Intelligent resource scheduling and provisioning method and system in heterogeneous computing environment

Abstract

The application relates to the technical field of resource scheduling and discloses an intelligent resource scheduling supply method and system in a heterogeneous computing environment, which comprises the following steps: acquiring core heterogeneous data, performing exponential smoothing denoising to obtain a smooth state value, combining an HAT extended tuple, and analyzing to obtain a comprehensive capability score; automatically adjusting contribution degree by similarity of static characteristics and dynamic characteristics; constructing a coupling correction function, calculating a hybrid precision matching degree, combining an exponential product to reflect nonlinear correlation between targets; constructing a dynamic penalty function, optimizing a quantum genetic algorithm fitness function, and performing multi-target collaborative optimization; constructing a complexity coupling and dynamic adaptation mechanism to dynamically adjust a task segmentation ratio; designing a frequency and reconfiguration time index ratio to optimize FPGA bit stream preloading; constructing a synchronous delay coupling model to perform state synchronization optimization; balancing multidimensional feedback through a combined reward function; and improving explainability by quantifying the influence of decision factors through contribution degree entropy.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, specifically to an intelligent resource scheduling and supply method and system in a heterogeneous computing environment. Background Technology

[0002] In existing technologies and patents, unified resource pools mostly focus on logical aggregation, failing to address the core issue of differences in physical resource characteristics. For example, Kubernetes uses DevicePlugin to access GPUs and FPGAs, supporting only whole-card / shard allocation. It cannot convert CPU computing cores, GPU CUDA cores, and FPGA logic unit LUTs into a unified computing power unit, making it impossible for the scheduler to recognize the differences in computing power cost-effectiveness between different hardware. While the intermediate instruction layer masks underlying differences, it does not optimize the compilation efficiency of virtual instructions to hardware native instructions. In industrial vision inspection scenarios, the instruction conversion time accounts for a higher proportion of the total task time than homogeneous scheduling. There is a time lag between computing power resource status acquisition and scheduling decisions, making it difficult for prediction models based on historical data to capture the resource demands of sudden tasks, resulting in insufficient or excessive resource reservation. Some dynamic scheduling strategies improve resource utilization but do not incorporate energy consumption into the objective function, leading to a decrease in energy efficiency under high load scenarios. Real-time tasks are sensitive to latency, but existing patents do not clearly distinguish task priorities, potentially causing critical tasks to be preempted by low-priority tasks.

[0003] Therefore, there is a need to provide intelligent resource scheduling and supply methods and systems for heterogeneous computing environments. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent resource scheduling and supply method and system in a heterogeneous computing environment. To solve the above-mentioned problems in the prior art, this invention achieves this through the following technical solution:

[0005] Firstly, the intelligent resource scheduling and supply method in a heterogeneous computing environment provided by embodiments of the present invention specifically includes the following steps:

[0006] Step 1: Collect heterogeneous data from the heterogeneous computing cluster to obtain core heterogeneous data, perform exponential smoothing and denoising to obtain smoothed state values, combine HAT extended tuples, use geometric mean fusion index to analyze and obtain comprehensive capability score, and determine whether the communication performance of the heterogeneous computing cluster meets the requirements.

[0007] Step 2: For heterogeneous computing clusters with compatible communication performance, the contribution is automatically adjusted by using the similarity between static and dynamic features through attention fusion and feature coupling mechanisms; a coupling correction function is constructed to calculate the mixed precision matching degree and determine the suitability between the task and the hardware precision.

[0008] Step 3: For heterogeneous computing clusters that are adapted to the task and hardware precision, construct a multi-objective coupling penalty mechanism, and combine exponential product to reflect the nonlinear correlation between objectives; construct a dynamic penalty function, optimize the fitness function of the quantum genetic algorithm, and perform multi-objective collaborative optimization;

[0009] Step 4: Based on the completion of collaborative optimization, construct a complexity coupling and dynamic adaptation mechanism, and dynamically adjust the task partitioning ratio using the nonlinear product of data volume and latency sensitivity; design the frequency and reconfiguration time exponential ratio to optimize FPGA bitstream preloading; construct a synchronous delay coupling model and perform state synchronization optimization.

[0010] Step 5: After the state synchronization optimization is completed, construct a reward coordination and loss linkage mechanism, balance multi-dimensional feedback by combining reward functions, and quantify the impact of decision factors by contribution entropy value to improve interpretability.

[0011] A further invention provides that the method for obtaining the comprehensive ability score is as follows:

[0012] Exponential smoothing denoising is performed on the core heterogeneous data to preserve real-time performance while suppressing high-frequency noise. The original core heterogeneous data at time t and the original core heterogeneous data at time t-1 are fused according to preset weights to calculate the smoothed state value and sort it.

[0013] Based on the obtained smooth state values, construct HAT tuples, assign architecture identifier codes to different hardware types, and generate globally unique resource IDs by combining HAT tuples.

[0014] For any smooth state value sequence, calculate the lower quartile and upper quartile values ​​of the smooth state value sequence, and calculate the difference between the upper quartile and lower quartile values ​​to obtain the anomaly threshold;

[0015] If the smoothed state value is less than or greater than the outlier threshold, the median of the sequence is used instead of the smoothed state value.

[0016] Combining HAT extended tuples and utilizing the geometric mean fusion index, the formula is:

[0017]

[0018] Calculate the comprehensive ability score ,in, To smooth the total number of dimensions of the state values, For dimensional indexing, The k-th dimension smoothed state value after preprocessing. To find the theoretical minimum extreme value, This is the maximum extremum of the standard theory;

[0019] A further invention, wherein the method for automatically adjusting the contribution level is as follows:

[0020] For static features, obtain task submission information and extract static features;

[0021] For dynamic features, monitor and acquire dynamic features during task execution;

[0022] Dependency features are obtained through task DAG graph analysis;

[0023] By employing an attention fusion and feature coupling mechanism, and using attention similarity fusion, key features are automatically highlighted through the formula:

[0024]

[0025] Calculate the task profile ,in, For static feature vectors, The cosine similarity between static and dynamic features ranges from [0,1]. The average value of the dynamic characteristics. For dependent feature vectors, Due to the complexity of the dependency, and ;

[0026] If the similarity between static and dynamic features is high, then the contribution ratio of static features is high; if the dependency complexity is high, then the contribution ratio of dependency feature vectors is low.

[0027] A further invention, wherein the method for calculating the mixed precision matching degree is as follows:

[0028] Obtain the hardware-supported accuracy and the task's theoretical accuracy. Quantify the compatibility between the task and hardware accuracy using a mixed-precision matching degree calculation, reflecting the non-linear correlation between hardware-supported accuracy and task-theoretical accuracy, through the formula: Calculate the mixed precision matching degree ,in, For the theoretical accuracy of the task, To ensure hardware accuracy, Let be the accuracy loss rate of task i on hardware j;

[0029] If the hardware-supported accuracy perfectly matches the theoretical accuracy of the task, then The greater the accuracy deviation or the higher the accuracy loss rate, the smaller the mixed accuracy matching degree.

[0030] The obtained mixed precision matching degree is compared with the preset matching degree threshold to analyze and determine the heterogeneous computing cluster that is compatible with the task and hardware precision.

[0031] A further invention, the method for constructing the multi-objective coupling penalty mechanism is as follows:

[0032] Construct a multi-objective coupled penalty mechanism: Objective constraints: extract the total latency upper limit and power budget from the task SLA; obtain the resource utilization safety threshold from the system configuration;

[0033] Predictive data: A random forest model is trained using historical scheduling records to predict the latency of task i on resource j. and energy consumption ;

[0034] Decision variables: decision Where 1 indicates that task i is assigned to resource j, and 0 indicates that task i is not assigned to resource j, the maximum utilization rate of the resource pool is obtained. ;

[0035] A further invention, the method for optimizing the fitness function of the quantum genetic algorithm is as follows:

[0036] By integrating latency, energy consumption, overload, and accuracy loss, a dynamic loss function is calculated to avoid subjective weighting, using the formula:

[0037]

[0038] The dynamic loss was calculated. ,in, For the total prediction delay, and , For the total SLA latency budget, For total predicted energy consumption, and , For the power budget of the resource pool, As an overload penalty, and , For accuracy penalty, and ;

[0039] Based on the obtained dynamic loss, the reciprocal of the dynamic loss is used as the fitness to construct an optimized quantum genetic algorithm fitness function, guiding the optimized quantum genetic algorithm to quickly converge to the optimal solution, through the formula: The fitness of the quantum genetic algorithm is calculated. The quantum genetic algorithm searches for the optimal allocation scheme through qubit encoding;

[0040] A further invention, the method for optimizing FPGA bitstream preloading is as follows:

[0041] Dynamically adjust the edge processing ratio to balance latency and transmission cost, using the formula:

[0042]

[0043] The proportion of edge tasks was calculated. ,in, To transmit data volume, For dynamic data thresholds;

[0044] Optimize the preloading strategy to reduce reconfiguration time, obtain the historical usage frequency of bitstream b and combine it with the exponential ratio of cluster reconfiguration time to minimum cluster reconfiguration time to obtain bitstream priority;

[0045] Bitstreams that are used frequently and reconfigured quickly have higher priority and are preloaded into the FPGA cache to reduce reconfiguration time.

[0046] A further invention, wherein the method for performing state synchronization optimization is as follows:

[0047] To balance synchronization latency and data reliability, optimize state synchronization: in, To verify the strength, To synchronize time, Error rate;

[0048] A further invention provides that the method for quantifying the influence of decision factors is as follows:

[0049] A combined reward function is constructed to balance multi-dimensional feedback, comprehensively considering feedback rewards based on latency, energy consumption, reliability, and accuracy to reflect a balance relationship. The reward value is then calculated using a formula. ;

[0050] Calculate the entropy value of the contribution of each factor to the scheduling decision, quantify the influence weight, and use the formula:

[0051]

[0052] The contribution entropy value was calculated. ,in, The contribution percentage of factor f in decision-making is calculated using attention weights. The lower the contribution entropy value, the more likely the factor is to dominate the decision-making process, thus improving interpretability.

[0053] By using the capability index and task profile output by resource perception as scheduling input, a multi-objective decision-making process is used to generate an allocation scheme. After cross-domain collaborative execution, feedback data is collected, and the strategy is updated through a reward function to achieve intelligent adaptive scheduling of resources in heterogeneous environments.

[0054] Secondly, the intelligent resource scheduling and supply system in a heterogeneous computing environment provided in the embodiments of the present invention specifically includes the following modules:

[0055] Data Acquisition and Fusion Module: Collects heterogeneous data from the heterogeneous computing cluster, obtains core heterogeneous data, performs exponential smoothing and noise reduction to obtain smoothed state values, combines HAT extended tuples, uses geometric mean fusion index to analyze and obtain comprehensive capability scores, and judges whether the communication performance of the heterogeneous computing cluster meets the requirements.

[0056] Adjustment and correction module: For heterogeneous computing clusters with compatible communication performance, the contribution is automatically adjusted by using the similarity between static and dynamic features through attention fusion and feature coupling mechanism; a coupling correction function is constructed to calculate the mixed precision matching degree and determine the suitability between task and hardware precision.

[0057] Collaborative optimization module: For heterogeneous computing clusters adapted to task and hardware precision, a multi-objective coupling penalty mechanism is constructed, which combines exponential product to reflect the nonlinear correlation between objectives; a dynamic penalty function is constructed, the fitness function of the quantum genetic algorithm is optimized, and multi-objective collaborative optimization is performed;

[0058] Coupling and Adaptation Module: Based on the completion of collaborative optimization, a complexity coupling and dynamic adaptation mechanism is constructed. The task segmentation ratio is dynamically adjusted using the nonlinear product of data volume and latency sensitivity. The frequency and reconfiguration time exponential ratio are designed to optimize FPGA bitstream preloading. A synchronous delay coupling model is constructed to optimize state synchronization.

[0059] Feedback Decision Module: Based on the completion of state synchronization optimization, a reward collaboration and loss linkage mechanism is constructed, and multi-dimensional feedback is balanced by combining reward functions; the influence of decision factors is quantified by contribution entropy value to improve interpretability.

[0060] The beneficial effects of this invention are:

[0061] 1. Dynamically time-warped and linearly interpolated multidimensional heterogeneous data are applied, and the comprehensive hardware capability is quantified using a geometric mean fusion index, highlighting the impact of bottleneck indicators such as communication latency. A HAT extended tuple is designed to generate globally unique resource IDs, enabling fast cross-platform resource retrieval. An attention similarity fusion model is proposed, dynamically adjusting the contribution ratio of static and dynamic features through cosine similarity to avoid redundant calculations. A nonlinear matching function is constructed to convert the deviation between hardware-supported accuracy and task theoretical accuracy into a quantifiable loss rate, identifying the suitability for accuracy-sensitive tasks. Multi-objective collaborative optimization of tasks is performed, and dependent features are extracted from the DAG multi-objective collaborative optimization graph. Combined with dynamic complexity indicators, the scheduling strategy is dynamically adjusted to adapt to changes in task structure and reduce scheduling latency.

[0062] 2. A dynamic loss function is designed to fuse latency, energy consumption, overload, and accuracy losses through exponential product, avoiding subjective weight allocation. A fitness function based on the reciprocal of the dynamic loss is proposed, combined with qubit encoding and rotating gate operation, to accelerate convergence while maintaining population diversity. An overload penalty term is automatically triggered to balance task allocation and hardware stability. The task partitioning ratio is dynamically adjusted based on the nonlinear product of data volume and latency sensitivity. A frequency-reconfiguration time exponential ratio optimization preloading strategy is proposed to prioritize high-frequency bitstreams with fast reconfiguration. A correlation function between verification strength, synchronization time, and error rate is constructed to ensure data reliability. A reward signal is designed by comprehensively considering task latency, energy consumption, success rate, and accuracy losses to achieve multi-objective balance. The influence weight of each factor on decision-making is quantified through entropy analysis. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0064] Figure 1 This is a flowchart of the steps of the intelligent resource scheduling and supply method in a heterogeneous computing environment provided in Embodiment 1 of the present invention;

[0065] Figure 2 This is a schematic diagram of the structure of the intelligent resource scheduling and supply system in a heterogeneous computing environment provided in Embodiment 2 of the present invention. Detailed Implementation

[0066] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0067] Example 1: As Figure 1 As shown in the embodiments of the present invention, the intelligent resource scheduling and supply method and system in a heterogeneous computing environment specifically includes the following steps:

[0068] Step 1: Collect heterogeneous data from the heterogeneous computing cluster to obtain core heterogeneous data, perform exponential smoothing and noise reduction to obtain smoothed state values, combine HAT extended tuples, use geometric mean fusion index to analyze and obtain a comprehensive capability score to determine whether the communication performance of the heterogeneous computing cluster meets the requirements.

[0069] In a specific embodiment, heterogeneous data is collected from CPUs, GPUs, FPGAs and edge MCUs in a heterogeneous computing cluster. A distributed perception agent is deployed to cover the heterogeneous resources of CPUs, GPUs, FPGAs and edge MCUs, and N types of core heterogeneous data are collected. The core heterogeneous data includes, but is not limited to: computing power, memory bandwidth, communication latency, power consumption, utilization rate and temperature.

[0070] Real-time data transmission is achieved through Kafka message queues, ensuring that timestamp errors meet the preset error range;

[0071] Linear interpolation is performed on core heterogeneous data of different frequencies to unify the granularity, using the formula:

[0072]

[0073] The core heterogeneous data at time t were calculated. ,in, for, for, for, for, for, To address the issue of time synchronization problems in multi-source data;

[0074] Exponential smoothing denoising is performed on the core heterogeneous data to preserve real-time performance while suppressing high-frequency noise. The original core heterogeneous data at time t and the original core heterogeneous data at time t-1 are fused according to preset weights to calculate the smoothed state value and sort it.

[0075] Based on the obtained smooth state values, HAT tuples are constructed to assign architecture identifier codes to different hardware types. For example, CPU=001, GPU=002, FPGA=003. Combined with HAT tuples, globally unique resource IDs are generated to realize cross-platform resource indexing.

[0076] The interquartile range is used to replace the 3σ method to adapt to smooth state values ​​that are not normally distributed. For any smooth state value sequence, the lower quartile and upper quartile values ​​of the smooth state value sequence are calculated. The difference between the upper quartile and lower quartile values ​​is calculated to obtain the anomaly threshold.

[0077] If the smoothed state value is less than or greater than the outlier threshold, the median of the sequence is used instead of the smoothed state value.

[0078] Combining HAT extended tuples and utilizing the geometric mean fusion index, the weakest link effect is reflected through the formula:

[0079]

[0080] Calculate the comprehensive ability score ,in, To smooth the total number of dimensions of the state values, For dimensional indexing, The k-th dimension smoothed state value after preprocessing. To find the theoretical minimum extreme value, This is the maximum extremum of the standard theory;

[0081] For example, if the GPU's 10 normalized smoothed state values ​​are [0.8, 0.7, 0.9, 0.85, 0.75, 0.82, 0.95, 0.6, 0.88, 0.78], then the overall capability score is obtained by rounding down to one decimal place. If the normalized value of the communication delay drops to 0.3, then It can keenly reflect the deterioration of communication performance;

[0082] Step 2: For heterogeneous computing clusters with compatible communication performance, the contribution is automatically adjusted by using the similarity between static and dynamic features through attention fusion and feature coupling mechanisms; a coupling correction function is constructed to calculate the mixed precision matching degree and determine the suitability between the task and the hardware precision.

[0083] In a specific embodiment, for static features, task submission information is obtained and static features are extracted. Static features include: task type, amount of input data, instruction set dependency, and theoretical accuracy requirement.

[0084] For dynamic features, dynamic features are monitored and obtained during task runtime. Dynamic features include: real-time CPU utilization, memory consumption, and I / O request frequency.

[0085] Dependency features are obtained by analyzing the task DAG graph. These features include: the number of subtasks extracted, the total amount of data dependencies, and the level of parallelism.

[0086] By employing an attention fusion and feature coupling mechanism, and using attention similarity fusion, key features are automatically highlighted through the formula:

[0087]

[0088] Calculate the task profile ,in, For static feature vectors, The cosine similarity between static and dynamic features ranges from [0,1]. The average value of the dynamic characteristics. For dependent feature vectors, Due to the complexity of the dependency, and ;

[0089] If the similarity between static and dynamic features is high, then the contribution ratio of static features is high; if the dependency complexity is high, then the contribution ratio of dependency feature vectors is low, so as to avoid excessive interference from dependency features.

[0090] For example, if the similarity between static and dynamic features in an AI training task is 0.85, then the contribution of static features is 85%. If the complexity depends on the dynamic feature, then the contribution of static features is 85%. Calculating the value to three decimal places, the contribution ratio of the dependent feature vector is obtained from the analysis. ;

[0091] Obtain the hardware-supported accuracy and the task's theoretical accuracy. Quantify the compatibility between the task and hardware accuracy using a mixed-precision matching degree calculation, reflecting the non-linear correlation between hardware-supported accuracy and task-theoretical accuracy, through the formula:

[0092]

[0093] Calculate the mixed precision matching degree ,in, For the theoretical accuracy of the task, To ensure hardware accuracy, Let be the accuracy loss rate of task i on hardware j;

[0094] If the hardware-supported accuracy perfectly matches the theoretical accuracy of the task, then The greater the accuracy deviation or the higher the accuracy loss rate, the smaller the mixed accuracy matching degree.

[0095] The obtained mixed precision matching degree is compared with the preset matching degree threshold to analyze and determine the heterogeneous computing cluster that is compatible with the task and hardware precision.

[0096] If the mixed precision matching degree is greater than or equal to the preset matching degree threshold, it indicates that the task of the heterogeneous computing cluster is compatible with the hardware precision; if the mixed precision matching degree is less than the preset matching degree threshold, it indicates that the task of the heterogeneous computing cluster is not compatible with the hardware precision.

[0097] Step 3: For heterogeneous computing clusters that are adapted to the task and hardware precision, construct a multi-objective coupling penalty mechanism, and combine exponential product to reflect the nonlinear correlation between objectives; construct a dynamic penalty function to avoid rigidity of fixed thresholds; optimize the fitness function of quantum genetic algorithm to perform multi-objective collaborative optimization;

[0098] In a specific embodiment, a multi-objective coupled penalty mechanism is constructed: objective constraints: extract the total latency upper limit and power budget from the task SLA; obtain the resource utilization safety threshold from the system configuration;

[0099] Predictive data: A random forest model is trained using historical scheduling records to predict the latency of task i on resource j. and energy consumption ;

[0100] Decision variables: decision Where 1 indicates that task i is assigned to resource j, and 0 indicates that task i is not assigned to resource j, the maximum utilization rate of the resource pool is obtained. ;

[0101] By integrating latency, energy consumption, overload, and accuracy loss, a dynamic loss function is calculated to avoid subjective weighting, using the formula:

[0102]

[0103] The dynamic loss was calculated. ,in, For the total prediction delay, and , For the total SLA latency budget, For total predicted energy consumption, and , For the power budget of the resource pool, As an overload penalty, and , For accuracy penalty, and ;

[0104] For example, total latency exceeding the total latency budget by 20% means... Energy consumption exceeding the power budget by 10% The exponent term is then rounded to two decimal places. If both are overloaded and Total dynamic loss In this case, latency and energy consumption should be optimized first.

[0105] Based on the obtained dynamic loss, the reciprocal of the dynamic loss is used as the fitness to construct an optimized quantum genetic algorithm fitness function, guiding the optimized quantum genetic algorithm to quickly converge to the optimal solution, through the formula:

[0106]

[0107] The fitness of the quantum genetic algorithm is calculated. The smaller the dynamic loss DLF, the higher the fitness. The quantum genetic algorithm searches for the optimal allocation scheme through qubit encoding, thus stabilizing the decision delay.

[0108] Step 4: Construct a complexity coupling and dynamic adaptation mechanism, dynamically adjust the task partitioning ratio using the nonlinear product of data volume and latency sensitivity; design the frequency and reconfiguration time exponential ratio to optimize FPGA bitstream preloading; construct a synchronous delay coupling model and perform state synchronization optimization.

[0109] In a specific embodiment, the edge processing ratio is dynamically adjusted to balance latency and transmission cost, using the formula:

[0110]

[0111] The proportion of edge tasks was calculated. ,in, To transmit data volume, The dynamic data threshold is determined by the level of latency sensitivity; the higher the latency sensitivity, the smaller the dynamic data threshold.

[0112] Optimize the preloading strategy to reduce reconfiguration time, obtain the historical usage frequency of bitstream b and combine it with the exponential ratio of cluster reconfiguration time to minimum cluster reconfiguration time to obtain bitstream priority;

[0113] Bitstreams that are used frequently and reconfigured quickly have higher priority and are preloaded into the FPGA cache to reduce reconfiguration time.

[0114] To balance synchronization latency and data reliability, optimize state synchronization: in, To verify the strength, To synchronize time, The longer the synchronization time or the higher the error rate, the stronger the verification strength.

[0115] Step 5: Based on the completion of state synchronization optimization, construct a reward coordination and loss linkage mechanism, balance multi-dimensional feedback by combining reward functions, and quantify the impact of decision factors by contribution entropy value to improve interpretability;

[0116] In a specific embodiment, feedback indicators are extracted from the execution log in real time. These feedback indicators include: actual task execution latency, power overrun ratio, task success rate, and total accuracy loss.

[0117] A combined reward function is constructed to balance multi-dimensional feedback, comprehensively considering feedback rewards based on latency, energy consumption, reliability, and accuracy, reflecting a balance relationship. This is achieved through the formula:

[0118]

[0119] Calculate the reward value ,in, The actual execution delay of the task. The delay specified by the SLA, The percentage of power overrun. To improve the success rate of the task, This represents the total loss of accuracy.

[0120] Calculate the entropy value of the contribution of each factor to the scheduling decision, quantify the influence weight, and use the formula:

[0121]

[0122] The contribution entropy value was calculated. ,in, The contribution percentage of factor f in decision-making is calculated using attention weights. The lower the contribution entropy value, the more likely the factor is to dominate the decision-making process, thus improving interpretability.

[0123] By using the capability index and task profile output by resource perception as scheduling input, a multi-objective decision-making process is used to generate an allocation scheme. After cross-domain collaborative execution, feedback data is collected, and the strategy is updated through a reward function to achieve intelligent adaptive scheduling of resources in heterogeneous environments.

[0124] Example 2: As Figure 2 As shown in the embodiment of the present invention, the intelligent resource scheduling and supply system in a heterogeneous computing environment specifically includes the following modules:

[0125] Data Acquisition and Fusion Module: Collects heterogeneous data from the heterogeneous computing cluster, obtains core heterogeneous data, performs exponential smoothing and noise reduction to obtain smoothed state values, combines HAT extended tuples, uses geometric mean fusion index to analyze and obtain comprehensive capability scores, and judges whether the communication performance of the heterogeneous computing cluster meets the requirements.

[0126] Adjustment and correction module: For heterogeneous computing clusters with compatible communication performance, the contribution is automatically adjusted by using the similarity between static and dynamic features through attention fusion and feature coupling mechanism; a coupling correction function is constructed to calculate the mixed precision matching degree and determine the suitability between task and hardware precision.

[0127] Collaborative optimization module: For heterogeneous computing clusters adapted to task and hardware precision, a multi-objective coupling penalty mechanism is constructed, which combines exponential product to reflect the nonlinear correlation between objectives; a dynamic penalty function is constructed, the fitness function of the quantum genetic algorithm is optimized, and multi-objective collaborative optimization is performed;

[0128] Coupling and Adaptation Module: Based on the completion of collaborative optimization, a complexity coupling and dynamic adaptation mechanism is constructed. The task segmentation ratio is dynamically adjusted using the nonlinear product of data volume and latency sensitivity. The frequency and reconfiguration time exponential ratio are designed to optimize FPGA bitstream preloading. A synchronous delay coupling model is constructed to optimize state synchronization.

[0129] Feedback Decision Module: Based on the completion of state synchronization optimization, a reward collaboration and loss linkage mechanism is constructed, and multi-dimensional feedback is balanced by combining reward functions; the influence of decision factors is quantified by contribution entropy value to improve interpretability.

[0130] The above provides a detailed description of one embodiment of the present invention, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. The above formulas are all dimensionless numerical calculations, and the formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world situation. The preset parameters in the formulas are set by those skilled in the art based on actual conditions and historical experience, and can be adjusted according to actual conditions. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A method for intelligent resource scheduling and supply in a heterogeneous computing environment, characterized in that, Includes the following steps: Heterogeneous data is collected from the heterogeneous computing cluster to obtain core heterogeneous data. Exponential smoothing and noise reduction are then performed to obtain smoothed state values. Combined with HAT extended tuples, geometric mean fusion index is used to analyze and obtain a comprehensive capability score to determine whether the communication performance of the heterogeneous computing cluster meets the requirements. The method for obtaining the comprehensive ability score is as follows: Exponential smoothing denoising is performed on the core heterogeneous data to preserve real-time performance while suppressing high-frequency noise. The original core heterogeneous data at time t and the original core heterogeneous data at time t-1 are fused according to preset weights to calculate the smoothed state value and sort it. Based on the obtained smooth state values, construct HAT tuples, assign architecture identifier codes to different hardware types, and generate globally unique resource IDs by combining HAT tuples. For any smooth state value sequence, calculate the lower quartile and upper quartile values ​​of the smooth state value sequence, and calculate the difference between the upper quartile and lower quartile values ​​to obtain the anomaly threshold; If the smoothed state value is less than or greater than the outlier threshold, the median of the sequence is used instead of the smoothed state value. Combining HAT extended tuples and utilizing the geometric mean fusion index, the formula is: Calculate the comprehensive ability score ,in, To smooth the total number of dimensions of the state values, For dimensional indexing, The k-th dimension smoothed state value after preprocessing. To find the theoretical minimum extreme value, This is the maximum extremum of the standard theory; For heterogeneous computing clusters with compatible communication performance, an attention fusion and feature coupling mechanism is used to automatically adjust the contribution of static and dynamic features based on their similarity. A coupling correction function is constructed to calculate the mixed precision matching degree and determine the compatibility between the task and the hardware precision. For heterogeneous computing clusters that are adapted to the task and hardware precision, a multi-objective coupling penalty mechanism is constructed, which combines exponential product to reflect the nonlinear correlation between objectives; a dynamic penalty function is constructed, the fitness function of the quantum genetic algorithm is optimized, and multi-objective collaborative optimization is performed. Based on the completion of collaborative optimization, a complexity coupling and dynamic adaptation mechanism is constructed, and the task partitioning ratio is dynamically adjusted by the nonlinear product of data volume and latency sensitivity; the frequency and reconfiguration time exponential ratio are designed to optimize FPGA bitstream preloading; a synchronous delay coupling model is constructed to optimize state synchronization. Based on the completion of state synchronization optimization, a reward coordination and loss linkage mechanism is constructed, and multi-dimensional feedback is balanced by combining reward functions; the influence of decision factors is quantified by contribution entropy value to improve interpretability.

2. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for automatically adjusting contribution is as follows: For static features, obtain task submission information and extract static features; For dynamic features, monitor and acquire dynamic features during task execution; Dependency features are obtained through task DAG graph analysis; By employing an attention fusion and feature coupling mechanism, and using attention similarity fusion, key features are automatically highlighted through the formula: Calculate the task profile ,in, For static feature vectors, The cosine similarity between static and dynamic features ranges from [0,1]. The average value of the dynamic characteristics. For dependent feature vectors, Due to the complexity of the dependency, and ; If the static and dynamic features have high similarity, then the static features contribute a higher proportion. If the dependency complexity is high, the contribution ratio of the dependency feature vector is low.

3. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for calculating the mixed precision matching degree is as follows: Obtain the hardware-supported accuracy and the task's theoretical accuracy, quantify the compatibility between the task and hardware accuracy through mixed-precision matching degree calculation, and reflect the non-linear correlation between hardware-supported accuracy and task-theoretical accuracy, using the formula: Calculate the mixed precision matching degree ,in, For the theoretical accuracy of the task, To ensure hardware accuracy, Let be the accuracy loss rate of task i on hardware j; If the hardware-supported accuracy perfectly matches the theoretical accuracy of the task, then The greater the accuracy deviation or the higher the accuracy loss rate, the smaller the mixed accuracy matching degree. The obtained mixed precision matching degree is compared with the preset matching degree threshold to analyze and determine the heterogeneous computing cluster that is compatible with the task and hardware precision.

4. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for constructing the multi-objective coupled penalty mechanism is as follows: Construct a multi-objective coupled penalty mechanism: Objective constraints: extract the total latency upper limit and power budget from the task SLA; obtain the resource utilization safety threshold from the system configuration; Predictive data: A random forest model is trained using historical scheduling records to predict the latency of task i on resource j. and energy consumption ; Decision variables: decision Where 1 indicates that task i is assigned to resource j, and 0 indicates that task i is not assigned to resource j, the maximum utilization rate of the resource pool is obtained. .

5. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for optimizing the fitness function of the quantum genetic algorithm is as follows: By integrating latency, energy consumption, overload, and accuracy loss, a dynamic loss function is calculated to avoid subjective weighting, using the formula: The dynamic loss was calculated. ,in, For the total prediction delay, and , For the total SLA latency budget, For total predicted energy consumption, and , For the power budget of the resource pool, As an overload penalty, and , For accuracy penalty, and ; Based on the obtained dynamic loss, the reciprocal of the dynamic loss is used as the fitness to construct an optimized quantum genetic algorithm fitness function, guiding the optimized quantum genetic algorithm to quickly converge to the optimal solution, through the formula: The fitness of the quantum genetic algorithm is calculated. The quantum genetic algorithm searches for the optimal allocation scheme by encoding qubits.

6. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for optimizing FPGA bitstream preloading is as follows: Dynamically adjust the edge processing ratio to balance latency and transmission cost, using the formula: The proportion of edge tasks was calculated. ,in, To transmit data volume, For dynamic data thresholds; Optimize the preloading strategy to reduce reconfiguration time, obtain the historical usage frequency of bitstream b and combine it with the exponential ratio of cluster reconfiguration time to minimum cluster reconfiguration time to obtain bitstream priority; Bitstreams that are used frequently and reconfigured quickly have higher priority and are preloaded into the FPGA cache to reduce reconfiguration time.

7. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for performing state synchronization optimization is as follows: To balance synchronization latency and data reliability, optimize state synchronization: in, To verify the strength, To synchronize time, This represents the error rate.

8. The intelligent resource scheduling and supply method in a heterogeneous computing environment according to claim 1, characterized in that, The method for quantifying the influence of decision-making factors is as follows: A combined reward function is constructed to balance multi-dimensional feedback, comprehensively considering feedback rewards based on latency, energy consumption, reliability, and accuracy to reflect a balance relationship. The reward value is then calculated using a formula. ; Calculate the entropy value of the contribution of each factor to the scheduling decision, quantify the influence weight, and use the formula: The contribution entropy value was calculated. ,in, The contribution percentage of factor f in decision-making is calculated using attention weights. The lower the contribution entropy value, the more likely the factor is to dominate the decision-making process, thus improving interpretability. By using the capability index and task profile output by resource perception as scheduling input, a multi-objective decision-making process is used to generate an allocation scheme. After cross-domain collaborative execution, feedback data is collected, and the strategy is updated through a reward function to achieve intelligent adaptive scheduling of resources in heterogeneous environments.

9. An intelligent resource scheduling and supply system for a heterogeneous computing environment, the supply system being used to execute the supply method according to any one of claims 1-8, characterized in that, include: Data Acquisition and Fusion Module: Collects heterogeneous data from the heterogeneous computing cluster, obtains core heterogeneous data, performs exponential smoothing and noise reduction to obtain smoothed state values, combines HAT extended tuples, uses geometric mean fusion index to analyze and obtain comprehensive capability scores, and judges whether the communication performance of the heterogeneous computing cluster meets the requirements. Adjustment and correction module: For heterogeneous computing clusters with compatible communication performance, the contribution is automatically adjusted by using the similarity between static and dynamic features through attention fusion and feature coupling mechanism; a coupling correction function is constructed to calculate the mixed precision matching degree and determine the suitability between task and hardware precision. Collaborative optimization module: For heterogeneous computing clusters adapted to task and hardware precision, a multi-objective coupling penalty mechanism is constructed, which combines exponential product to reflect the nonlinear correlation between objectives; a dynamic penalty function is constructed, the fitness function of the quantum genetic algorithm is optimized, and multi-objective collaborative optimization is performed; Coupling and Adaptation Module: Based on the completion of collaborative optimization, a complexity coupling and dynamic adaptation mechanism is constructed. The task segmentation ratio is dynamically adjusted using the nonlinear product of data volume and latency sensitivity. The frequency and reconfiguration time exponential ratio are designed to optimize FPGA bitstream preloading. A synchronous delay coupling model is constructed to optimize state synchronization. Feedback Decision Module: Based on the completion of state synchronization optimization, a reward collaboration and loss linkage mechanism is constructed, and multi-dimensional feedback is balanced by combining reward functions; the influence of decision factors is quantified by contribution entropy value to improve interpretability.

Images