Autonomous task allocation and dynamic optimization system and method based on agent cooperation

Abstract

The application relates to the technical field of intelligent agent cooperative task allocation, and particularly discloses an autonomous task allocation and dynamic optimization system and method based on intelligent agent cooperation, which constructs a dynamic perception basis by acquiring migration historical trajectory data of a target memory page within a preset time window, access feature vectors of each computing node, and a bus delay matrix between nodes. A shock kinetic energy value is calculated according to the migration history, a cooperative damping instruction is generated when the value exceeds a preset threshold, and a virtual mass parameter is obtained by fusing a random backoff factor. The virtual mass parameter and the bus delay matrix are combined to calculate a migration power consumption threshold, and a gravitational potential difference value is calculated based on the access feature vectors, so that a livelock state is accurately determined through numerical comparison. Finally, a memory locking instruction is generated to prohibit migration and trigger a task migration signal, effectively solving the ping-pong migration livelock problem under the NUMA architecture, and significantly improving the system throughput and resource utilization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent cooperative task allocation technology, and relates to an autonomous task allocation and dynamic optimization system and method based on intelligent agent cooperation. Background Technology

[0002] In non-uniform memory access architectures, the asymmetry of memory access latency is one of the core factors affecting the performance of multi-core processor systems. Each CPU core accesses its local memory node much faster than it accesses a remote node. The operating system's memory management unit optimizes this access locality by dynamically migrating memory pages, aiming to place data near the cores accessed most frequently. This mechanism plays an irreplaceable role in maintaining cache coherency and reducing the load on the cross-node interconnect bus, directly affecting the processing efficiency of large-scale parallel computing tasks and the overall throughput of the system.

[0003] When computationally intensive tasks on multiple different nodes simultaneously compete for access to the same shared memory page, the operating system's automatic migration mechanism may fall into a peculiar competitive deadlock. Specifically, node A migrates a memory page to its local machine due to frequent access, then node B experiences a local missing memory page due to its own access requests, and migrates the page back to its own node, and so on, forming a "migrating ping-pong." This two-way feedback loop caused by resource contention appears to be the operating system dynamically responding to changes in access patterns, but in reality, it has evolved into an inefficient form of system friction: memory pages move frequently between two physical nodes, consuming the bandwidth of the interconnect bus, while the actual computational instructions are stalled due to waiting for data consistency synchronization.

[0004] However, traditional migration strategies are mostly based on simple counts of access frequency or threshold triggers, relying excessively on fixed frequency statistics to determine page ownership, while ignoring the crucial dimension of the "disorder" of access patterns. When the access distribution of a page across different nodes tends to be uniform and high-frequency, existing mechanisms cannot distinguish between benign data sharing and malicious resource contention, and mechanically executing migration instructions only exacerbates bus congestion. This lagging and coarse response mode means that the system can only passively remedy the situation after "migration oscillations" occur, and cannot predict and intervene in the early stages of hidden cost accumulation. This causes the data bus, which should be serving computation, to be occupied by a large amount of consistency maintenance traffic, ultimately resulting in CPU core idling, increased power consumption, and a precipitous drop in processing power. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides an autonomous task allocation and dynamic optimization system and method based on intelligent agent collaboration to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for autonomous task allocation and dynamic optimization based on agent cooperation, the method comprising: Obtain the migration history trajectory data of the target memory page within a preset time window, the access feature vector of each computing node, and the bus delay matrix between nodes; The oscillation kinetic energy value of the target memory page is calculated based on the migration history trajectory data, and a cooperative damping instruction is generated when the oscillation kinetic energy value exceeds the preset oscillation threshold. In response to the cooperative damping command, a random backoff factor is generated, and the random backoff factor is weighted and fused with the oscillation kinetic energy value to obtain the virtual mass parameters of the target memory page; The migration power consumption threshold is obtained by multiplying the virtual mass parameter with the bus delay matrix, and the gravitational potential energy difference between the current node and the requesting node is calculated based on the access feature vector. The gravitational potential energy difference is compared with the migration power consumption threshold. If the gravitational potential energy difference is less than the migration power consumption threshold, the current system is determined to be in a livelock state. A memory lock instruction is generated to prohibit the migration of the target memory page, and a task migration signal to schedule the computing task to the current node is generated based on the memory lock instruction.

[0007] A second aspect of the present invention provides an autonomous task allocation and dynamic optimization system based on agent cooperation, the system comprising: Damping instruction generation module: acquires the migration history trajectory data of the target memory page within a preset time window, the access feature vector of each computing node, and the bus delay matrix between nodes; calculates the oscillation kinetic energy value of the target memory page based on the migration history trajectory data, and generates a cooperative damping instruction when the oscillation kinetic energy value exceeds a preset oscillation threshold; Potential energy difference calculation module: In response to the cooperative damping command, it generates a random backoff factor and performs a weighted fusion calculation with the random backoff factor and the oscillation kinetic energy value to obtain the virtual mass parameter of the target memory page; it multiplies the virtual mass parameter with the bus delay matrix to obtain the migration power consumption threshold, and calculates the gravitational potential energy difference between the current node and the requesting node based on the access feature vector; Task migration scheduling module: It compares the gravitational potential energy difference with the migration power consumption threshold. If the gravitational potential energy difference is less than the migration power consumption threshold, it determines that the current system is in a livelock state, generates a memory lock instruction to prohibit the migration of the target memory page, and generates a task migration signal to schedule the computing task to the current node based on the memory lock instruction.

[0008] As described above, the autonomous task allocation and dynamic optimization system and method based on agent collaboration provided by the present invention has at least the following beneficial effects: This invention constructs a multi-dimensional dynamic data perception model by acquiring real-time migration history data of the target memory page within a preset time window, access feature vectors of each computing node, and bus delay matrices between nodes. Based on the migration history data, it quantifies and calculates the oscillation kinetic energy value of the target memory page, and then generates a cooperative damping instruction when the oscillation kinetic energy value exceeds a preset oscillation threshold. This approach overcomes the lag inherent in traditional memory management, which relies solely on static access counts for judgment. It can sensitively capture the unstable state of memory pages repeatedly jumping between nodes, intervene promptly, and effectively avoid the risks of invalid memory bus resource occupation and jitter caused by the system blindly responding to high-frequency access requests.

[0009] This invention, responding to a cooperative damping command, generates a random backoff factor and weights it with the oscillation kinetic energy value to obtain a virtual mass parameter. The migration power consumption threshold is determined by multiplying the virtual mass parameter with the bus delay matrix. Simultaneously, the gravitational potential energy difference between nodes is calculated based on the access feature vector, thus establishing a dynamic migration cost evaluation system based on physical field theory. This feature breaks the rigid pattern of fixed or only physically distance-related migration costs in existing technologies, endowing memory pages with a "mass" attribute that adaptively changes with oscillation intensity. The random backoff mechanism effectively disrupts the synchronization of multi-node contention, significantly reducing unnecessary consumption of system bus bandwidth while ensuring the advantage of data locality.

[0010] This invention achieves accurate determination of system livelock status by numerically comparing the gravitational potential energy difference with the migration power consumption threshold. Upon determining that livelock has occurred, a memory locking instruction is generated to prevent the migration of the target memory page, while simultaneously triggering a task migration signal to schedule the computation task to the current node. This logic cleverly transforms "data migration" into "task migration," completely solving the livelock problem caused by the "ping-pong migration" of memory pages under NUMA architecture. It not only frees up wasted bus bandwidth resources but also optimizes system execution efficiency through the re-adaptation of computation and data, significantly improving system throughput and operational stability in high-concurrency scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0012] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention.

[0013] Figure 2 This is a schematic diagram showing the connections of the various modules in the system of the present invention. Detailed Implementation

[0014] The following description, in conjunction with the implementation of the present invention, is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0015] In traditional NUMA-based memory management systems, fixed migration cost thresholds and static access counting rules are ill-suited to the dynamic competition environment of concurrent access in multi-core processors. When the access frequency of multiple compute nodes to the same memory page fluctuates non-linearly within a time window, the system cannot establish a dynamic correlation between migration decisions and system stability, leading to a mismatch between migration decision logic and real-time load competition. This static threshold mechanism induces a "ping-pong effect," causing memory pages to repeatedly jump between different nodes, resulting in bus bandwidth resources being occupied by ineffective migration traffic, ultimately affecting the execution efficiency of computing tasks and system throughput.

[0016] For example, in high-concurrency computing scenarios, two computing nodes frequently and alternately access the same target memory page within a short period of time, causing a surge in the number of page migrations within a millisecond-level time window, reaching a migration frequency of over 20 times per second. In this situation, traditional systems still use a preset access heat counter threshold to determine ownership, leading to frequent cross-node page migration instructions being incorrectly treated as legitimate operations. The operating system kernel cannot recognize this high-frequency oscillation pattern, continuously consuming valuable inter-chip interconnect bandwidth for data movement instead of performing actual computing tasks, resulting in a significant decrease in the system's effective computing power.

[0017] If these issues are not addressed, the repeated oscillations and migrations of memory pages will cause the system to fall into a livelock state, wasting a significant number of CPU cycles waiting for memory data synchronization, and in severe cases, even causing system responsiveness to stall. A rigid migration decision-making mechanism will hinder the system from capturing the critical state where contention access patterns transform into system performance bottlenecks, delaying the optimal timing for scheduling intervention. High-frequency cross-node data transfers will also cause cache coherency protocol traffic storms, crowding out communication resources for other normal tasks, ultimately creating a negative feedback loop that affects the stability of the entire computing node.

[0018] To address the aforementioned issues, this application first considers establishing a dynamic correlation mechanism between migration decisions and system stability. Traditional systems use a fixed cost to determine the necessity of migration, leading to page oscillations caused by intense competition. To resolve this, this application attempts to quantify historical migration trajectory data into oscillation kinetic energy values ​​and fuse them with a random backoff factor to generate virtual mass parameters, thereby dynamically adjusting the migration power consumption threshold. Further analysis reveals that relying solely on static bandwidth occupancy analysis is insufficient to accurately determine livelock states; therefore, it is necessary to introduce node access feature vectors to calculate gravitational potential energy differences and compare them with dynamically generated migration power consumption thresholds. By designing a linkage mechanism between cooperative damping instructions and task migration signals, the memory management strategy adaptively adjusts according to the intensity of system competition, thereby solving the problems of invalid bus resource occupation and low computational efficiency.

[0019] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Example

[0020] Please see Figure 1 As shown, the autonomous task allocation and dynamic optimization method based on agent cooperation includes the following steps: Obtaining the migration history data of the target memory page within a preset time window, the access feature vectors of each computing node, and the bus latency matrix between nodes specifically includes the following steps: The memory monitoring unit records the resident node identifiers of the target memory page in real time within a preset time window, and serializes the recorded multiple resident node identifiers in chronological order to generate migration history trajectory data of the target memory page's spatial location change process. Collect the access instruction count and cache miss count of each computing node for the target memory page, and then weight and combine the access instruction count and cache miss count to construct the access feature vector of each computing node. The signal transmission delay between the resident node currently holding the target memory page and other computing nodes is measured, and the signal transmission delay is filled into a grid data structure of a preset dimension to form a bus delay matrix between nodes.

[0021] In one specific embodiment, a memory monitoring unit at the operating system kernel level is used to poll the physical address mapping status of the target memory page at a preset sampling frequency, identify the physical computing node currently residing in the memory page, and record it in a time-node pair format. To capture the dynamic drift trend of the memory page within the system, a preset time window is set as an adjustable range of 500 milliseconds to 2000 milliseconds. The logic for setting this range is that if the window is less than 500 milliseconds, it is easily affected by transient noise, leading to data jitter; if it is greater than 2000 milliseconds, it will cause the system to lag in responding to sudden load changes. All the resident node identifiers collected within this window are serialized and arranged according to the chronological order of the time points to generate migration history trajectory data of the target memory page's spatial location change process.

[0022] The system collects access instruction counts and cache miss counts for the target memory page from each compute node. These counts are then weighted and combined to construct an access feature vector reflecting the load urgency of each compute node. The processor's performance monitoring unit then reads the total number of accesses to the page by each node within a preset time window. and the number of last-level cache misses Considering that the latency of remote memory access caused by a cache miss is much higher than the latency of a local cache hit, the two cannot be simply added together. Therefore, the following weighted combination formula is used to construct the access feature vector value of node j. : ; in, The base access weight, which defaults to 1, represents the load base brought by a standard hit; This is the penalty coefficient for a miss, calculated using the formula... It was calculated in advance that, among which For average remote memory access latency, Delayed access to local cache; To normalize the time factor, thus generating the access feature vector for each computing node. .

[0023] Finally, the signal transmission delay between the resident node currently holding the target memory page and other computing nodes is measured, and this delay is filled into a grid data structure of a preset dimension to form a bus delay matrix between nodes. The one-way physical transmission delay between node i and node j is obtained by sending lightweight probe packets to the system bus or directly reading SLIT data from the ACPI table. To eliminate random errors caused by network jitter, a moving average filtering algorithm is used to analyze the most recent five measurements. Perform smoothing. Build. Dimensional bus delay matrix , where matrix elements This represents the communication overhead required to migrate data from node i to node j.

[0024] ; The above formula is the verification formula for matrix elements, where For memory page size, This refers to the bandwidth of the link between nodes. Based on basic line delay, This is a bandwidth utilization correction factor, a dimensionless parameter with a value of 0-1, which is dynamically adjusted according to the bus congestion level.

[0025] The oscillation kinetic energy value of the target memory page is calculated based on the migration history trajectory data, and a cooperative damping instruction is generated when the oscillation kinetic energy value exceeds a preset oscillation threshold. The specific steps include: The migration history trajectory data is traversed in time series to filter out records with inconsistent resident node identifiers at adjacent times as valid migration events, and the corresponding time decay weight is determined based on the time span between the time of occurrence of the valid migration event and the current time. All selected valid migration events are weighted and summed according to time decay weights to obtain the oscillation kinetic energy value of the reciprocating movement intensity of the target memory page between computing nodes. The oscillation kinetic energy value is compared with the preset oscillation threshold. If the oscillation kinetic energy value is greater than the preset oscillation threshold, a coordinated damping command is generated.

[0026] Preferably, the corresponding time decay weight is determined based on the time span between the occurrence of the effective migration event and the current time, specifically including: Extract the generation time point when each valid migration event is recorded, calculate the time difference between the generation time point and the current system time, and obtain the time span value of the event; The time span value is mapped using a preset negative exponential decay logic, so that the smaller the time span value, the larger the mapped value. The value obtained after mapping is determined as the time decay weight corresponding to the effective migration event, and a one-to-one correspondence between the time decay weight and the effective migration event is established.

[0027] In one specific embodiment, migration history trajectory data A time-series traversal is performed to filter records where the identifiers of the resident nodes at adjacent time points are inconsistent, identifying them as valid migration events. The specific filtering logic is as follows: a traversal pointer k is set from the start position to the end position of the trajectory data, and the resident nodes at the k-th time point are compared. The dwell node at time point k-1 .like If a physical memory page transfer occurs within that time slot, it is determined to be the i-th valid migration event, and the time point of the k-th time point is extracted as the occurrence time of this event. .

[0028] Next, based on the time span between the occurrence of a valid migration event and the current time, a corresponding time decay weight is determined. All selected valid migration events are then weighted and summed according to this weight to obtain the oscillation kinetic energy value of the target memory page. The core idea is that migrations closer to the current time have a greater destructive impact on system stability and should therefore be assigned a higher weight; while earlier migrations have had their cache invalidation impact gradually diminish over time, so their weight should decay rapidly. Oscillation Kinetic Energy Value The calculation formula is constructed as follows: ; Where M is the total number of valid migration events detected within the preset time window; The current system time is in milliseconds. The time of occurrence of the i-th valid migration event is expressed in milliseconds. This refers to the time span, representing the degree of novelty or aging of an event; The basic impact coefficient is a dimensionless constant with a preset value of 1.0, representing the normalized intensity of a standard migration event at the moment of occurrence. It is the half-life constant, used to control the rate at which memories are forgotten, and is measured in milliseconds.

[0029] Regarding parameters The default logic is that it usually will The timeout period is set to between 100ms and 300ms. This range is determined based on the context switching overhead and cache warm-up period for cross-NUMA node access in modern CPUs. If the load is too high, the system will become overly resentful, causing long-subsided oscillations to still limit normal load balancing; if... If the value is too small, the system will be insensitive to high-frequency oscillations. Optimal setting. =200ms.

[0030] The final calculation result is a dimensionless scalar value, which physically represents the density of residual migration activity at the current moment.

[0031] Finally, the oscillation kinetic energy value is compared with the preset oscillation threshold. The preset oscillation threshold is set based on the system's bus bandwidth margin. Typically, the preset oscillation threshold is set to 2.5, meaning that if more than 2.5 full-intensity migrations occur within the equivalent decay period at the calculation time (i.e., more than 3 ping-pong migrations occur in a very short time), the system will be judged to be in an unstable oscillating state. When the oscillation kinetic energy value is greater than the preset oscillation threshold, a cooperative damping command is generated. This command forcibly increases the cost threshold for the next migration, thereby using hysteresis to suppress the system's turbulent behavior.

[0032] In response to the cooperative damping command, a random backoff factor is generated, and the random backoff factor is weighted and fused with the oscillating kinetic energy value to obtain the virtual mass parameters of the target memory page. The specific steps include: Extract the unique hardware identifier of the current computing node and input it into the preset parameter mapping model as the initial iteration parameter. After a preset number of iterations, generate a numerical sequence and select the current iteration value from the numerical sequence as a random backoff factor. Logarithmic compression mapping is performed on the oscillation kinetic energy value to reduce the difference in numerical magnitude. The processed oscillation kinetic energy value is determined as the kinetic energy gain coefficient. The random retreat factor is multiplied with the kinetic energy gain coefficient to obtain the dynamic weight value of the damping strength at the current moment. The dynamic weight value is accumulated with the preset base unit value to generate the mass expansion ratio at the current moment, and the historical virtual mass parameters of the target memory page stored in the previous control cycle are retrieved. The historical virtual mass parameters are multiplied and amplified using the mass expansion ratio, and the resulting value is updated to the virtual mass parameters of the target memory page under the current control cycle.

[0033] Preferably, the preset logic of the parameter mapping model is as follows: A nonlinear polynomial feedback function structure is constructed with the current iteration value as the independent variable and the next iteration value as the dependent variable. A mapping control parameter is set in the function structure to adjust the degree of numerical divergence, thus forming the basic architecture of the parameter mapping model. Establish a sample set containing multiple sets of historical concurrent conflict events. For each set of historical concurrent conflict events, extract the instruction execution timing data of each computing node when the conflict occurs, and calculate the cross-correlation coefficient of the numerical sequences generated by each computing node. Mark the state with the cross-correlation coefficient lower than the preset safety threshold as a valid divergence state. The mapping control parameters are used as variables to be optimized. The Monte Carlo simulation algorithm is used to randomly select test parameter values ​​within the preset numerical domain and substitute them into the nonlinear polynomial feedback function structure for simulation iteration. The probability distribution of the numerical sequence generated under different test parameter values ​​is in an effective divergent state is statistically analyzed. Select the test parameter value associated with the maximum probability value in the probability distribution, and solidify it as the standard mapping control parameter of the parameter mapping model to complete the model training; The logic for setting the iteration parameters is as follows: obtain the static hardware physical address code and dynamic clock jitter frequency value of the current computing node, perform bitwise XOR operation and normalization on the static hardware physical address code and dynamic clock jitter frequency value, and determine the floating-point value within the preset value range after processing as the initial iteration parameter of the parameter mapping model.

[0034] Preferably, after a preset number of iterative calculations, a numerical sequence is generated, including: The initial iteration parameters are set to the current input values ​​of the first round of operation, and the standard mapping control parameters are retrieved into the operation register to start the single-step operation logic based on the nonlinear polynomial feedback function structure. The single-step operation logic includes calculating the difference between the preset unit value and the current input value, performing continuous multiplication of the difference with the current input value and the standard mapping control parameter to obtain the single-step operation result; The result of the single-step operation is updated to the current input value of the next round of operation, and the single-step operation logic is repeated until the number of operations reaches the preset transient elimination threshold, thereby eliminating non-stationary values ​​with high autocorrelation characteristics in the initial stage. After the number of operations exceeds the transient elimination threshold, the single-step operation logic continues to be executed, and the single-step operation results generated each time are stored in a first-in-first-out buffer queue of a preset capacity until the queue is full. All values ​​in the first-in-first-out buffer queue are uniformly truncated with floating-point precision to eliminate precision errors caused by differences in hardware architecture, and the processed set of values ​​is encapsulated into a numerical sequence.

[0035] Preferably, a sample set containing multiple sets of historical concurrent conflict events is established. For each set of historical concurrent conflict events, the instruction execution timing data of each computing node at the time of the conflict is extracted, and the cross-correlation coefficient of the numerical sequences generated by each computing node is calculated. States with cross-correlation coefficients lower than a preset safety threshold are marked as valid divergence states. Specifically, this includes: Traverse the underlying logs of the multi-agent system, filter out records in which multiple computing nodes initiate bus occupancy requests for the same physical memory address within a preset clock cycle window, aggregate the filtered records to define a group of historical concurrent conflict events, and accumulate multiple such groups of events to construct the sample set. For each set of historical concurrent conflict events in the sample set, the precise instruction submission time point when each computing node issues a bus occupancy request is analyzed, the time offset of the instruction submission time point relative to the start time of the set of events is calculated, and the time offset is determined as the instruction execution timing data of each computing node. The instruction execution timing data is substituted into the nonlinear polynomial feedback function structure as an analog input. The virtual numerical sequence corresponding to each computing node is obtained by iterative generation while keeping the mapping control parameters unchanged. The sliding dot product operation and normalization are performed on the virtual numerical sequences of any two computing nodes in the same set of events to obtain the cross-correlation coefficient representing the waveform similarity between the two. The absolute value of the cross-correlation coefficient is compared with a preset safety threshold. If the absolute value of the cross-correlation coefficient is less than the preset safety threshold, it is determined that the current parameter configuration can effectively disrupt the synchronous oscillation behavior between nodes, and the system state under this parameter configuration is marked as an effective divergence state.

[0036] Preferably, the mapping control parameters are used as variables to be optimized. A Monte Carlo simulation algorithm is used to randomly select test parameter values ​​within a preset numerical domain and substitute them into a nonlinear polynomial feedback function structure for simulation iteration. The probability distribution of the numerical sequences generated under different test parameter values ​​being in an effective divergent state is statistically analyzed, specifically including: Based on the mathematical characteristics of the nonlinear polynomial feedback function structure, a range of values ​​that can maintain the iterative operation without numerical overflow and have non-periodic characteristics is determined. The range of values ​​is set as a preset numerical domain, and the domain is divided into several equally spaced parameter statistical intervals. A large number of discrete values ​​are extracted from the preset numerical domain using a uniform distribution random sampling algorithm as test parameter values. Each test parameter value is then filled into the coefficient position of the nonlinear polynomial feedback function structure in sequence, and the virtual iterative operation is started by calling the initial iteration parameters to generate the corresponding virtual numerical sequence. For each test parameter value, a virtual numerical sequence is generated. Based on the cross-correlation coefficient calculation logic, a divergence check is performed. If the check result determines that the virtual numerical sequence is in a valid divergence state, a valid hit mark is generated and mapped to the parameter statistical interval to which the test parameter value belongs. The cumulative number of valid hit markers within each parameter statistical interval is counted, and the ratio of this number to the total number of test parameter values ​​falling within the interval is calculated. This yields the success rate of the system generating an effective divergence state induced by each parameter statistical interval. The set of success rate values ​​from all parameter statistical intervals is then used to determine the probability distribution.

[0037] Preferably, the range of values ​​that can maintain the iterative operation without numerical overflow and possess aperiodic characteristics is determined, specifically including: We construct the analytical expression of the first derivative for the nonlinear polynomial feedback function structure, use the analytical expression of the first derivative to calculate the distribution of the extreme points of the function curve, and determine the theoretical convergence boundary that makes the function output value always in a bounded state based on the distribution of the extreme points. Within the range defined by the theoretical convergence boundary, select a number of test points for exploratory iteration, monitor the numerical magnitude of the iteration results in real time, and eliminate the numerical regions that would cause the calculation results to exceed the preset hardware register bit width limit, thereby defining a safe operating range that can avoid numerical overflow. The system iterates through the values ​​within the safe operating range, calculates the trajectory separation index of adjacent values ​​after a preset number of iterations, merges continuous value segments with trajectory separation indices greater than a preset judgment threshold, and thus determines a preset value domain with non-periodic characteristics.

[0038] In one specific embodiment, the process of generating the random backoff factor relies on a preset parameter mapping model, which is essentially a parameter-optimized nonlinear chaos generator. Specifically, the system first extracts the unique hardware identifier of the current computing node (e.g., the 48-bit MAC address of the network interface controller or the CPU's serial number) as a static component, and simultaneously reads the nanosecond-level clock jitter value of the high-precision event timer at the current moment as a dynamic component. The system performs a bitwise XOR operation on the static and dynamic components, and maps the result to a modulo operation. The interval is used to obtain the initial iteration parameters. The purpose of this step is to ensure that even if multiple nodes trigger the damping mechanism simultaneously, their initial seeds are completely orthogonal, thus avoiding seed collisions in the pseudo-random number generator that could cause synchronization of damping behavior.

[0039] Next, a numerical sequence is generated using a pre-trained parameter mapping model. This model employs a variant of the classic Logistic mapping as the nonlinear polynomial feedback function structure, and its core iterative formula is: ; in, This is the value for the current iteration. The value for the next iteration. These are standard mapping control parameters.

[0040] In actual calculations, Substituting into the above formula, perform a preset number of warm-up iterations. The logic for setting the preset number of iterations is usually based on the transient elimination threshold. (For example, setting it to 64 rounds) is to eliminate transient trajectories caused by the initial value sensitivity of chaotic systems, ensuring that the numerical sequence enters the completely disordered chaotic attractor region. After completing... After each iteration, perform a single-step calculation, and the resulting numerical value is the random backoff factor at the current moment. .

[0041] It should be added that the standard mapping control parameters The acquisition and model training process is typically completed during the system initialization phase: 1. Domain Determination: Based on the mathematical characteristics of the Logistic mapping, the range of values ​​that can maintain the system in a chaotic state without numerical overflow is determined. Within this range, the system calculates the first derivative and the Lyapunov exponent (i.e., the trajectory separation exponent), removes the periodic window, and divides the remaining interval into equally spaced parameter statistical intervals according to a preset step size.

[0042] 2. Monte Carlo simulation: The Monte Carlo algorithm is used to randomly sample test parameter values ​​within each of the above intervals. For each Simulates a multi-node concurrent conflict scenario, generates multiple virtual numerical sequences, and calculates the cross-correlation coefficients between the sequences. .

[0043] 3. Valid divergence determination: If If the value is below the preset safety threshold, it is determined that the parameter can effectively disrupt synchronous oscillation, that is, it is in an effective divergence state.

[0044] 4. Parameter solidification: Calculate the probability of generating a valid divergence state in each interval, and select the median of the interval with the highest probability as the parameter. This ensures the generation It has the highest decorrelation.

[0045] Subsequently, dynamic weight values ​​are calculated and virtual mass parameters are updated. The system acquires the oscillating kinetic energy value and performs logarithmic compression mapping on it. The reason for using a logarithmic function is that the oscillating kinetic energy may increase exponentially due to sudden traffic surges; direct use would lead to excessive damping and freezing of memory pages, while logarithmic mapping can smooth out numerical differences. The kinetic energy gain coefficient is then calculated. The formula is: ; Adding 1 here is to ensure that when The gain is 0, and the ln function guarantees the normalization of the output dimensions.

[0046] Next, the random backoff factor With kinetic energy gain coefficient By fusing the data, the dynamic weighted value of the damping intensity at the current moment is obtained. : ; Here The purpose is to introduce random disturbances, causing different nodes to compete for the same hot page in a random way, thereby breaking the "deadlock" ping-pong effect.

[0047] Finally, the mass expansion ratio is calculated based on the dynamic weight values. The formula is ,in This is a preset base unit value used to control the damping response sensitivity. The system retrieves historical virtual quality parameters from the target memory page stored in the previous control cycle. If there is no historical value, the initial value is 1.0, and a multiplication amplification operation is performed: ; Calculated This refers to the updated virtual quality parameters.

[0048] It needs to be explained that the logic behind using logarithmic compression mapping to process the oscillation kinetic energy value is to map the potentially exponentially explosive raw oscillation data to a linearly controllable range. This prevents the damping parameter from overflowing or the system from overreacting due to excessively large instantaneous values, thus ensuring that the kinetic gain coefficient can both characterize the oscillation intensity and maintain numerical stability. Subsequently, the random backoff factor is multiplied with the kinetic gain coefficient. This aims to use the unpredictability of random numbers to disrupt the decision-making synchronization of multiple nodes when processing the same hot page, avoiding secondary oscillations caused by multiple nodes making consistent decisions due to the same oscillation perception. This allows the dynamic weight value to have both intensity characterization and decorrelation characteristics. Finally, the mass expansion ratio is calculated based on the dynamic weight value and multiplied with the historical virtual mass parameter for update. This operation simulates the principle of physical inertia accumulation, that is, the longer the duration and the greater the intensity of the oscillation, the greater the "virtual mass" of the memory page, and the higher the "energy" threshold required for migration. This design logically forms a closed loop: by gradually increasing the migration threshold, invalid migration is suppressed until the access mode changes or the contention disappears.

[0049] It should be added that the Monte Carlo algorithm is used to randomly sample test parameter values ​​within each of the above intervals. And for each The process of simulating multi-node concurrent conflict scenarios to calculate cross-correlation coefficients aims to verify through numerical simulation whether chaotic sequences under specific control parameters have sufficient desynchronization capabilities, thereby preventing multiple computing nodes from colliding again during the backoff process.

[0050] The process begins by using a pseudo-random number generator to perform uniform distribution sampling within each interval based on parameter statistical intervals, thereby extracting specific test parameter values. To simulate a realistic concurrent conflict environment, the system constructs a simulation model containing two virtual nodes (node ​​A and node B), and assigns different initial iteration seeds to these two nodes. and The generation logic for these two initial seeds simulates the physical address differences in actual hardware, such as setting... This represents two different physical nodes that collide at the same time. Then, the selected test parameter values ​​are... Substituting into the nonlinear polynomial feedback function structure ( ), synchronously iterate through nodes A and B respectively to generate two virtual numerical sequences of length L. and Where L is the preset sequence sampling length, and it is recommended to take a value of 50 to 100. The reason for setting this range is that if the length is too short, the statistical characteristics are not significant; if the length is too long, it will exceed the effective time window for actual memory access backoff, resulting in a waste of computing resources.

[0051] After generating the sequences, calculate the cross-correlation coefficients between the sequences. This coefficient is used to quantify the linear correlation between the backoff factor sequences generated by two nodes. The calculation uses a discretized form of the Pearson product-moment correlation coefficient, and its specific formula is as follows: ; in, and Sequences and The value of the k-th iteration; and These are the arithmetic means of the two sequences, respectively; the denominator is the product of the standard deviations of the two sequences.

[0052] The formula is introduced to ensure that: when When it is close to 0, it indicates that for the current parameter Even if two nodes have different initial seeds, the backoff sequences they generate are orthogonal and do not interfere with each other, effectively dispersing concurrent requests; if | A value close to 1 indicates that the sequence is highly synchronized, which may cause nodes to retry simultaneously after backing away, triggering another conflict.

[0053] It should be added that the operation logic for parameter solidification is as follows: The defined numerical domain is divided into K consecutive parameter statistical intervals according to a preset sampling resolution step size. For each interval, the system iterates through all Monte Carlo simulation records performed within that interval and counts the total number of test parameter values ​​falling within that interval. And the number of hits that were determined to be in a valid divergence state (i.e., the absolute value of the cross-correlation coefficient was less than the safety threshold). .

[0054] Simultaneously calculate the divergence probability density for each interval. This metric quantifies the reliability of the system in generating high-quality chaotic sequences within this parameter range. The calculation formula is as follows: ; in, and The count value is in the dimension "number", while the comparison value is a dimensionless probability. This term represents the standard deviation of the cross-correlation coefficients of all test samples within the interval. It is introduced to penalize intervals that have low average correlation but drastic fluctuations. The pre-defined fluctuation penalty factor is usually set to 1.0. This formula considers not only the success rate of divergence but also the stability of divergence quality. If most samples in a certain interval diverge, but occasionally there are outliers with extremely high correlation (meaning it is on the verge of chaos), its score will be lowered by the standard deviation term.

[0055] Subsequently, global optimization and parameter fixing are performed. The system processes all calculated parameters. Sort the data and identify the target interval corresponding to the global maximum point of the probability density curve. To ensure the robustness of the parameters and prevent them from drifting to nearby non-chaotic regions due to floating-point operation errors, the boundary values ​​of the interval are not used directly. Instead, the geometric center position of the target interval is calculated and determined as the final standard mapping control parameter. The calculation formula is: ; in, For the target interval The lower bound value, This represents the interval width.

[0056] It needs to be explained that the system iterates through the divergence probability calculation results of all statistical intervals, executes the maximum value retrieval algorithm, and obtains the interval index max(k) with the highest probability value by comparison. Then, based on this index, it locates the corresponding parameter domain range and locks this range as the target interval.

[0057] The migration power threshold is obtained by multiplying the virtual mass parameter with the bus delay matrix. Simultaneously, the gravitational potential energy difference between the current node and the requesting node is calculated based on the access feature vector. This process includes the following steps: The communication link between the resident node currently holding the target memory page and the external node that initiated the access request is located in the bus delay matrix. The signal transmission delay time corresponding to the communication link is extracted as the path impedance coefficient. The path impedance coefficient and the virtual quality parameter are multiplied by a scalar to obtain the migration power consumption threshold that is required to overcome the energy limit of the current system damping. The access feature vector is parsed to separate the first feature component belonging to the resident node and the second feature component belonging to the external node. The access frequency value and instruction cycle cost value are extracted for each feature component. The preset gravity weight coefficient table is retrieved, and the access frequency value is weighted and corrected using the instruction cycle overhead value. The original gravity value of the node's dependence on memory page computation is calculated, and the bus load rate of the current node is introduced to normalize the original gravity value, generating the local adsorption potential energy value of the resident node and the remote traction potential energy value of the external node respectively. Perform a difference operation by subtracting the local adsorption potential energy value from the remote traction potential energy value, and determine the numerical result as the gravitational potential energy difference that drives the target memory page to undergo spatial displacement.

[0058] The gravitational potential energy difference is compared with the migration power consumption threshold. If the gravitational potential energy difference is less than the migration power consumption threshold, the current system is determined to be in a livelock state. A memory lock instruction is generated to prohibit the migration of the target memory page, and a task migration signal to schedule the computing task to the current node is generated based on the memory lock instruction.

[0059] In one specific embodiment, virtual mass parameters are invoked. Then, in the bus delay matrix, the coordinate index between the resident node currently holding the target memory page and the external node initiating the access request is located. The element value at this index position is extracted as the path impedance coefficient. This coefficient characterizes the signal delay of data transmission over the physical link. To quantify the energy floor required to overcome current system damping, a migration power threshold is defined. The calculation formula is as follows: ; in, This is a preset energy conversion coefficient, with its unit of flavor-normalized cost per nanosecond. A value of 1 is recommended to map the time delay dimension to the system cost dimension. The larger the virtual mass of a page (i.e., the more violent the oscillation), or the farther the physical distance of a node, the higher the migration barrier it forms. Only nodes with extremely strong access needs can overcome this barrier.

[0060] Next, access feature vectors Perform dimensional analysis to separate the nodes belonging to the resident nodes. First characteristic component and belonging to external nodes The second characteristic component Where f represents the access frequency per unit time, and c represents the instruction cycle overhead. A preset gravity weighting coefficient table is retrieved, and the access frequency value is weighted and corrected using the instruction cycle overhead value. The "computational dependency" of the compute node on the memory page, i.e., the original gravity value, is then calculated. : ; This represents the frequency with which node i accesses the target memory page per unit of time. It reflects the "demand intensity" of the node for this data resource; the higher the access frequency, the more the node needs this memory page. This represents the instruction cycle overhead when node i accesses the target memory page.

[0061] Here, the subscript i is a node index variable used to distinguish different compute nodes. When i=s, it represents a resident node, and when i=r, it represents an external request node. This is an overhead sensitivity index used to nonlinearly amplify the gravitational weights resulting from high latency.

[0062] Further introduce the bus load rate of the current node The original gravity value is normalized to generate the actual potential energy value, where the bus load rate is a percentage obtained through a hardware counter. The logic here is that when the node bus load is too high, its effective adsorption capacity decreases due to queuing effects. Therefore, the local adsorption potential energy value of the resident node... Remote traction potential energy value of external nodes The calculation is as follows: ; This represents the potential energy value of node i; By substituting the node roles, the formula is applied to the calculations obtained by the resident nodes. direct assignment The formula is applied to the calculations obtained by the external request node. direct assignment .

[0063] Finally, the difference in gravitational potential energy is obtained by performing a difference operation. : This difference represents the net driving force that "pulls" memory pages from the resident node to the external node.

[0064] The difference in gravitational potential energy The value is compared numerically with the migration power consumption threshold. If the gravitational potential energy difference is determined to be less than the migration power consumption threshold, and... When the value is >0, the system generates a memory lock instruction, temporarily prohibiting changes to the physical address of the page, and generates a task migration signal based on the memory lock instruction. This signal will trigger the operating system kernel's task scheduler, which, based on the principle of "data does not move, computation moves," migrates the thread or computation task context that initiated the request on the external node to the resident node for execution, thereby eliminating remote access latency at the physical level and completely avoiding memory thrashing.

[0065] It needs to be explained that if we do not calculate... Using the raw gravity value directly ignores the limitation of memory service capacity caused by the node's own bus congestion—even if a heavily loaded node initiates high-frequency access, its efficiency in processing migrated memory pages will be significantly reduced due to queuing effects, and blindly migrating to that node will actually exacerbate system latency. Therefore, calculating using the formula... In essence, this involves performing a "load penalty" operation, correcting the theoretical access gravity into effective potential energy in a real-world environment. It is precisely because of this... (resident nodes) )and (Request node) These two corrected values, and the subsequent differential operations, reflect the net driving force for memory page movement after taking into account the differences in load capacity between the two sides, thus ensuring that the migration decision does not overload the system by ignoring node load. Example

[0066] like Figure 2 As shown, an autonomous task allocation and dynamic optimization system based on agent collaboration is presented. The system includes a damping command generation module, a potential energy difference calculation module, and a task migration scheduling module. The various modules are connected via wired and / or wireless connections to enable data transmission between them; Damping instruction generation module: acquires the migration history trajectory data of the target memory page within a preset time window, the access feature vector of each computing node, and the bus delay matrix between nodes; calculates the oscillation kinetic energy value of the target memory page based on the migration history trajectory data, and generates a cooperative damping instruction when the oscillation kinetic energy value exceeds a preset oscillation threshold; Potential energy difference calculation module: In response to the cooperative damping command, it generates a random backoff factor and performs a weighted fusion calculation with the random backoff factor and the oscillation kinetic energy value to obtain the virtual mass parameter of the target memory page; it multiplies the virtual mass parameter with the bus delay matrix to obtain the migration power consumption threshold, and calculates the gravitational potential energy difference between the current node and the requesting node based on the access feature vector; Task migration scheduling module: It compares the gravitational potential energy difference with the migration power consumption threshold. If the gravitational potential energy difference is less than the migration power consumption threshold, it determines that the current system is in a livelock state, generates a memory lock instruction to prohibit the migration of the target memory page, and generates a task migration signal to schedule the computing task to the current node based on the memory lock instruction.

[0067] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0068] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0070] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for autonomous task allocation and dynamic optimization based on agent collaboration, characterized in that, include: Obtain the migration history trajectory data of the target memory page within a preset time window, the access feature vector of each computing node, and the bus delay matrix between nodes; The oscillation kinetic energy value of the target memory page is calculated based on the migration history trajectory data, and a cooperative damping instruction is generated when the oscillation kinetic energy value exceeds the preset oscillation threshold. In response to the cooperative damping command, a random backoff factor is generated, and the random backoff factor is weighted and fused with the oscillation kinetic energy value to obtain the virtual mass parameters of the target memory page; The migration power consumption threshold is obtained by multiplying the virtual mass parameter with the bus delay matrix, and the gravitational potential energy difference between the current node and the requesting node is calculated based on the access feature vector. The gravitational potential energy difference is compared with the migration power consumption threshold. If the gravitational potential energy difference is less than the migration power consumption threshold, the current system is determined to be in a livelock state. A memory lock instruction is generated to prohibit the migration of the target memory page, and a task migration signal to schedule the computing task to the current node is generated based on the memory lock instruction.

2. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 1, characterized in that, The oscillation kinetic energy value of the target memory page is calculated based on the migration history trajectory data, and a cooperative damping instruction is generated when the oscillation kinetic energy value exceeds a preset oscillation threshold. The specific steps include: The migration history trajectory data is traversed in time series to filter out records with inconsistent resident node identifiers at adjacent times as valid migration events, and the corresponding time decay weight is determined based on the time span between the time of occurrence of the valid migration event and the current time. All selected valid migration events are weighted and summed according to time decay weights to obtain the oscillation kinetic energy value of the reciprocating movement intensity of the target memory page between computing nodes. The oscillation kinetic energy value is compared with the preset oscillation threshold. If the oscillation kinetic energy value is greater than the preset oscillation threshold, a coordinated damping command is generated.

3. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 2, characterized in that, The time decay weight is determined based on the time span between the occurrence of a valid migration event and the current time, specifically including: Extract the generation time point when each valid migration event is recorded, calculate the time difference between the generation time point and the current system time, and obtain the time span value of the event; The time span value is mapped using a preset negative exponential decay logic, so that the smaller the time span value, the larger the mapped value. The value obtained after mapping is determined as the time decay weight corresponding to the effective migration event, and a one-to-one correspondence between the time decay weight and the effective migration event is established.

4. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 1, characterized in that, Obtaining the virtual quality parameters of the target memory page involves the following steps: Extract the unique hardware identifier of the current computing node and input it into the preset parameter mapping model as the initial iteration parameter. After a preset number of iterations, generate a numerical sequence and select the current iteration value from the numerical sequence as a random backoff factor. Logarithmic compression mapping is performed on the oscillation kinetic energy value, and the processed oscillation kinetic energy value is determined as the kinetic energy gain coefficient. The random retreat factor is multiplied with the kinetic energy gain coefficient to obtain the dynamic weight value of the damping strength at the current moment. The dynamic weight value is accumulated with the preset base unit value to generate the mass expansion ratio at the current moment, and the historical virtual mass parameters of the target memory page stored in the previous control cycle are retrieved. The historical virtual mass parameters are multiplied and amplified using the mass expansion ratio, and the resulting value is updated to the virtual mass parameters of the target memory page under the current control cycle.

5. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 4, characterized in that, The default logic of the parameter mapping model is as follows: A nonlinear polynomial feedback function structure is constructed with the current iteration value as the independent variable and the next iteration value as the dependent variable. A mapping control parameter is set in the function structure to adjust the degree of numerical divergence, thus forming the basic architecture of the parameter mapping model. Establish a sample set containing multiple sets of historical concurrent conflict events. For each set of historical concurrent conflict events, extract the instruction execution timing data of each computing node when the conflict occurs, and calculate the cross-correlation coefficient of the numerical sequences generated by each computing node. Mark the state with the cross-correlation coefficient lower than the preset safety threshold as a valid divergence state. The mapping control parameters are used as variables to be optimized. The Monte Carlo simulation algorithm is used to randomly select test parameter values ​​within the preset numerical domain and substitute them into the nonlinear polynomial feedback function structure for simulation iteration. The probability distribution of the numerical sequence generated under different test parameter values ​​is in an effective divergent state is statistically analyzed. Select the test parameter value associated with the maximum probability value in the probability distribution, and solidify it as the standard mapping control parameter of the parameter mapping model to complete the model training; The logic for setting the iteration parameters is as follows: obtain the static hardware physical address code and dynamic clock jitter frequency value of the current computing node, perform bitwise XOR operation and normalization on the static hardware physical address code and dynamic clock jitter frequency value, and determine the floating-point value within the preset value range after processing as the initial iteration parameter of the parameter mapping model.

6. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 4, characterized in that, After a predetermined number of iterative calculations, a numerical sequence is generated, including: The initial iteration parameters are set to the current input values ​​of the first round of operation, and the standard mapping control parameters are retrieved into the operation register to start the single-step operation logic based on the nonlinear polynomial feedback function structure. The single-step operation logic includes calculating the difference between the preset unit value and the current input value, performing continuous multiplication of the difference with the current input value and the standard mapping control parameter to obtain the single-step operation result; Update the single-step calculation result with the current input value for the next round of calculation, and repeat the single-step calculation logic until the number of calculations reaches the preset transient elimination threshold; After the number of operations exceeds the transient elimination threshold, the single-step operation logic continues to be executed, and the single-step operation results generated each time are stored in a first-in-first-out buffer queue of a preset capacity until the queue is full. All values ​​in the first-in-first-out buffer queue are uniformly truncated with floating-point precision to eliminate precision errors caused by differences in hardware architecture, and the processed set of values ​​is encapsulated into a numerical sequence.

7. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 5, characterized in that, A sample set containing multiple sets of historical concurrent conflict events is established. For each set of historical concurrent conflict events, the instruction execution timing data of each computing node at the time of the conflict is extracted, and the cross-correlation coefficient of the numerical sequences generated by each computing node is calculated. States with cross-correlation coefficients lower than a preset safety threshold are marked as valid divergence states. Specifically, this includes: Traverse the underlying logs of the multi-agent system, filter out records in which multiple computing nodes initiate bus occupancy requests for the same physical memory address within a preset clock cycle window, aggregate the filtered records to define a group of historical concurrent conflict events, and accumulate multiple such groups of events to construct the sample set. For each set of historical concurrent conflict events in the sample set, the precise instruction submission time point when each computing node issues a bus occupancy request is analyzed, the time offset of the instruction submission time point relative to the start time of the set of events is calculated, and the time offset is determined as the instruction execution timing data of each computing node. The instruction execution timing data is substituted into the nonlinear polynomial feedback function structure as an analog input. The virtual numerical sequence corresponding to each computing node is obtained by iterative generation while keeping the mapping control parameters unchanged. The sliding dot product operation and normalization are performed on the virtual numerical sequences of any two computing nodes in the same set of events to obtain the cross-correlation coefficient representing the waveform similarity between the two. The absolute value of the cross-correlation coefficient is compared with a preset safety threshold. If the absolute value of the cross-correlation coefficient is less than the preset safety threshold, it is determined that the current parameter configuration can effectively disrupt the synchronous oscillation behavior between nodes, and the system state under this parameter configuration is marked as an effective divergence state.

8. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 5, characterized in that, Using the mapping control parameters as variables to be optimized, the Monte Carlo simulation algorithm is used to randomly select test parameter values ​​within a preset numerical domain and substitute them into a nonlinear polynomial feedback function structure for simulation iteration. The probability distribution of the numerical sequences generated under different test parameter values ​​being in an effective divergent state is statistically analyzed, specifically including: Based on the mathematical characteristics of the nonlinear polynomial feedback function structure, a range of values ​​that can maintain the iterative operation without numerical overflow and have non-periodic characteristics is determined. The range of values ​​is set as a preset numerical domain, and the domain is divided into several equally spaced parameter statistical intervals. A large number of discrete values ​​are extracted from the preset numerical domain using a uniform distribution random sampling algorithm as test parameter values. Each test parameter value is then filled into the coefficient position of the nonlinear polynomial feedback function structure in sequence, and the virtual iterative operation is started by calling the initial iteration parameters to generate the corresponding virtual numerical sequence. For each test parameter value, a virtual numerical sequence is generated. Based on the cross-correlation coefficient calculation logic, a divergence check is performed. If the check result determines that the virtual numerical sequence is in a valid divergence state, a valid hit mark is generated and mapped to the parameter statistical interval to which the test parameter value belongs. The cumulative number of valid hit markers within each parameter statistical interval is counted, and the ratio of this number to the total number of test parameter values ​​falling within the interval is calculated. This yields the success rate of the system generating an effective divergence state induced by each parameter statistical interval. The set of success rate values ​​from all parameter statistical intervals is then used to determine the probability distribution.

9. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 8, characterized in that, Determine the range of values ​​that can maintain the iterative operation without numerical overflow and possess aperiodic characteristics, specifically including: We construct the analytical expression of the first derivative for the nonlinear polynomial feedback function structure, use the analytical expression of the first derivative to calculate the distribution of the extreme points of the function curve, and determine the theoretical convergence boundary that makes the function output value always in a bounded state based on the distribution of the extreme points. Within the range defined by the theoretical convergence boundary, select a number of test points for exploratory iteration, monitor the numerical magnitude of the iteration results in real time, and eliminate the numerical regions that would cause the calculation results to exceed the preset hardware register bit width limit, thereby defining a safe operating range that can avoid numerical overflow. The system iterates through the values ​​within the safe operating range, calculates the trajectory separation index of adjacent values ​​after a preset number of iterations, merges continuous value segments with trajectory separation indices greater than a preset judgment threshold, and thus determines a preset value domain with non-periodic characteristics.

10. The autonomous task allocation and dynamic optimization method based on agent collaboration according to claim 1, characterized in that, The calculation of the gravitational potential energy difference between the current node and the requesting node based on the access feature vector includes the following steps: The communication link between the resident node currently holding the target memory page and the external node that initiated the access request is located in the bus delay matrix. The signal transmission delay time corresponding to the communication link is extracted as the path impedance coefficient. The path impedance coefficient and the virtual quality parameter are multiplied by a scalar to obtain the migration power consumption threshold that is required to overcome the energy limit of the current system damping. The access feature vector is parsed to separate the first feature component belonging to the resident node and the second feature component belonging to the external node. The access frequency value and instruction cycle cost value are extracted for each feature component. The preset gravity weight coefficient table is retrieved, and the access frequency value is weighted and corrected using the instruction cycle overhead value. The original gravity value of the node's dependence on memory page computation is calculated, and the bus load rate of the current node is introduced to normalize the original gravity value, generating the local adsorption potential energy value of the resident node and the remote traction potential energy value of the external node respectively. Perform a difference operation by subtracting the local adsorption potential energy value from the remote traction potential energy value, and determine the numerical result as the gravitational potential energy difference that drives the target memory page to undergo spatial displacement.

11. An autonomous task allocation and dynamic optimization system based on agent collaboration, characterized in that, It is implemented based on the autonomous task allocation and dynamic optimization method based on agent cooperation as described in any one of claims 1-10, and includes: Damping instruction generation module: acquires the migration history trajectory data of the target memory page within a preset time window, the access feature vector of each computing node, and the bus delay matrix between nodes; calculates the oscillation kinetic energy value of the target memory page based on the migration history trajectory data, and generates a cooperative damping instruction when the oscillation kinetic energy value exceeds a preset oscillation threshold; Potential energy difference calculation module: In response to the cooperative damping command, it generates a random backoff factor and performs a weighted fusion calculation with the random backoff factor and the oscillation kinetic energy value to obtain the virtual mass parameter of the target memory page; it multiplies the virtual mass parameter with the bus delay matrix to obtain the migration power consumption threshold, and calculates the gravitational potential energy difference between the current node and the requesting node based on the access feature vector; Task migration scheduling module: It compares the gravitational potential energy difference with the migration power consumption threshold. If the gravitational potential energy difference is less than the migration power consumption threshold, it determines that the current system is in a livelock state, generates a memory lock instruction to prohibit the migration of the target memory page, and generates a task migration signal to schedule the computing task to the current node based on the memory lock instruction.

Images