0038] During the scheduling period Internally, the system utilizes the aligned set of node observations. Construct the reference observation trajectory vector This serves as the expected representation of the global state. The construction method employs a weighted average strategy:

[0039] in, For the first Alignment observations of individual nodes, As a weighting factor for nodes, The system uses equal weights by default.

[0040] The system measures the observation offset of each node based on the Euclidean distance between each node and the reference trajectory, and calculates the current system observation score signal. It describes the degree of consistency of the observation states of multiple nodes at the current moment.

[0041]

[0042] The smaller the value, the higher the aggregation of observations at each node, the better the state convergence, and the more stable and reliable the system as a whole. The larger the value, the greater the difference in observations, indicating that the system is in a disturbed or abnormal state.

[0043] To transform the deviation results between the observations of each node and the reference trajectory into a judgment signal that the scheduling system can recognize and execute, the system will use continuous deviation indicators. Mapped to standardized gated inputs This serves as the direct input for subsequent state determination mechanisms.

[0044]

[0045] The offset scoring threshold set for the system is used to determine whether the current observation state is reliable. If the system deems the current observation reliable, it includes it in the state estimation process; otherwise, it skips that dimension or reduces its weight during fusion to avoid errors introduced by high-noise data. This mechanism ensures the overall robustness and controllability of the state estimation input even when the observation quality at edge nodes is unstable.

[0046] To further enhance the system's steady-state perception and scheduling decision-making capabilities in dynamic environments, the system needs to integrate historical states and current reliable observations to generate an optimal estimate of the system state and its uncertainty quantification index. This step is based on Kalman filtering for modeling, and achieves dynamic estimation and confidence assessment of the global state by recursively fusing observations and predictions over time series.

[0047] Define the global state vector of the system as Indicates the scheduling period The system's multidimensional resource state at any given time (e.g., CPU utilization, memory usage, I / O throughput, task latency, etc.). The system's state evolves over time, influenced by changes in task load and noise disturbances. This evolution can be modeled as a linear state transition model:

[0048] in, This is the state transition matrix, which characterizes the temporal dependencies of the system states. Typically, an identity matrix can be used. This indicates that the system state evolves smoothly over a short period of time; The covariance matrix represents the inherent disturbance process noise of the system. It reflects the intensity of system fluctuations.

[0049] To balance estimation smoothness and sensitivity, the system... Instead of directly modeling it as the full covariance matrix, a scaling update mechanism is used for dynamic adjustment: the baseline is first estimated offline. Each of them Represents the state vector of the th The variance in each dimension, characterizing the typical fluctuation range of the resource metric (such as CPU, memory, or latency) in the historical state sequence, can be estimated from the first-order difference variance of the state residuals:

[0050] These variance terms are estimated offline from historical trajectories before the system goes live, forming the system's baseline noise level. During online operation, only a global scaling factor is updated. The noise level is adjusted based on the variance changes of observed innovation items:

[0051]

[0052]

[0053] The deviation between current observations and predictions is defined as follows:

[0054] In each cycle, the system obtains a time-aligned set of observations. Each node's observation follows the following observation model, where The observation matrix is ​​usually taken as the identity matrix. Indicates the state of the system being directly observed; To characterize observation noise, we define nodal observation errors. Observation noise reflects random disturbances during the measurement and reporting process, and its covariance matrix... It can be obtained through statistical analysis of historical observation residuals at nodes.

[0055]

[0056] The system in the current cycle Based on the estimate from the previous period and Perform prediction and make prior estimates.

[0057]

[0058] Next, according to The system dynamically decides whether to perform a state update. When this indicates that the current observation is reliable, the system performs observation fusion; when In such cases, the system skips updates and directly retains the prediction results to prevent noisy observations from shifting the estimation.

[0059] During the fusion process, the system first calculates weighted observations based on node weights (default equal weight).

[0060] The comprehensive observation noise covariance is then formed using the nodal observation error covariance matrix. Then, the Kalman gain is calculated:

[0061] Kalman gain The relative weights of prediction and observation are determined during the fusion process. If the observation noise is large, the update magnitude is small; if the system prediction uncertainty is high, the observation correction weight is increased.

[0062] The fusion update process is defined as follows: calculate and output the posterior state estimate. and its covariance .

[0063]

[0064]

[0065] when At that time, directly order , This ensures the system remains stable under low-confidence observation conditions.

[0066] Finally, the system output is represented in the form of triples. As the core input of the confidence gating system, it provides a continuous, stable, and interpretable basis for its state determination and scheduling control, realizing adaptive closed-loop control of "perception-evaluation-scheduling".

[0067] In one possible implementation, the gated state machine is a confidence-gated two-state finite state machine with a hysteresis mechanism, wherein the two-state finite state machine is used to generate a global gated state based on the triplet in each control cycle.

[0068] For example, the gated state output in each cycle It does not directly participate in resource scheduling decisions, but rather acts as the "master control signal" for the entire system's strategy switching and behavior regulation, determining the scheduler's execution mode. Its core function is to dynamically switch the system's operating strategy between "steady-state exploration" and "conservative avoidance" based on the system's current observation quality and state uncertainty, forming a confidence-based hierarchical scheduling linkage system to achieve a balance between stability, adaptability, and robustness in resource scheduling. This step focuses on explaining the behavior switching logic and linkage mechanism under gating conditions.

[0069] When the gate status is At this point, the system is determined to be in a confident stable state, and both the current observation quality and the state fusion evaluation results meet the standards, allowing the scheduler to run the high-performance optimization mode: The system employs a reinforcement learning scheduler based on the policy gradient method. It scores multiple candidate scheduling actions through a policy network to quantify the system's preference for each action in the current environment. Based on this, the system selects the scheduling action to achieve multi-objective optimization of latency, energy consumption, resource migration overhead, and SLA default. The policy network is kept active during training, allowing parameters to be updated at runtime. And perform real-time return calculations and advantage function estimations online; To fully unleash the expressive power of the strategy space, the system employs performance-prioritized rhythm control parameters to adjust the sensitivity and frequency of resource control, such as the maximum rate of resource requests. Minimum start / stop time window for tasks Concurrent migration limit Trajectory smoothing coefficient ; Before scheduling execution, the system uses a safety pruning mechanism. Policy network The output action set undergoes feasibility screening, using rules such as resource availability thresholds and popularity limits as constraints to exclude actions with high potential risks, reducing their probability of being sampled and executed. Under stable conditions, A relatively lenient filtering threshold is adopted to ensure the exploratory nature and diversity of the strategy.

[0070] When the system enters an unstable state (gated state is...) In order to mitigate the systemic risks that may arise from observation drift or decision bias, the system will immediately switch to conservative mode. The reinforcement learning scheduler switches to degrade mode, and the policy network... The output will no longer participate in the actual scheduling decision, parameters Updates will also be suspended, freezing the online learning process to avoid erroneous learning under unstable observations; The scheduler switches to a lightweight heuristic strategy (heuristic rule), which is a simple allocation and scheduling based on "idle node priority". It does not require model training and has stronger short-term robustness and fast response capability. Synchronously adjust the rhythm control parameters to suppress system fluctuations, such as lowering the upper limit of resource request rate. Extend the minimum start / stop time window for tasks Reduce the maximum number of concurrent migrations Improve trajectory smoothness coefficient ; Safe clipping mechanism Implement more conservative filtering thresholds, including resource availability and node popularity, to narrow the action space and reduce the probability of potentially uncertain behaviors being adopted.

[0071] A dynamic linkage was established between the gating state and the execution of the scheduling strategy: on the one hand, the system will control the gating state... As a control signal for the scheduler, it drives the switching between reinforcement learning and heuristic policies. On the other hand, the system coordinates and configures rhythm adjustment parameters to ensure that the scheduling behavior has clear policy boundaries and execution constraints under different stability conditions. This mechanism enhances the system's ability to perceive and respond to abnormal states, enabling it to effectively avoid triggering high-risk policies when the observation confidence level drops significantly.

[0072] S130. Based on the global gating state, the resource capacity range and rhythm control parameters are determined by the current state using a rolling optimization method in the finite time domain. In one possible implementation, prior to the step of determining the resource capacity range and rhythm control parameters using the current state based on the global gating state, the method further includes: The objective function for a finite-time rolling optimization method is constructed based on the path execution cost term, service risk penalty term, and rhythm smoothing term. The mathematical representation of the objective function is as follows:

[0073] For path execution cost, To serve risk penalty items, For rhythmic smoothness, These are the weighting coefficients. For task type In the cycle Assigned to the The scheduling volume of each node. This represents the rhythm boundary.

[0074] For example, within each control cycle, this layer receives input triples. The gating state is triggered and obtained through hysteresis determination. This is then uniformly mapped to parameter settings and resource boundary constraints for lower-level scheduling.

[0075] Current cycle rhythm constraint The definition is as follows:

[0076] in, This sets an upper limit on the rate of resource request / release, controlling the magnitude of resource fluctuations within a single cycle; This represents the minimum start / stop time window for a task, used to limit frequent starts and stops and suppress scheduling jitter. It is a trajectory smoothing factor that controls the smoothness of changes in scheduling instructions; This sets a limit on the number of concurrent migrations within the same period. A capacity control parameter is also defined. :

[0077] Indicates the first Resource types in cycle The target occupancy or available limit within, This represents the total allocation for the resource pool.

[0078] When the gate status is At that time, the system is determined to be in a confident stable state, and performance-prioritized rhythm control parameters are adopted, i.e. When the gate state is When the system enters an unstable state, the rhythm constraint is adjusted synchronously to suppress system fluctuations. The parameter configuration must satisfy the following inhibition relationship:

[0079] A path-sensitive migration cost modeling framework is introduced, using "task type → target node" as the basic unit to model the migration overhead and potential risks in the task scheduling process.

[0080] System defines flow-level allocation variables This indicates that within the current period, the planned quantity is... "Task type is The task request is scheduled to the target node. Task type Defined as a set of jobs with similar resource configurations and service attributes (such as latency-sensitive streams, general-purpose computing, low-priority tasks, etc.), used to characterize the statistical behavior of resource flow. For task types... Dispatch from the original resident node to the node Let the path have a cost of . , representing the execution cost of a typical migration operation under this path, is calculated using the following formula:

[0081] Among them, three components The time consumption, energy consumption, and disturbance of the migration operation are measured separately. The weighting coefficients configured for the strategy are used to reflect the system's priorities in terms of performance, energy efficiency, and steady-state defense.

[0082] Time-consuming items Measuring tasks Assigned to nodes The average latency from triggering to convergence (e.g., resource readiness, cold start completion) is used to extract all actions from historical execution records. (i.e., migration path and) Matching scheduling), based on its single-time consumption. Calculate the expected value:

[0083] Energy consumption items Indicates the completion of one cycle from arrive The energy consumed by the migration operation is typically affected by the task size and link characteristics. Its estimation formula is shown below, where... Indicates task type Average data volume It is the energy consumption factor per unit of data on the path, which depends on the physical bandwidth of the path and the network structure (such as whether remote communication is involved).

[0084]

[0085] Disturbance term The potential risks that the path migration poses to the stability of the target system are assessed, mainly including target node pressure, task sensitivity, and path risk factors.

[0086]

[0087] It can be configured in conjunction with business strategies to reflect the system's sensitivity to "steady-state disturbances"; For nodes Current resource disruption metrics, such as burst load rate and load fluctuations; For task type Delay sensitivity level; The system identifies the environmental vulnerability of the path, such as whether it is a hotspot path, whether it has experienced failures, and whether there are rate limits. All three components support online updates and threshold adjustments.

[0088] In each control cycle, based on state estimation and its covariance Perform a forward prediction step to obtain the prior state estimate for the next cycle. :

[0089] Since the upper layer does not directly use all state components, but instead focuses on aggregated resource load (such as the CPU, memory, storage, and bandwidth requirements of the resource pool), the system further uses a linear readout matrix. The state is mapped to the resource dimension to obtain the load forecast for the next cycle:

[0090] in It is a resource load vector. It is the covariance of its predictive uncertainty. Matrix Each line defines how to extract or weight multiple state variables from the system state vector to form the load of a certain resource dimension. For example, if The dimensional state contains the CPU utilization of multiple nodes. The corresponding rows can be used to calculate their average, weighted, or maximum values, thereby constructing system-level CPU demand forecasts.

[0091] To improve prediction stability and buffer short-term noise, the system will perform a lightweight exponentially weighted smoothing on the load prediction to obtain a smoothed load estimate. ( The smoothing coefficient is the initial value. Historical averages can be used. It serves as the actual input for rolling optimization and participates in path scheduling planning and rhythm adjustment.

[0092]

[0093] The system in each control cycle A finite-time domain rolling optimization method is adopted, based on rhythm constraints. Path cost and lightweight load prediction sequences Construct a robust optimal control problem. This problem has a length of... By jointly considering path execution cost, service risk penalty, and rhythm smoothing term within the rolling time domain, the optimal allocation and rhythm plan for the current cycle are obtained. Only the result of the current step is executed, and the remaining plans are used as references and updated over time.

[0094] In this rolling solution process, the state prediction uses the same system dynamics as obtained above, namely the state transition matrix. Covariance of process noise and the fusion state Pushing into the future Step to get Using a linear readout matrix Mapping the predicted state to the resource load dimension yields the predicted demand trajectory. And obtain its covariance. ,in .

[0095] Based on this, the system defines the decision variable as the rhythm parameter trajectory for each step in the prediction time domain. and task migration / scheduling plan matrix

[0096] This indicates the "task type" "In the cycle" Assigned to target node The current scheduling volume.

[0097] The objective function for rolling optimization includes a path execution cost term, a service risk penalty term, and a rhythm smoothing term. Path execution cost term By weighted summing of all possible combinations of task types and target nodes within the prediction interval, the resource consumption and path loss cost of the overall task migration behavior are measured.

[0098] in, Indexed by task type. Index the target node. During the control cycle. Internally, the system's scheduling plan for each type of task. This will result in actual consumption of various system resources. Considering the differences in resource requirements among different types of tasks, a task-resource mapping matrix is ​​defined. :

[0099] This indicates that each scheduling type is The types of resources consumed by the task The unit resource quantity (e.g., Mbps bandwidth). According to the scheduling plan. Calculate the resource requirements of all scheduled tasks within this period. Total usage:

[0100] By comparing the resource usage with the system's resource type Forecast demand Determine whether the current scheduling plan meets the predicted load demand. If This indicates that there is insufficient scheduling of this type of resource in this cycle. Definition , indicating the difference:

[0101] This represents resource demands that are not yet met under the current scheduling scheme. To penalize these unmet demands, a service risk penalty term is added to the optimization objective. :

[0102] This refers to the risk weights for each type of resource, reflecting the system's sensitivity adjustment mechanism for resource types with high prediction uncertainty. The rhythm smoothing term represents the risk weights at each time step. Compared to the previous step The amplitude of rhythm parameter changes, according to Norm measure (i.e., the sum of the absolute values ​​of the element-wise differences), coefficient Controlling sensitivity to non-smooth changes:

[0103] Finally, the objective function for rolling optimization is defined as follows, reflecting the trade-off preferences between path cost, risk control, and rhythm stability.

[0104]

[0105] The rolling optimization problem includes resource capacity constraints and pacing boundary constraints. Resource capacity constraints are implemented using a task-resource mapping matrix. The task scheduling load is converted into the total demand for resource class k, ensuring that it does not exceed the capacity limit for that period. :

[0106] Rhythm boundary constraints are based on the gating state. Choose different rhythm control parameters:

[0107] The parameter configurations satisfy the following relationship:

[0108] S140. Based on the global gating state, select the target strategy generation method from the learning-based strategy generation rules and the heuristic rules, and generate a resource scheduling strategy for the task to be processed using the target strategy generation method, with the resource capacity range, rhythm control parameters and the safety threshold corresponding to the global gating state as constraints.

[0109] In one possible implementation, the step of generating a resource scheduling strategy for the task to be processed using the target strategy generation method, with resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints, includes: When the target strategy generation method is a learning-based strategy generation rule, the set of tasks to be scheduled in the current task queue is obtained, and a task state vector for the task to be processed is constructed. Heterogeneous edge nodes that meet the activity requirements are designated as active nodes, and the node state vectors of the active nodes are constructed. The task state vector and the node state vector are concatenated to form state combination data, and the state combination data is input into the reinforcement learning policy network of the learning policy generation rule to obtain a set of candidate actions for the task to be processed. Using resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints, actions within the candidate action set are filtered to generate a resource scheduling strategy for the task to be processed.

[0110] For example, the received rhythm constraint quantity Resource capacity limitations The strategy outputs actions by combining confidence information with the real-time status of the nodes. Feasibility trimming is performed to obtain a controlled, practically executable subset of actions, avoiding the risk of violating resource constraints or causing overload during actual execution. (Set of executable actions) The definition is as follows:

[0111] in, A set of actions that satisfy the following conditions: Resource feasibility: target node Current remaining available resources The total load across each resource category shall not exceed the resource capacity limit for this period.

[0112] Node popularity limit: Current popularity of the target node If the system is in an unstable state If the threshold is not met, the protection will be automatically lowered to enhance protection.

[0113] Rhythm boundary constraints: determined by the upper-level rhythm constraint quantity Unified restrictions include upper limits on the number of migrations, instantaneous rate limits, and jitter smoothing rules.

[0114] For each task Strategy output The system constructs a safe pruning function. Perform safety constraint trimming and feasibility modification:

[0115] If candidate action Then, candidate actions generated using heuristic strategies are used. : In the set of executable actions The node with the highest strategy score is selected from the internal nodes; if If empty, skip the current cycle scheduling or retain it in the original node.

[0116] If the system is in a confident stable state The scheduler enters normal operating mode, and the policy network operates in the original action space. The middle outputs candidate actions for each task to be scheduled. Subsequently, Perform feasibility trimming on the action set to obtain an executable action set. And select the final action from them. ,at this time, Use a lenient filtering threshold.

[0117] If the system is in an unstable state The scheduler switches to degraded operation mode, adopting a lightweight heuristic strategy, i.e., a simple allocation and scheduling based on "idle node priority," requiring no model training. The policy network's parameters are updated ( The online learning (database) is frozen, and its output is used only as a reference signal, no longer for the primary action selection. At this time, Use stricter feasibility filtering thresholds to prioritize system stability and prevent high-risk actions from being adopted.

[0118] Ultimately, the set of scheduling actions Distribute to the edge resource orchestration engine. System records are as row nodes. Feedback on whether a degradation strategy has been enabled, actual resource usage, and node status changes is used for strategy updates, forming a scheduling learning loop.

[0119] It receives the output of capacity control and scheduling strategies, executes the generated resource configuration instructions and scheduling action sets as system operations, and collects operational feedback to build real-time reports. This drives the online updates of strategies and models, forming a complete closed-loop control chain of "perception-decision-execution-feedback", continuously improving the system's adaptability and stability under heterogeneous nodes, multi-tenant loads and dynamic resource changes.

[0120] In each control cycle, the system receives the rhythm constraint output by the upper-level planning module. and the set of scheduling actions output by the policy network. Capacity control commands will be parsed into specific resource change behaviors, such as scaling up, releasing, or cold starting edge nodes, while scheduling actions... This corresponds to the binding and migration of tasks with edge nodes.

[0121] To avoid inconsistencies in state caused by resource execution delays, the system writes all control information to a "control instruction buffer awaiting activation" and maintains the execution status of each instruction in real time, including: task... Task Type Target Node Execution status (Pending execution, in progress, completed, failed) and response confirmation time This information ultimately forms the control log structure. This provides a basis for execution alignment for subsequent feedback collection and strategy learning.

[0122]

[0123] To prevent drastic changes in the upper-level rhythm parameters within adjacent control cycles, which could lead to instability and sudden jitter in the execution layer, the system imposes rhythm constraints on the output of the rolling strategy optimization.

[0124] Component-by-component speed limiting control is introduced to ensure that the variation range is limited to a reasonable range. The upper limit of the difference is defined for each rhythm component as follows:

[0125]

[0126] in Upper limit of the difference between each component From gated state Control. In unstable states, the system adopts a more conservative update strategy. This ensures the stability of parameter evolution. The rate limiting mechanism reduces the risk of execution-level instability caused by sudden parameter tuning by suppressing abrupt changes in policy updates.

[0127] In one possible implementation, the method further includes: When the global gating state is 1, the operation index data during the execution of scaling up / down and resource scheduling actions are statistically analyzed, and the operation index data is weighted to obtain an immediate report. Based on the immediate feedback, the reinforcement learning policy network of the learning policy generation rules is updated using an online reinforcement learning update method based on PPO, so as to achieve continuous updating of the reinforcement learning policy network. When the global gating state is 0, the parameters of the reinforcement learning policy network in the learning policy generation rule remain unchanged.

[0128] For example, after resource operations and task scheduling are completed, the system collects a set of feedback metrics in real time, including: average task response latency. Energy consumption Rhythm fluctuation (measure (such as the resulting uneven trajectory), and indicators of whether a Service Level Agreement (SLA) breach has been triggered. ,in This indicates that a task response latency exceeds its SLA threshold within the period, otherwise it is 0. Based on the above metrics, the system defines an immediate reward function:

[0129] Where the coefficient The adjustable weights reflect business objectives such as latency sensitivity, energy consumption control, and stability preference. This reward value will be used to construct the advantage function in subsequent policy network training. Simultaneously, the system outputs a set of KPI indicators for the current period. As a key input for strategy performance analysis.

[0130] The system calls a policy update algorithm based on near-end policy optimization to update the policy parameters. Fine-tuning will be performed. The specific optimization objective is:

[0131] When the system is in a confident stable state Only when the condition is met should parameter updates be performed on the model; otherwise, the current parameters should be frozen to avoid low-quality observations contaminating the learning process.

[0132] Finally, the system will output the updated policy parameters. It also pushes periodic KPI indicators, strategy change status, and abnormal alarms (such as SLA default, resource drift, etc.) to the upper-layer module to form an adaptive optimization closed loop, continuously enhancing the scheduling robustness and system stability of the resource pool in the edge cloud environment.

[0133] Specifically, the parameters of the policy network The system is optimized using the Proximal Policy Optimization (PPO) loss function. The system employs the PPO's pruning objective function. The probability ratio between the current policy and the previous policy is defined as:

[0134] Constructing the advantage function This indicates how much better the current action is than the average effect of the strategy. This represents the expected reward of the current task-action pair (which can be calculated from the cumulative discounted reward in the actual execution trajectory). This represents the baseline value estimate in the current global state, used to reduce variance: it incorporates all task-node interaction features. After summarizing, an average pooling operation is performed to obtain the global context vector, which is then processed by a small feedforward neural network of 1-2 layers. Perform regression prediction and output scalar values. As the expected scheduling value of the current global state.

[0135] Based on the above definition, the objective function of PPO is:

[0136] Used to limit the policy ratio The update range, This is the shear coefficient (generally set between 0.1 and 0.2).

[0137] In each control cycle of the system, the original motion space is first constructed. , Includes tasks Scheduled to node This includes all possible scheduling behaviors, such as deferring scheduling. The policy network in... Output motion score and candidate motion And continuously optimize strategy parameters In order to learn better scheduling preferences.

[0138] Secondly, such as Figure 3As shown, this application provides a cross-layer resource scheduling device, including: The data processing module 201 is used to acquire multi-source observation data of heterogeneous edge nodes and perform time alignment processing on the multi-source observation data to obtain aligned data. The mapping module 202 is used to perform state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector, and to map the uncertainty-aware signal into a global gated state through a gated state machine. The limiting module 203 is used to determine the resource capacity range and rhythm control parameters based on the global gating state and the current state using a finite-time-domain rolling optimization method. The strategy generation module 204 is used to select a target strategy generation method from learning-based strategy generation rules and heuristic rules according to the global gating state, and generate a resource scheduling strategy for the task to be processed by using the target strategy generation method, with resource capacity range, rhythm control parameters and the safety threshold corresponding to the global gating state as constraints.

[0139] In one possible implementation, such as Figure 4 As shown, this application embodiment provides a terminal device 300, including: a memory 310, a processor 320, and a first computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the first computer program 311, it performs the following steps: acquiring multi-source observation data of heterogeneous edge nodes, performing time alignment processing on the multi-source observation data to obtain aligned data; performing state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector, and mapping the uncertainty-aware signal to a global gated state through a gating state machine; determining the resource capacity range and rhythm control parameters based on the global gated state using a finite-time-domain rolling optimization method; selecting a target strategy generation method from learning-based strategy generation rules and heuristic rules according to the global gated state, and generating a resource scheduling strategy for the task to be processed using the target strategy generation method, with the resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gated state as constraints.

[0140] In one possible implementation, such as Figure 5As shown, this application embodiment provides a computer-readable storage medium 400, on which a second computer program 411 is stored. When the second computer program 411 is executed by a processor, it implements the following steps: acquiring multi-source observation data of heterogeneous edge nodes, performing time alignment processing on the multi-source observation data to obtain aligned data; performing state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector, and mapping the uncertainty-aware signal to a global gating state through a gating state machine; determining the resource capacity range and rhythm control parameters through the current state using a finite-time-domain rolling optimization method based on the global gating state; selecting a target strategy generation method from learning-based strategy generation rules and heuristic rules according to the global gating state, and generating a resource scheduling strategy for the task to be processed through the target strategy generation method, using the resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints.

[0141] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

[0142] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A cross-layer resource scheduling method, characterized in that, include: Acquire multi-source observation data of heterogeneous edge nodes, and perform time alignment processing on the multi-source observation data to obtain aligned data; The aligned data is used to perform state estimation to obtain an uncertainty-aware signal with confidence and a global state estimation vector. The uncertainty-aware signal is then mapped to a global gated state through a gated state machine. Based on the global gating state, the resource capacity range and rhythm control parameters are determined using a finite-time-domain rolling optimization method through the current state. Based on the global gating state, a target strategy generation method is selected from the learning-based strategy generation rules and heuristic rules. Using the target strategy generation method, a resource scheduling strategy for the task to be processed is generated with constraints such as the resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state. When the global gating state is 1, the operation index data during the execution of scaling up / down and resource scheduling actions are statistically analyzed, and the operation index data is weighted to obtain an immediate report. Based on the immediate feedback, the reinforcement learning policy network of the learning policy generation rules is updated using the PPO-based online reinforcement learning update method to achieve continuous updating of the reinforcement learning policy network, and a resource scheduling policy for the task to be processed is generated based on the updated reinforcement learning policy network. When the global gating state is 0, the parameters of the reinforcement learning policy network in the learning policy generation rule remain unchanged, and a resource scheduling policy for the task to be processed is generated using heuristic rules.

2. The cross-layer resource scheduling method as described in claim 1, characterized in that, The step of performing time alignment processing on the multi-source observation data to obtain aligned data includes: Record the sending and receiving times of the state packets of each heterogeneous edge node, use the difference between the sending and receiving times as the delay observation, and calculate the delay estimate using an exponentially weighted moving average method; Align the observation time axis using the time delay estimate; If the value is not an integer, linear interpolation is used to calculate the interpolation weights, and time-aligned observations are generated based on the interpolation weights to obtain aligned data.

3. The cross-layer resource scheduling method as described in claim 1, characterized in that, The step of performing state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector includes: A weighted averaging strategy is used to construct a reference observation trajectory based on the aligned data; Calculate the Euclidean distance between each heterogeneous edge node and the reference observation trajectory, calculate the observation score signal based on the Euclidean distance, and obtain the discrete confidence signal through thresholding; When the discrete confidence signal indicates that the alignment data is reliable, the historical state estimate of the previous cycle is fused with the alignment data through Kalman filtering to obtain the global state estimation vector, the covariance matrix of the current state, and the trace value of the covariance matrix. The discrete confidence signal, the covariance matrix, and the trace of the covariance matrix are combined into a triplet, and the triplet is used as an uncertainty-aware signal with confidence.

4. The cross-layer resource scheduling method as described in claim 3, characterized in that, The gated state machine is a confidence-gated two-state finite state machine with a hysteresis mechanism. The two-state finite state machine is used to generate a global gated state based on the triplet in each control cycle.

5. The cross-layer resource scheduling method as described in claim 1, characterized in that, Before the step of determining the resource capacity range and rhythm control parameters using a finite-time-domain rolling optimization method based on the global gating state and the current state, the method further includes: The objective function for a finite-time rolling optimization method is constructed based on the path execution cost term, service risk penalty term, and rhythm smoothing term. The mathematical representation of the objective function is as follows: For path execution cost, To serve risk penalty items, For rhythmic smoothness, These are the weighting coefficients. For task type In the cycle Assigned to the The scheduling volume of each node. This represents the rhythm boundary.

6. The cross-layer resource scheduling method as described in claim 1, characterized in that, The step of generating a resource scheduling strategy for the task to be processed using the target strategy generation method, with resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints, includes: When the target strategy generation method is a learning-based strategy generation rule, the set of tasks to be scheduled in the current task queue is obtained, and a task state vector for the task to be processed is constructed. Heterogeneous edge nodes that meet the activity requirements are designated as active nodes, and the node state vectors of the active nodes are constructed. The task state vector and the node state vector are concatenated to form state combination data, and the state combination data is input into the reinforcement learning policy network of the learning policy generation rule to obtain a set of candidate actions for the task to be processed. Using resource capacity range, rhythm control parameters, and the safety threshold corresponding to the global gating state as constraints, actions within the candidate action set are filtered to generate a resource scheduling strategy for the task to be processed.

7. A cross-layer resource scheduling device, characterized in that, include: The data processing module is used to acquire multi-source observation data of heterogeneous edge nodes and perform time alignment processing on the multi-source observation data to obtain aligned data. The mapping module is used to perform state estimation on the aligned data to obtain an uncertainty-aware signal with confidence and a global state estimation vector, and to map the uncertainty-aware signal into a global gated state through a gated state machine. The limiting module is used to determine the resource capacity range and rhythm control parameters based on the global gating state and the current state using a finite-time-domain rolling optimization method. The strategy generation module is used to select a target strategy generation method from learning-based strategy generation rules and heuristic rules based on the global gating state, and generate a resource scheduling strategy for the task to be processed by using the target strategy generation method, with resource capacity range, rhythm control parameters and the safety threshold corresponding to the global gating state as constraints. When the global gating state is 1, the operation index data during the execution of scaling up / down and resource scheduling actions are statistically analyzed, and the operation index data is weighted to obtain an immediate report. Based on the immediate feedback, the reinforcement learning policy network of the learning policy generation rules is updated using the PPO-based online reinforcement learning update method to achieve continuous updating of the reinforcement learning policy network, and a resource scheduling policy for the task to be processed is generated based on the updated reinforcement learning policy network. When the global gating state is 0, the parameters of the reinforcement learning policy network in the learning policy generation rule remain unchanged, and a resource scheduling policy for the task to be processed is generated using heuristic rules.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the cross-layer resource scheduling method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the cross-layer resource scheduling method as described in any one of claims 1 to 6.

Images