An agent cluster flow scheduling and fault-tolerant control method for complex scenes

Abstract

This invention relates to the field of collaborative control in artificial intelligence and distributed systems, specifically disclosing a method for agent cluster process scheduling and fault-tolerant control in complex scenarios. This method constructs a system entropy model based on the weighted fusion of multi-dimensional state variables to quantify cluster disorder in real time. When the entropy growth rate exceeds a threshold, a virtual inertia control mechanism is triggered, dynamically injecting a virtual buffer pool with adjustable capacity to absorb sudden traffic. A physical damping mechanism is used to constrain the frequency and amplitude of task migration, and task allocation weights are nonlinearly adjusted based on entropy feedback. If the system remains unstable, a three-level degradation and gradual recovery strategy is initiated. Through the above technical solutions, this invention achieves proactive fault perception, impact buffering, suppression of scheduling jitter, and continuous assurance of core business operations, improving the stability and robustness of agent clusters in high-concurrency complex scenarios.

Description

Technical Field

[0001] This invention belongs to the field of collaborative control of artificial intelligence and distributed systems, specifically involving a method for agent cluster process scheduling and fault-tolerant control in complex scenarios. Background Technology

[0002] With the rapid development of distributed computing and artificial intelligence collaborative technologies, agent clusters, as the core carriers for handling large-scale concurrent tasks and complex business logic, have demonstrated extremely high application value in fields such as internet architecture, autonomous driving, and intelligent manufacturing. Through collaborative division of labor and dynamic resource allocation among multiple nodes, agent clusters integrate computing power with business flexibility, providing underlying architectural support for coping with highly dimensional and ever-changing business environments. In the process of distributed collaborative work, the stability, real-time performance, and resilience to unexpected situations of the cluster are important standards for measuring system reliability.

[0003] Process scheduling and fault tolerance control are core technologies for maintaining the high-performance operation of agent clusters. Their fundamental goal is to ensure a balanced load distribution across nodes and continuous system availability through scientific resource scheduling algorithms and defensive control logic. This technology requires the system to accurately model real-time task flows, data flows, and node health status, and to achieve smooth transmission of logical instructions within the business topology. To maintain system order in complex environments, the scheduling system must possess the ability to sensitively identify potential risks and the capacity for self-adjustment and balancing when subjected to disturbances.

[0004] Existing technologies often suffer from system avalanche and cascading failures when faced with sudden traffic surges due to a lack of proactive state awareness. Traditional rate limiting and degradation strategies are typically based on static feedback mechanisms, whose response speed is insufficient to cover the critical time from local overload to overall system collapse, causing faults to spread rapidly within the cluster. Existing dynamic load balancing algorithms often suffer from severe jitter when handling drastically fluctuating task flows due to a lack of physical damping characteristics, resulting in frequent and blind migration of tasks between different nodes, consuming system bandwidth and computing resources. Existing scheduling architectures generally lack elastic buffering logic, making it difficult to absorb and smooth the impact energy of massive data spikes, resulting in a severe lack of cluster robustness in extreme scenarios. This presents a critical technical challenge requiring a solution: an agent cluster process scheduling and fault-tolerant control method for complex scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method for agent cluster process scheduling and fault tolerance control in complex scenarios, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for agent cluster process scheduling and fault tolerance control in complex scenarios, comprising the following specific steps: Step 1: Construct a system state awareness model for the Agent cluster, using the real-time load, task queue length, response latency, and resource utilization of each Agent node as basic state variables, and generate a system entropy value that represents the overall orderliness of the system through weighted fusion. Step 2: Based on the trend of the system entropy value change, determine whether there is an entropy increase phenomenon caused by local node anomalies. When the rate of change of entropy value exceeds a preset threshold, trigger the scheduling early warning mechanism. Step 3: After the scheduling warning mechanism is triggered, the virtual inertia control module is activated, and a virtual buffer pool is dynamically injected into the task scheduling path. This virtual buffer pool is used to temporarily accept sudden high-concurrency traffic and delay its direct transmission to downstream nodes. Step 4: Based on the real-time feedback of the virtual buffer pool's fill status and system entropy, adjust the task allocation weights so that newly arriving tasks are preferentially routed to Agent nodes with lower system entropy and sufficient resource reserves. Step 5: During the process of releasing tasks from the virtual buffer pool, a physical damping mechanism is introduced to impose dual constraints on the frequency and magnitude of task migration, preventing jitter caused by frequent rescheduling. Step 6: Continuously monitor whether the system entropy value has fallen back to the stable range. If it does not recover within the predetermined time, start the tiered degradation strategy and restrict the scheduling permissions of non-core tasks in stages according to the preset priority to ensure the continuous operation of critical business links.

[0007] Preferably, in step 1, the calculation of the system entropy value adopts the normalization processing result of multidimensional state variables. Each state variable is assigned a different weight coefficient according to its influence on system stability. The weight coefficient is obtained by training with historical fault data and is updated online regularly to adapt to the evolution of business scenarios.

[0008] Preferably, in step 2, the determination of the entropy change rate adopts the differential analysis method within a sliding time window. The window length is dynamically adjusted according to the cluster size to ensure that the inflection point of entropy increase can be accurately captured under different load scales.

[0009] Preferably, the capacity of the virtual buffer pool in step 3 is not a fixed value, but dynamically expands and contracts according to the ratio of the current system entropy value to the historical peak entropy value. When the system is in a high-risk state, it automatically expands to enhance its ability to absorb sudden traffic.

[0010] Preferably, the adjustment of task allocation weights in step 4 adopts a nonlinear mapping function, which makes the task increment obtained by nodes with lower system entropy values ​​grow exponentially, accelerating the convergence process of the system from disorder to order.

[0011] Preferably, in step 5, the physical damping mechanism suppresses ineffective oscillations of tasks between neighboring nodes by setting an upper limit on the number of migrations of the same task within a continuous scheduling cycle and setting a minimum interval for the distance between target nodes in two adjacent scheduling cycles.

[0012] Preferably, the graded degradation strategy in step 6 includes three control levels: the first level restricts the creation of low-priority background tasks, the second level suspends the execution of non-real-time interactive tasks, and the third level only retains basic tasks that maintain the system heartbeat and core communication links. Each level has independent triggering and recovery conditions.

[0013] Preferably, during the task pausing period, the virtual buffer pool performs lightweight preprocessing on the tasks, including task type identification, dependency resolution, and execution context compression, to reduce the computational overhead of subsequent scheduling decisions.

[0014] Preferably, a time decay factor is introduced in the calculation of the system entropy value, so that the influence of historical states on the current entropy value decays exponentially over time, highlighting the dominant role of recent states in the system disorder.

[0015] Preferably, each node in the Agent cluster is deployed with a local entropy proxy module, which is used to collect local status in real time and participate in the distributed calculation of global entropy, avoiding single-point bottlenecks caused by centralized monitoring.

[0016] Preferably, the physical damping mechanism works in conjunction with the virtual inertia control module to synchronously adjust the damping coefficient when the virtual buffer pool releases the task, so that the release rate is dynamically matched with the bearing capacity of the downstream node.

[0017] Preferably, the recovery process of the graded degradation strategy adopts a gradual rollback mechanism, that is, after the system entropy value has been continuously lower than the safety threshold for a predetermined time, the task restrictions are gradually lifted to prevent secondary impact caused by premature restoration of all functions.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a system state perception model by introducing thermodynamic entropy change theory, which realizes a forward-looking quantitative assessment of the disorder of the Agent cluster and overcomes the shortcomings of traditional flow limiting and degradation strategies with lag in response. By drawing on the power grid virtual inertia technology to construct a virtual buffer pool, it provides a critical time buffer window under sudden flow impact, blocking the path of single-point failure to spread to the entire cluster.

[0019] 2. This invention reduces the jitter effect in the dynamic load balancing process by integrating physical damping mechanisms to constrain task migration behavior, thereby improving the smoothness of the scheduling process and the robustness of the system. Through the combination of tiered degradation strategies and progressive recovery mechanisms, it ensures the continuity of core business while avoiding secondary oscillations in the system during the recovery phase. The overall solution deeply integrates physical principles with software engineering practices, forming a new fault-tolerant control system that combines predictability, buffering, and stability, suitable for high-concurrency, highly volatile, and complex Agent cluster application scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of the core principle framework of the scheduling early warning mechanism based on system entropy perception according to the present invention; Figure 3 This is a logical flowchart of the virtual inertia control module and the dynamic expansion and contraction of the virtual buffer pool according to the present invention. Figure 4 This is a flowchart illustrating the task migration scheduling logic framework that integrates nonlinear weight adjustment and physical damping mechanisms according to the present invention. Figure 5 This is a flowchart illustrating the logical flow of the tiered degradation strategy and gradual recovery mechanism according to the present invention. Detailed Implementation

[0021] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0022] In the Agent cluster process scheduling and fault-tolerant control method for complex scenarios provided by this invention, the core of the entire control logic lies in the deep cross-border integration of thermodynamic entropy change theory in physics and virtual inertia control technology in power systems, so as to provide Agent clusters with a scheduling system that is predictive, buffering and stable when facing high concurrency and strong fluctuations in traffic.

[0023] In the aforementioned method for agent cluster process scheduling and fault tolerance control for complex scenarios, step 1 specifically involves: constructing a system state awareness model for the agent cluster. The system first performs comprehensive state monitoring on each independent execution unit (Agent node) within the agent cluster. These basic state variables cover various dimensions of computing resources, including but not limited to real-time load data such as CPU utilization, memory usage, network bandwidth utilization, and disk I / O operation frequency. The system also needs to obtain the task queue length of each Agent node in real time, i.e., the number of tasks waiting to be executed, and the response latency, i.e., the time taken for a task to complete execution and return a result after entering the queue.

[0024] Through the aforementioned multi-dimensional basic state variables, the system constructs a multi-dimensional state vector. To eliminate differences between different physical dimensions and make variables of different properties comparable, the calculation of the system entropy value in step 1 uses the normalized result of the multi-dimensional state variables. For each basic state variable, the system uses the difference between its current real-time value and its historical minimum value, and then divides it by the difference between its historical maximum value and its historical minimum value, mapping all variables to the interval between 0 and 1.

[0025] After normalization, each state variable is assigned a different weight coefficient based on its impact on system stability. For example, in computationally intensive task scenarios, CPU utilization and response latency are assigned higher weight coefficients, while in data transmission intensive scenarios, network bandwidth utilization takes precedence.

[0026] These weighting coefficients are not fixed but are derived through deep learning training on historical failure data of the Agent cluster. The system analyzes the fluctuation characteristics of various state variables before historical failures to identify which variables' abnormal fluctuations are most likely to cause system crashes and dynamically adjusts the weights. These weighting coefficients are also updated online periodically to adapt to the evolution of business logic and application scenarios. After weight allocation, the system multiplies each normalized state variable by its corresponding weighting coefficient and sums the results to generate a system entropy value that characterizes the current level of chaos or disorder in the system.

[0027] To improve the sensitivity and accuracy of perception, this embodiment also introduces a time decay factor in the calculation of the system entropy value. When calculating the entropy value at the current moment, the system combines the historical states from multiple previous sampling moments. The closer the historical state is to the current moment, the greater its influence on the current entropy value; the farther away the historical state is, the more exponentially its weight decays over time. This design allows the system to more sensitively capture recent state change trends, prevent interference from outdated historical data, and more accurately reflect the dominant state of system disorder. To avoid single-point performance bottlenecks caused by centralized monitoring, each node in the Agent cluster deploys a local entropy proxy module. This local entropy proxy module is responsible for collecting local basic state variables in real time and using a distributed consensus algorithm to participate in the parallel calculation of the global entropy value, ensuring real-time performance and reliability in a large-scale cluster environment.

[0028] In the aforementioned Agent cluster process scheduling and fault-tolerant control method for complex scenarios, step 2 specifically involves: determining whether there is entropy increase caused by local node anomalies based on the trend of system entropy value changes. After obtaining a continuous sequence of system entropy values, the system uses a differential analysis method within a sliding time window to determine the rate of entropy change. The system sets a time window of fixed length, which slides forward as the sampling time progresses. Within each window period, the system calculates the rate of change of entropy value relative to time, i.e., the difference between the current entropy value and the entropy value at the beginning of the window, divided by the corresponding time span. When the calculated rate of change of entropy value exceeds a preset threshold, the system determines that an abnormal entropy increase has occurred, meaning that the orderliness within the cluster is rapidly decreasing, which is highly likely to trigger a systemic avalanche effect. At this time, a scheduling early warning mechanism is immediately triggered.

[0029] It should be noted that the sliding time window length in step 2 is not fixed, but dynamically adjusted according to the current size of the cluster. When the number of cluster nodes increases, the system automatically extends the window length to smooth data noise and improve decision stability; when the cluster size is small, the system shortens the window length to improve response speed. This dynamic adjustment mechanism ensures that the inflection point in the entropy increase process—the key trigger point of the transition from order to disorder—can be accurately captured under different load scales and cluster densities. Once the warning mechanism is triggered, the system immediately enters a defense preparation state, no longer relying solely on traditional passive rate limiting, but actively activating subsequent virtual inertia control logic.

[0030] In the aforementioned Agent cluster process scheduling and fault-tolerant control method for complex scenarios, step 3 specifically involves activating the virtual inertia control module after the scheduling early warning mechanism is triggered. This virtual inertia control module draws on the principle of maintaining grid frequency stability using rotational inertia in power systems, dynamically injecting a virtual buffer pool into the task scheduling path. The virtual buffer pool, acting as an intermediate layer, has the core function of temporarily handling sudden, extremely high-concurrency traffic far exceeding the cluster's processing capacity. By temporarily storing newly arrived task requests in the virtual buffer pool, the system can delay the direct transmission of this traffic to downstream Agent nodes, freeing up critical processing buffer time for downstream nodes and preventing instantaneous high-voltage surges from causing direct node downtime.

[0031] In this embodiment, the capacity of the virtual buffer pool in step 3 is highly dynamic, rather than a statically allocated storage space. Its capacity dynamically scales based on the ratio of the currently observed real-time system entropy value to the historical peak entropy value during system operation. When the system entropy value rapidly approaches the historical peak or exceeds a safety threshold, the virtual inertia control module automatically expands the logical capacity of the virtual buffer pool to enhance its ability to absorb and handle sudden large traffic spikes. During task caching, the virtual buffer pool does not simply store data but performs lightweight preprocessing on the tasks. This includes: quickly identifying task types, resolving dependencies between tasks using feature matching technology, compressing the execution context, and removing redundant information from task request packets. This preprocessing mechanism reduces the computational overhead of subsequent scheduling decisions and also reduces bandwidth consumption during task transmission within the cluster.

[0032] In the aforementioned Agent cluster process scheduling and fault-tolerant control method for complex scenarios, step 4 specifically involves adjusting task allocation weights based on the real-time feedback of the virtual buffer pool's fill status and system entropy. When distributing tasks from the virtual buffer pool to backend Agent nodes, the system no longer uses traditional round-robin or random allocation methods, but instead employs a weight adjustment algorithm based on a nonlinear mapping function. The system calculates the available resource margin and local entropy value of each target Agent node in real time and inputs these parameters into the nonlinear mapping function. The design logic of this nonlinear mapping function ensures that nodes with lower system entropy and more abundant resource margins receive exponentially increasing allocation weights; conversely, nodes with higher entropy or near-saturated resources have their allocation weights rapidly reduced.

[0033] Through this non-linear adjustment, newly arriving tasks are preferentially routed to the most stable and healthy nodes, accelerating the system's convergence from its current disordered state to an ordered state. Simultaneously, the scheduler continuously monitors the virtual buffer pool's fill ratio. When the fill ratio exceeds a warning level, the system further compresses the allocation window through a feedback mechanism, forcibly reducing the rate of task transmission downstream until the system entropy shows a significant downward trend. This closed-loop feedback control mechanism ensures that, under extreme traffic pressure, scheduling decisions always guide the cluster towards a higher degree of order.

[0034] In the aforementioned agent cluster process scheduling and fault tolerance control method for complex scenarios, step 5, during the release of tasks from the virtual buffer pool, introduces a physical damping mechanism. To prevent the jitter effect caused by frequent task rescheduling leading to repeated migrations between different nodes during dynamic load balancing, this invention imposes dual constraints on the task migration frequency and migration amplitude. The physical damping mechanism first limits the migration behavior of the same task within a continuous scheduling cycle by setting an upper limit on the number of migrations. The system allocates a migration counter to each task. Once the number of times a task is rescheduled from one node to another reaches the preset upper limit, the task will be locked on the current node or forcibly returned to the virtual buffer pool, strictly prohibiting endless secondary jumps.

[0035] Furthermore, the physical damping mechanism sets a minimum interval requirement for the distance between target nodes in two adjacent scheduling operations. This distance can refer to the number of hops in the network topology or the logical communication delay between nodes. If the logical distance between the next potential target node and the current node is less than the preset minimum interval threshold, the system will determine that this migration is highly likely to cause local network congestion or invalid oscillations, and suppress the migration behavior. The physical damping mechanism works collaboratively with the virtual inertia control module.

[0036] When releasing tasks from the virtual buffer pool, the system dynamically adjusts the damping coefficient based on the real-time capacity of downstream nodes. When a node's capacity is low, the damping coefficient is increased to slow down the frequency of task releases and migrations; when a node finishes processing and the load decreases, the damping coefficient is decreased accordingly to restore scheduling sensitivity. This control method, which simulates physical damping, improves the smoothness of the scheduling process and the overall robustness of the system.

[0037] In the aforementioned Agent cluster process scheduling and fault-tolerant control method for complex scenarios, step 6 specifically involves: continuously monitoring whether the system entropy value has fallen back to a stable range. If the system entropy value fails to recover below the safety threshold within a predetermined observation time, it indicates that the current virtual inertia control and weight adjustment are insufficient to cope with the current traffic volume. At this time, the system will initiate a tiered degradation strategy. The tiered degradation strategy in this embodiment includes three rigorous control levels.

[0038] The first level of degradation control primarily targets low-priority background tasks, such as non-real-time data statistics and historical log compression, restricting the creation of new requests for these tasks. The second level of degradation control further suspends all non-real-time interactive business logic, maintaining only synchronous calls to core user services. The highest level, the third level of degradation control, enters a minimalist operating mode, retaining only heartbeat messages, core communication links, and the highest level of security protection tasks necessary to maintain basic system survival.

[0039] Each degradation level has independent trigger and recovery conditions. The system continuously evaluates the priority weight of various services and restricts the scheduling permissions of non-core tasks step by step according to the preset priority tiers to ensure the continuous operation of critical business links to the greatest extent. When the system entropy value finally begins to consistently fall below the safety threshold for a predetermined duration, the system will not immediately fully restore all functions, but will adopt a gradual rollback mechanism. The system will first lift the third-level restriction, and after observing for a period of time, if the entropy value remains stable, it will then gradually lift the second-level and first-level restrictions. This phased and rhythmic recovery process can prevent secondary traffic surges caused by premature full restoration of functions, ensuring that the system can smoothly return to its initial operating state after experiencing large fluctuations.

[0040] Example 2: Building upon Example 1, Example 2 further refines the Agent cluster scheduling implementation scheme in a heterogeneous hardware environment. Considering that in actual deployments, the nodes in the Agent cluster may be deployed on servers with different configurations, the differences in hardware specifications can lead to different system entropy effects for the same load value.

[0041] In step 1 of Example 2, the system introduces an additional hardware benchmark correction factor when performing multidimensional state variable normalization. The system weights and corrects the collected real-time load and response latency based on the hardware benchmark scores or computing power ratings of each node. For nodes with stronger computing power, the same CPU utilization value is assigned a smaller sensitivity coefficient when calculating the entropy value, allowing the system entropy value to reflect the true stress situation across hardware platforms.

[0042] In step 3 of Embodiment 2, a predictive pre-expansion mechanism is introduced for the dynamic scaling logic of the virtual buffer pool. The system establishes a predictive model based on a long short-term memory network to anticipate future flow trends within the sliding window. If the predictive model output indicates a sustained entropy increase trend within the next few seconds, the virtual inertia control module will pre-adjust the capacity of the virtual buffer pool and inject a certain amount of damping force, even if the current entropy value has not yet triggered the warning threshold. This proactive action can consume sudden energy earlier, making the curve of the entire control process smoother.

[0043] In step 5 of Embodiment 2, regarding the implementation of the physical damping mechanism, this embodiment adds a deep constraint on task data consistency. During task migration, the physical damping module checks the consistency checksum of the task execution status. If the data synchronization cost during migration exceeds a preset benefit threshold, the damping mechanism will forcibly intervene, even if the migration frequency constraint is met, by increasing "logical stickiness" to prevent the task from leaving the original node. This cost-effective damping adjustment further optimizes the overall performance of the system in dynamic scheduling.

[0044] In step 6 of Example 2, the gradual recovery mechanism is given more intelligent stepping logic. The system automatically adjusts the observation period for lifting restrictions based on the historical recovery success rate. If, in the historical record, the recovery of a certain type of service has repeatedly caused a secondary rebound in entropy, then in the current recovery process, the system will automatically extend the silent observation period for that service level and adopt a smaller traffic release ratio. Through this adaptive gradual recovery, the system can achieve a higher level of intelligent fault-tolerant control.

[0045] Example 3: This example provides a special implementation method for Agent clusters in edge computing scenarios. In edge computing environments, due to bandwidth limitations and more frequent network fluctuations, the system entropy perception model in step 1 adopts a local priority principle during distributed computing.

[0046] Each edge agent node first calculates its local entropy value locally. Only when the change in local entropy exceeds a certain range will it synchronize and update information to the regional central node. This approach reduces the management traffic required to maintain global entropy awareness. In the entropy increase determination step 2, the system introduces geographical correlation analysis. When an abnormal entropy increase occurs in a certain edge region, the early warning mechanism automatically scans the surrounding regions geographically or network-routingly adjacent to it, and pre-activates low-intensity virtual inertia control in the scheduling paths of these adjacent regions, forming a "defense circle" effect to prevent the fault from spreading to adjacent regions.

[0047] In step 3 of embodiment 3, the virtual buffer pool is designed with a multi-level structure. In addition to the local micro-buffer pools on each node, a regional-level shared virtual buffer pool is also provided. When a sudden overload occurs on an edge node, tasks are first pushed into the local buffer pool. If the local buffer pool reaches saturation, tasks overflow to the regional-level shared buffer pool via a high-speed dedicated line link. This multi-level buffer system provides stronger fault tolerance for uncertainties in edge environments.

[0048] In step 5 of Embodiment 3, the physical damping mechanism employs a calculation logic based on an energy loss model. The system treats each task migration as an energy loss process, the magnitude of which is proportional to the amount of data carried by the task and the physical distance of the migration. The physical damper accumulates the total energy loss per unit time in real time. When the total energy loss value reaches a preset safety limit, the system will forcibly stop all cross-regional task migrations by increasing the damping coefficient, ensuring the basic connectivity of the edge network.

[0049] In step 6 of Example 3, the tiered degradation strategy enhances the protection of local autonomy. During the three-level degradation process, even under the highest level of restrictions, edge agent nodes retain the authority to process the most basic local sensor signals, without relying entirely on cloud feedback. This design ensures that even in extreme network isolation situations, the edge cluster can maintain a basic controlled state, and after the network recovers, it can reconnect to the global scheduling system through a gradual fallback mechanism.

[0050] To further illustrate the technical logic of this invention, the mathematical processing involved in the above method is described in plain text below.

[0051] In the process of generating the system entropy value, the system first performs linear scaling on each state dimension. For any state observation, its normalized value is equal to the difference between the state observation and the preset minimum boundary value, divided by the difference between the preset maximum boundary value and the minimum boundary value. The system calculates a weighted sum, that is, multiplying each normalized value by its corresponding assigned weight coefficient and accumulating all the product terms. Based on this, to reflect the temporal correlation, the system multiplies the current accumulated result by a decay constant between 0 and 1, and sums it with the remainder of the entropy value result from the previous time step multiplied by the decay constant, to obtain the final smoothed entropy value.

[0052] When determining the entropy growth rate, the system performs a time-domain slope calculation. Specifically, it selects the entropy value at the current sampling moment as the endpoint value and the entropy value at the start of the sliding window as the starting point value. The difference between the endpoint value and the starting point value is calculated. If the difference is positive and the ratio of the difference to the window duration is greater than a preset critical rate of change, the logical judgment result is true, and the subsequent defense process is immediately triggered.

[0053] In the virtual buffer pool's capacity scaling logic, the target capacity is calculated by multiplying the system's default base capacity value by a scaling factor. This scaling factor is determined by dividing the current real-time entropy value by the historical highest entropy value. To ensure that the scaling process is not too drastic, the system performs a non-linear smoothing process on this quotient, for example, by using the quotient as an exponent for power operations. This results in slow capacity expansion when the entropy value is low, and rapid capacity expansion when the entropy value is close to the critical point.

[0054] During the task allocation weight adjustment process, the implementation logic of the nonlinear mapping function is as follows: For any candidate Agent node, the system first calculates its resource sufficiency. This resource sufficiency is equal to 1 minus the node's normalized load value. The system calculates the target allocation weight for this node, and the calculation logic is as follows: take a preset base constant, use the node's resource sufficiency as the exponent for exponentiation, and divide the result by the sum of the exponentiation results of all nodes in the cluster. In this way, weight allocation achieves a nonlinear shift towards nodes with high resource sufficiency.

[0055] In the mathematical logic of the physical damping mechanism, the dynamic adjustment of the damping coefficient follows this rule: the damping coefficient of the current period equals the initial damping coefficient plus a correction term. The magnitude of this correction term is positively correlated with the current fill rate of the virtual buffer pool. Subtracting the safety water level threshold from the fill rate, if the difference is positive, multiplies this difference by a preset proportional adjustment coefficient and adds it to the initial damping coefficient. In the constraint logic of task migration, the system calculates the shortest path length between the current node and the target node in the network topology graph. If this shortest path length is less than a preset resistance radius, the probability of the migration request taking effect is forcibly set to zero.

[0056] In the triggering judgment of tiered degradation, the system adopts a duration accumulation logic. That is, the system starts a timer, and when the system entropy value remains in the high range, the timer increments the time value. When the timer value successively reaches the first preset duration, the second preset duration, and the third preset duration, the first, second, and third level degradation strategies are triggered respectively. The recovery process is the opposite. The system requires that the entropy value must be below the safety threshold for a longer cooling-off period before a step-by-step rollback operation can be executed.

[0057] Through the aforementioned refined logical design, this invention achieves a highly biomimetic fault-tolerant control scheme with physical characteristics. Compared to traditional simple threshold triggering mechanisms, this scheme improves the survivability and business continuity of the Agent cluster under extremely complex conditions through entropy-aware "soft triggering" and virtual inertia "soft buffering."

[0058] In practice, each layer of the Agent cluster works closely together. The bottom-level local entropy agent module periodically reports its status to the upper-level scheduler via lightweight heartbeat messages. The message format uses compact binary encoding and includes a node identifier, a timestamp sequence of multi-dimensional state vectors, and a locally preprocessed fragment of the initial entropy value. The middle-level virtual inertia control module maintains a high-performance task mapping table, recording the entry time, priority label, dependency topology, and preprocessed context index of each task in the virtual buffer pool.

[0059] When a traffic spike occurs, the scheduler's main control logic intervenes rapidly. It first performs a differential calculation of the global entropy value within a nanosecond timescale. Once it determines that the entropy increase has exceeded the limit, the logic valves of the virtual buffer pool immediately switch. Newly arriving request data streams are redirected to a temporary storage area composed of high-speed random access memory. Within this storage area, tasks are reordered according to priority and expected execution duration. For tasks with extremely high computational intensity and relatively low real-time requirements, the system deeply compresses their execution context, retaining only the seed parameters necessary to trigger execution.

[0060] During the task release phase, the physical damper monitors the traffic density of each exit path in real time. If an increase in packet loss rate or latency fluctuation is detected in a certain network segment, the damper will immediately adjust the damping coefficient to limit the task release rate to that network segment. This adjustment process is smooth and continuous, avoiding system oscillations caused by "all on" or "all off" switching in traditional scheduling.

[0061] To ensure the fairness and scientific rigor of the tiered degradation strategy, the system maintains a business importance score table. When executing a degradation operation, the system assesses the negative impact on the entire business system in real time based on the accumulated scores of currently restricted tasks. If the business value loss caused by a degradation decision exceeds the set tolerance limit, the system automatically attempts to find alternative fault-tolerant paths, such as cross-datacenter resource scheduling or requesting higher-level resource quotas, achieving a dynamic balance between ensuring system security and maintaining business value.

[0062] During the system recovery phase, the gradual rollback mechanism operates with very rigorous steps. The system first sends a small amount of probe traffic to the restricted task categories and simultaneously activates a high-frequency entropy monitoring window. If, within a predetermined period after the probe traffic is released, the system entropy fluctuation remains within a minimal noise range, it indicates that the throughput capacity of downstream nodes has substantially recovered. At this point, the system will gradually reopen service functions in a preset increment, such as increasing the allowance by 10% each time. At each step, the system performs a complete stability check; only after passing checks for multiple consecutive periods will it proceed to the next release phase.

[0063] This fault-tolerant control method based on physical principles is applicable not only to general-purpose agent clusters on an internet scale, but also to specialized and complex scenarios with extremely high reliability requirements, such as industrial control, smart grid scheduling, and autonomous driving collaborative processing. Its core advantage lies in transforming originally uncontrollable random bursts of traffic into a controlled data stream that is predictable, bufferable, and smoothly processed under the constraints of a physical model, fundamentally eliminating the physical conditions that allow cascading failures to occur.

[0064] In summary, this invention constructs a comprehensive, multi-layered protection system through the deep synergy of six core steps: system entropy perception, sliding window differential judgment, dynamic scaling of the virtual buffer pool, nonlinear mapping weight adjustment, physical damping frequency constraints, and graded degradation-based gradual recovery. This protection system can adaptively cope with various unknown and complex traffic patterns, ensuring that the Agent cluster maintains a high degree of order and business stability even under extreme pressure, demonstrating strong engineering practicality and technological foresight.

[0065] Finally, it should be noted that the various functional modules in this invention can be deployed on physical server clusters or in cloud-native environments based on containerization technology. In containerized scenarios, the basic state variables also include container quota utilization, inter-container communication latency, and image repository access speed. The virtual buffer pool can be implemented using a high-performance distributed caching system, ensuring its horizontal scalability during expansion. This flexible deployment method enables this invention to be widely compatible with existing mainstream software architectures, giving it high potential for widespread application.

[0066] In the control processes described in Examples 1 to 3, all parameter settings, threshold selections, and initial values ​​of weighting coefficients were optimized after simulating thousands of different types of traffic fluctuation models in a large number of laboratory simulation environments. In practical applications, the system also has self-learning and evolution capabilities, and can automatically fine-tune the internal physical model parameters according to the specific traffic patterns of different industries, making the scheduling strategy increasingly aligned with the needs of actual business operations, and achieving true intelligent autonomous fault tolerance.

[0067] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for agent cluster process scheduling and fault tolerance control in complex scenarios, characterized in that, Includes the following steps: Step 1: Construct a system state awareness model for the Agent cluster, collect real-time load, task queue length, response latency, and resource utilization of each Agent node, and perform weighted fusion processing to generate a system entropy value that characterizes the overall orderliness of the system. Step 2: Based on the trend of the system entropy, identify the entropy increase phenomenon caused by local node anomalies by judging the rate of entropy change, and trigger the scheduling early warning mechanism when the rate of entropy change exceeds a preset threshold; Step 3: After the scheduling early warning mechanism is triggered, the virtual inertia control module is activated to dynamically inject a virtual buffer pool with dynamic capacity into the task scheduling path to handle sudden traffic and delay the transmission of traffic to the travel node. Step 4: Adjust the task allocation weight based on the real-time feedback of the virtual buffer pool's fill status and the system entropy value, and prioritize routing newly arriving tasks to Agent nodes where the system entropy value is below the safety threshold and has sufficient resource reserves; Step 5: During the process of releasing tasks from the virtual buffer pool, a physical damping mechanism is introduced to suppress task jitter between nodes by applying dual constraints on the task migration frequency and migration amplitude. Step 6: Continuously monitor the system entropy value. If the system entropy value does not fall back to the stable range within a predetermined time, then activate the graded degradation strategy and restrict the scheduling permissions of non-core tasks level by level according to preset priorities.

2. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, The process of generating the system entropy value in step 1 includes: Normalization is performed on the real-time load, task queue length, response latency and resource utilization rate, mapping each state variable to a range between 0 and 1, where the normalized value is equal to the difference between the current real-time value and the historical minimum value divided by the difference between the historical maximum value and the historical minimum value. Different weighting coefficients are assigned to each state variable based on its impact on system stability. These weighting coefficients are obtained by performing deep learning training on historical fault data of the Agent cluster and are updated online. The process of calculating the entropy value of the system introduces a time decay factor. By combining the entropy value calculation result at the current moment with the historical state at the previous sampling moment, the influence weight of the historical state on the current entropy value decays exponentially over time, and the historical state closer to the current moment has a greater weight. Each node in the Agent cluster deploys a local entropy proxy module, which collects local state variables in real time and uses a distributed consensus algorithm to participate in the parallel calculation of global entropy, thus avoiding the performance bottleneck of centralized monitoring.

3. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, The specific process for determining the rate of change of entropy in step 2 includes: A fixed-length time window is set and slides forward as the sampling time progresses. The length of the time window is dynamically adjusted according to the number of nodes in the Agent cluster. The length of the time window is extended when the number of nodes increases and shortened when the number of nodes decreases. Within each window period, the rate of change of the system entropy with respect to time is calculated. Specifically, the difference between the entropy value at the current moment and the entropy value at the beginning of the time window is calculated, and then the difference is divided by the corresponding time span to obtain the rate of change of the entropy value. The rate of change of the entropy value is compared with a preset critical rate of change. If the rate of change of the entropy value is greater than the critical rate of change, the logical judgment result is true, and the subsequent defense process is immediately initiated.

4. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, The process of dynamically injecting the virtual buffer pool in step 3 includes: Obtain the current hardware benchmark score or computing power rating of the Agent cluster, and introduce a hardware benchmark correction factor to perform weighted correction on the collected real-time load and response latency; Calculate the target capacity of the virtual buffer pool, which is obtained by multiplying the base capacity value by the scaling factor, and the scaling factor is determined by dividing the current system entropy value by the historical highest entropy value; The scaling factor is subjected to nonlinear smoothing, and the quotient is used as an exponent for power operation, so that the capacity expansion rate of the virtual buffer pool is low when the system entropy value is low, and the capacity expansion rate of the virtual buffer pool is high when the system entropy value reaches the critical point. A prediction model based on a long short-term memory network is established to predict the flow trend within the future sliding window. When the prediction results show a continuous entropy increase trend, the capacity of the virtual buffer pool is adjusted in advance and damping force is injected.

5. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, Step 3 further includes performing lightweight preprocessing on the tasks temporarily stored in the virtual buffer pool, specifically including: Feature matching technology is used to identify the task type of the task and to parse the dependency topology between the tasks; The execution context of the task is compressed to remove redundant information from the task request packet, retaining only the seed parameters required to trigger execution, thereby reducing the computational overhead and bandwidth usage of subsequent scheduling decisions.

6. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, The specific process of adjusting task allocation weights in step 4 includes: Calculate the resource sufficiency of each candidate Agent node in the Agent cluster, where the resource sufficiency is equal to one minus the normalized load value of the corresponding node. The target allocation weight of each candidate Agent node is calculated using a nonlinear mapping function. Specifically, a preset base constant is selected, the resource sufficiency is used as the exponent for exponentiation, and the result is divided by the sum of the exponentiation results of all nodes in the cluster to achieve a nonlinear shift of weight towards nodes with high resource sufficiency. The system monitors the filling ratio of the virtual buffer pool in real time. When the filling ratio exceeds the preset warning level, the allocation window is compressed through a feedback mechanism to reduce the rate of sending tasks downstream until the system entropy value drops.

7. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, The specific process of implementing the dual constraints in step 5 includes: Each task is assigned a migration counter to record the number of migrations of the task within a continuous scheduling cycle. Once the number of migrations reaches the preset upper limit, the task is locked at the current node or forcibly returned to the virtual buffer pool. Set a minimum interval threshold between target nodes in two adjacent scheduling operations, where the minimum interval threshold includes the number of hops in the network topology or logical communication delay; Calculate the shortest path length between the current node and the target node in the network topology graph. If the shortest path length is less than the preset resistance radius, it is determined that this migration will cause local network congestion, and the migration behavior is forcibly stopped. During the task migration process, the consistency checksum of the task execution status is checked, and the data synchronization cost during the migration process is calculated. If the data synchronization cost exceeds the preset benefit threshold, logical stickiness is added to prevent the task from leaving the original node.

8. The method for agent cluster process scheduling and fault tolerance control in complex scenarios according to claim 1, characterized in that, Step 5 also includes the collaborative working process of the physical damping mechanism and the virtual inertia control module: Real-time monitoring of packet loss rate and latency fluctuation rate of each exit path, and dynamic adjustment of damping coefficient based on the real-time carrying capacity of downstream nodes; Calculate the damping coefficient for the current cycle. The damping coefficient for the current cycle is equal to the initial damping coefficient plus a correction term. The magnitude of the correction term is obtained by multiplying the difference between the current filling rate of the virtual buffer pool and the safe water level threshold by a preset proportional adjustment coefficient. A computational logic based on an energy loss model is established, treating each task migration as an energy loss process. The magnitude of the energy loss is proportional to the amount of data carried by the task and the physical distance of the migration. When the total energy loss per unit time reaches a preset safety limit, the cross-regional task migration is forcibly stopped by increasing the damping coefficient.

9. The method for agent cluster process scheduling and fault tolerance control for complex scenarios according to claim 1, characterized in that, The specific execution process of the tiered degradation strategy in step 6 includes: Establish a degradation system with three control levels and maintain a business importance score table to assess the negative impact on business systems when performing degradation operations; Start a timer to accumulate the time value when the system entropy value is in the high range. When the timer value reaches the first preset duration, trigger the first level of degradation control to restrict the creation of new requests for low-priority background tasks. When the timer value reaches the second preset duration, trigger the second level of degradation control to suspend all non-real-time interactive business logic and only maintain the synchronous call of the user's core business. When the timer reaches the third preset duration, the highest level of third-level degradation control is triggered, entering a simplified operation mode, which retains only the heartbeat messages to maintain system survival, core communication links, and the highest level of security protection tasks.

10. The method for agent cluster process scheduling and fault tolerance control for complex scenarios according to claim 1, characterized in that, Step 6 also includes a gradual rollback process after the system entropy value is restored: After the system entropy value remains below the safety threshold and reaches the preset cooling-off period value, a small amount of probe traffic is sent to the restricted task category, and a high-frequency entropy monitoring window is opened. If, within a predetermined period after the probe flow is released, the fluctuation range of the system entropy value is detected to be within a preset noise range, it is determined that the throughput capacity of the downstream node has been restored. The task restrictions are gradually lifted and the traffic quota is released in a preset step ratio. Stability checks are performed at each step. Only after multiple consecutive cycles of checks pass can the next release stage be entered. The observation period for lifting restrictions is automatically adjusted based on the historical recovery success rate. If the recovery of a business in the historical record has caused the entropy value to rebound, the silent observation period for that business level will be automatically extended, and a smaller traffic release ratio will be adopted.

Images