Software load computing resource virtualization allocation method and device, equipment and medium

Abstract

The invention relates to a software load computing resource virtualization allocation method and device, equipment and a medium. The method comprises the following steps: acquiring a task description vector of a task and real-time state data of computing resources; according to the task description vector of the task and a preset historical performance database, computing resources needed when the task is started are analyzed, and a starting feature vector is obtained; according to the starting feature vector and the real-time state data, calculating a computing power hotspot distribution diagram of the starting task; the computing power hotspot distribution map is used for representing the use condition of each computing resource at different time points when the task is started; based on the computing power hotspot distribution map, the task description vector and the real-time state data, simulating a competition condition when the task is started, and obtaining resource competition intensity quantitative data; generating a task pre-allocation scheme according to the resource competition intensity quantitative data and the real-time state data; the pre-allocation scheme is used for representing the allocation relationship between the task and each computing resource. By adopting the method, the computing power resource utilization rate can be improved.

Description

Technical Field

[0001] This invention belongs to the field of computing resource allocation, and in particular relates to a software-based virtualized allocation method, apparatus, equipment and medium for computing resources. Background Technology

[0002] With the development of computing resource allocation technology, dynamic resource allocation technology based on software-based payloads and virtualization has emerged. This technology forms a shared resource pool by abstracting hardware resources, which enables the isolated operation and rapid deployment of multiple payload tasks, and in turn gives rise to resource allocation methods represented by static resource partitioning and priority-based dynamic scheduling.

[0003] Existing priority-based dynamic scheduling and static resource partitioning resource allocation technologies typically employ static resource reservation or allocation strategies based on historical average resource usage before task startup. Specifically, the resource manager pre-allocates a fixed number of CPU cores and memory space in a dedicated resource partition based on the maximum resource requirements declared by the task or its recent average usage. When multiple tasks start simultaneously, the scheduler allocates available resources sequentially according to preset task priorities, following a first-come, first-served or priority preemption principle, and dynamically adjusts resource quotas by monitoring system load in real time.

[0004] However, existing priority-based dynamic scheduling and static resource partitioning resource allocation technologies lack fine-grained perception of resource consumption patterns during the load startup phase and cannot effectively predict instantaneous resource hotspots. In high-concurrency startup scenarios, the inability to accurately predict the distribution of resource hotspots before task startup makes it difficult for the system to prepare for the upcoming resource storms, leading to frequent resource contention and response delays during the critical task initialization phase, resulting in low utilization of computing resources. Summary of the Invention

[0005] Therefore, it is necessary to provide a software-based virtualization allocation method, apparatus, equipment, and medium for computing resources that can improve the utilization rate of computing resources, addressing the aforementioned technical problems.

[0006] Firstly, this application provides a software-based method for virtualizing and allocating computing resources for payloads, including:

[0007] Obtain the task description vector and real-time status data of computing resources for the task; the task description vector includes the task type parameter, completion time limit parameter, and computing power requirement parameter.

[0008] Based on the task description vector and the preset historical performance database, the computing resources required to start the task are analyzed to obtain the startup feature vector; the historical performance database includes the task type parameters and computing resource consumption time series of historical tasks.

[0009] Based on the startup feature vector and real-time status data, a computing power hotspot distribution map of the startup task is calculated; the computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points during the startup task.

[0010] Based on the computing power hotspot distribution map, task description vector and real-time status data, the competition situation at task startup is simulated to obtain quantitative data on resource competition intensity.

[0011] Based on the quantitative data of resource competition intensity and real-time status data, a pre-allocation scheme for tasks is generated; the pre-allocation scheme is used to characterize the allocation relationship between tasks and various computing resources.

[0012] Furthermore, based on the task description vector and a pre-defined historical performance database, the computational resources required to start the task are analyzed to obtain a startup feature vector, including:

[0013] Select historical tasks from the historical performance database that have the same task type parameters as the task to form a reference task set;

[0014] For each historical task in the reference task set, the computational resource consumption time series of the historical task is time-aligned to obtain the time-aligned computational resource consumption sequence of each historical task.

[0015] For each historical task in the reference task set, the consumption of computing resources when starting the historical task is quantified based on the time-aligned computing resource consumption sequence, and the start-up consumption feature vector of each historical task is obtained.

[0016] The startup feature vector is obtained based on the startup consumption feature vector of each historical task in the reference task set.

[0017] Furthermore, based on the startup feature vector and real-time status data, a computing power hotspot distribution map of the startup task is calculated, including:

[0018] Based on the initiation feature vector and real-time state data, the spatiotemporal cross feature vector is constructed using the following formula:

[0019]

[0020] in, It is a spatiotemporal cross feature vector. It is the initiation feature vector. It is real-time status data. It is the element-wise multiplication symbol;

[0021] Based on the initiation feature vector, the rate of change of computing resources is calculated to obtain the time-series derived feature vector;

[0022] The spatiotemporal cross feature vector, time-series derived feature vector, startup feature vector, and real-time status data are input into the trained computing power hotspot distribution prediction model to obtain the computing power hotspot distribution map of the startup task.

[0023] The computing power hotspot distribution prediction model was obtained through the following method:

[0024] Obtain preset historical startup data, which includes startup feature vectors of historical tasks, real-time status data, and corresponding computing power hotspot distribution maps;

[0025] Using the historical task startup feature vector and real-time status data as input features, and the corresponding computing power hotspot distribution map as the prediction target, the historical startup data is divided into a model training set and a model validation set according to a preset ratio.

[0026] An initial computing power hotspot distribution prediction model is constructed. The pseudo residuals between the input features in the model training set and the prediction target are calculated. A decision tree is iteratively constructed based on the pseudo residuals to optimize the initial computing power hotspot distribution prediction model until the preset iteration termination condition is met, thus obtaining the preliminary computing power hotspot distribution prediction model.

[0027] The preliminary computing power hotspot distribution prediction model is tested using a model validation set until it meets the preset accuracy requirements, thus obtaining a trained computing power hotspot distribution prediction model.

[0028] Furthermore, the loss function of the computing power hotspot distribution prediction model is:

[0029]

[0030] in, It is the model loss. It is the total number of historical startup data. Any historical task , It is a historical mission The result of the loss It is the total number of decision trees. Any decision tree The total number of leaf nodes, It is a decision tree The output vector, It is the quantity penalty coefficient. It is the weighted penalty coefficient;

[0031] The first loss term is obtained using the following formula:

[0032]

[0033] in, It is a historical mission The result of the loss It is a computing power hotspot distribution prediction model for historical tasks The predicted value, It is a historical mission The distribution map of computing power hotspots It is the threshold for result loss.

[0034] Furthermore, based on the computing power hotspot distribution map, task description vectors, and real-time status data, the competition situation at task startup is simulated to obtain quantitative data on resource competition intensity, including:

[0035] Based on preset simulation scenario coefficients, adjust real-time status data to obtain scenario status data;

[0036] For each simulated scenario coefficient, based on the computing power hotspot distribution map and the scenario status data, the competition intensity value of the simulated scenario coefficient is calculated; the competition intensity value is used to characterize the computing resource usage intensity corresponding to the start of the task.

[0037] Based on the competition intensity values ​​of each simulated scenario coefficient, the local extrema of the competition intensity values ​​are detected to obtain the local extrema data of each competition intensity value;

[0038] The competition intensity values ​​of each simulated scenario coefficient are sorted based on the local extreme value data of each competition intensity value. Combined with the local extreme value data corresponding to the competition intensity values, the quantitative data of resource competition intensity is obtained.

[0039] Furthermore, based on quantitative data on resource competition intensity and real-time status data, a pre-allocation plan for tasks is generated, including:

[0040] Based on the quantitative data of resource competition intensity, the preset start time threshold, and the preset computing resource capacity threshold, the allocation constraints between tasks and each computing resource are constructed.

[0041] Using tasks and computing resources as graph nodes, and based on the computing power hotspot distribution map and real-time status data, fill the connection edges between graph nodes to obtain a complete allocation bipartite graph;

[0042] Based on the allocation constraints, the matching relationship between tasks and computing resources is solved to obtain candidate pre-allocation schemes; the candidate pre-allocation schemes form a candidate pre-allocation scheme set.

[0043] Based on the completion time limit parameters, computing power requirement parameters, and real-time status data, a candidate pre-allocation scheme is selected from the candidate pre-allocation scheme set to obtain the task pre-allocation scheme.

[0044] Furthermore, the method also includes:

[0045] Obtain execution feedback data for the pre-allocation scheme; the execution feedback data is used to characterize the changes in the real-time status data of computing resources when the pre-allocation scheme is executed;

[0046] Based on execution feedback data, computing power hotspot distribution map and resource competition intensity quantification data, the prediction accuracy of computing resource usage at task startup is quantified to obtain prediction accuracy data of pre-allocation scheme;

[0047] Based on prediction accuracy data, task description vectors, and execution feedback data, the historical performance database is updated to obtain the updated historical performance database.

[0048] Secondly, this application also provides a software-based payload computing resource virtualization allocation device, comprising:

[0049] The data acquisition module is used to acquire the task description vector and real-time status data of computing resources; the task description vector includes the task type parameter, completion time limit parameter and computing power requirement parameter.

[0050] The resource consumption fitting module is used to analyze the computing resources required to start a task based on the task description vector and a preset historical performance database, and obtain the startup feature vector. The historical performance database includes the task type parameters and computing resource consumption time series of historical tasks.

[0051] The resource hotspot calculation module is used to calculate the computing power hotspot distribution map of the startup task based on the startup feature vector and real-time status data; the computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points during the startup task.

[0052] The competition calculation module is used to simulate the competition situation when a task starts based on the computing power hotspot distribution map, task description vector and real-time status data, and obtain quantitative data on resource competition intensity.

[0053] The scheme generation module is used to generate pre-allocation schemes for tasks based on quantification data of resource contention intensity and real-time status data; the pre-allocation schemes are used to characterize the allocation relationship between tasks and various computing resources.

[0054] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the software-based load computing resource virtualization allocation methods described in the first aspect of this application.

[0055] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the software-based load computing resource virtualization allocation methods described in the first aspect of this application.

[0056] The aforementioned software-based payload virtualization allocation method, apparatus, device, and medium acquires the task description vector and real-time status data of computing resources. The task description vector includes task type parameters, completion time limit parameters, and computing power requirement parameters. Based on the task description vector and a pre-set historical performance database, the computing resources required to start the task are analyzed to obtain a startup feature vector. The historical performance database includes task type parameters and computing resource consumption time series of historical tasks. Based on the startup feature vector and real-time status data, a computing power hotspot distribution map of the startup task is calculated. The computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points when the task starts. Based on the computing power hotspot distribution map, task description vector, and real-time status data, the competition situation at task startup is simulated to obtain quantified data of resource competition intensity. Based on the quantified data of resource competition intensity and real-time status data, a pre-allocation scheme for the task is generated. The pre-allocation scheme is used to characterize the allocation relationship between the task and each computing resource. It effectively solves the core problems of traditional resource allocation methods, such as long startup delays and fierce resource competition caused by a lack of forward-looking perspective. It significantly improves the response speed and resource utilization efficiency when dealing with high-concurrency task startup scenarios, while enhancing the determinism and reliability of task execution under complex load environments. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart illustrating a software-based virtualization allocation method for computing resources is provided in one embodiment of this application.

[0059] Figure 2 A schematic diagram of a software-based load computing resource virtualization allocation device provided in one embodiment of this application;

[0060] Figure 3 This is a schematic diagram of a computer device for a software-based virtualization allocation method for computing resources, provided as an embodiment of this application. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] In one embodiment, such as Figure 1 As shown, a software-based virtualization allocation method for computing resources is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S101-S105, wherein:

[0063] S101, obtain the task description vector and real-time status data of computing resources; the task description vector includes the task type parameter, completion time limit parameter and computing power requirement parameter.

[0064] Specifically, the terminal acquires the task description vector of the task to be executed and real-time status data of the computing resources available to execute the task and drive the software-based payload. The task description vector is the metadata description of the task, including task type parameters, completion deadline parameters, and computing power requirement parameters. The task type parameter characterizes the type of task and the type of payload required. The completion deadline parameter describes the deadline for task execution. The computing power requirement parameters describe the task's requirements for CPU core count, memory capacity, and cache capacity. Computing resources include, but are not limited to, CPU, memory system, and cache system. Real-time status data includes, but is not limited to, CPU utilization, available memory capacity, and cache status parameters at various levels. For example, the task could be remote sensing imaging and signal processing, and the task description vector carries a task label to indicate that it represents the task corresponding to the task label. The terminal can acquire task description vectors for multiple tasks simultaneously. Optionally, the real-time status data carries resource labels to distinguish the real-time status data of different computing resources available to execute the task, such as... CPU core utilization rate It is a memory node Available memory capacity It is a memory node Available memory capacity, of which , and It is a resource tag.

[0065] S102, based on the task description vector and the preset historical performance database, analyze the computing resources required to start the task and obtain the startup feature vector; the historical performance database includes the task type parameters and computing resource consumption time series of historical tasks.

[0066] Specifically, the terminal retrieves historical tasks with the same task type parameter from a preset historical performance database based on the task type parameter in the task description vector. It then obtains the computational resource consumption time series of these historical tasks and extracts the computational resource consumption features from these time series to obtain the startup feature vector of the current task. The preset historical performance database includes the task type parameter and computational resource consumption time series of historical tasks. The computational resource consumption time series characterizes the changes in CPU utilization, available memory capacity, and cache status parameters of the computational resources used to start the historical task over time. The preset historical performance database can be compiled based on actual work. For example, the computational resource consumption time series can take the following form: ,in, It calculates the time series of resource consumption. It is the first A historical mission It is the first A historical task at the start time node CPU utilization at that time It is the first A historical task at the start time node Available memory capacity at that time It is the first A historical task at the start time node The cache state parameters at each level at that time, among which , It is the first The total time for the launch of each historical mission.

[0067] S103, Calculate the computing power hotspot distribution map of the startup task based on the startup feature vector and real-time status data; the computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points when the startup task is launched.

[0068] Specifically, the terminal will initiate the fusion calculation of feature vectors and real-time status data to generate a computing power hotspot distribution map. This map is a spatiotemporal two-dimensional data matrix. The vertical dimension represents different types of computing resources, including but not limited to CPU utilization, memory capacity, and cache status parameters at various levels with resource labels. The horizontal dimension represents the time node after task initiation, and the matrix element values ​​characterize the expected usage intensity of the corresponding resource at a specific time node. The terminal uses the gradient boosting decision tree algorithm for prediction. Gradient boosting decision tree is an ensemble learning method that iteratively constructs multiple weak decision trees and combines them with weights. Leveraging the feature combination advantages and nonlinear fitting capabilities of the gradient boosting decision tree algorithm, it learns the spatiotemporal patterns of resource usage from the input features and outputs a probabilistic prediction of the hotspot distribution.

[0069] S104, based on the computing power hotspot distribution map, task description vector and real-time status data, simulates the competition situation when the task starts and obtains quantitative data on the intensity of resource competition.

[0070] Specifically, the terminal performs multi-dimensional competition simulation based on the computing power hotspot distribution map. It overlays the hotspot distribution map with real-time status data of current computing resources to simulate the competition among resource nodes in a multi-task concurrent scenario, outputting quantitative data on resource competition intensity. This quantitative data reflects the changes in competition among CPU utilization, memory capacity, and cache status parameters (labeled with resource tags) as the task starts.

[0071] S105: Based on the resource competition intensity quantification data and real-time status data, generate a pre-allocation scheme for tasks; the pre-allocation scheme is used to characterize the allocation relationship between tasks and various computing resources.

[0072] Specifically, the terminal uses quantified data on resource contention intensity and real-time status data, combined with preset constraints, to obtain a pre-allocation scheme for tasks using the Hungarian algorithm. This pre-allocation scheme indicates how tasks will be assigned to corresponding computing resources for execution. The Hungarian algorithm is a combinatorial optimization algorithm that solves for maximum weight matching in a bipartite graph in polynomial time. For example, the preset constraints indicate the matching conditions between tasks and computing resources, and may include, but are not limited to, time constraints, capacity constraints, and avoidance constraints, which can be set according to the actual workload.

[0073] This embodiment provides a software-based virtualized allocation method for computing resources. It acquires the task description vector and real-time status data of computing resources; accurately grasps the resource consumption pattern of the task using historical data-driven feature extraction to obtain a startup feature vector; utilizes machine learning algorithms to accurately predict future resource hotspots to obtain a computing power hotspot distribution map; and proactively assesses resource competition risks through simulation to obtain quantitative data on resource competition intensity. Based on optimization algorithms, a preventative resource allocation strategy is generated, resulting in a pre-allocation scheme for the task. This effectively solves the core problems of traditional resource allocation methods, such as long startup delays and intense resource competition due to a lack of forward-looking perspective. It improves the response speed and resource utilization efficiency in high-concurrency task startup scenarios, while enhancing the determinism and reliability of task execution under complex load environments.

[0074] In one embodiment, based on the task description vector and a preset historical performance database, the computing resources required to start the task are analyzed to obtain a startup feature vector, including:

[0075] S201. Select historical tasks from the historical performance database that have the same task type parameters as the task to form a reference task set.

[0076] Specifically, the terminal uses the task type parameter of the current task to filter historical tasks with the same task type parameter from the historical performance database to form a reference task set.

[0077] S202, for each historical task in the reference task set, time-align the computational resource consumption time series of the historical task to obtain the time-aligned computational resource consumption sequence of each historical task.

[0078] Specifically, for each historical task in the reference task set, the terminal uses a dynamic time warping algorithm to map the computational resource consumption time series with different start times onto a unified relative time axis. This eliminates the sequence length differences caused by different task start times, resulting in a time-aligned computational resource consumption sequence for each historical task. The dynamic time warping algorithm is a time-series processing method that achieves non-linear sequence alignment by finding the path with the minimum cumulative distance, preserving the key morphological features of the sequence from the impact of time scaling. The resulting time-aligned computational resource consumption sequences for each historical task ensure comparability across all historical tasks on the same relative time scale. For example, it normalizes the start-up processes of tasks with different durations to relative progress from 0% to 100%, facilitating subsequent unified feature quantification analysis.

[0079] S203. For each historical task in the reference task set, the consumption of computing resources when starting the historical task is quantified based on the time-aligned computing resource consumption sequence, and the start-up consumption feature vector of each historical task is obtained.

[0080] Specifically, for each historical task in the reference task set, the terminal extracts feature indicators that characterize the resource usage pattern during the task's startup phase, generating a startup consumption feature vector for each historical task. This startup consumption feature vector includes at least the peak sequence of the CPU utilization sequence, the integral area of ​​the memory bandwidth sequence, and the fluctuation frequency of the cache miss rate sequence. The peak position of the CPU utilization sequence is identified using a peak-finding algorithm, representing the relative time point of the computationally intensive phase during task startup, with the peak height reflecting the maximum computational load of that phase. The integral area of ​​the memory bandwidth sequence is obtained by calculating the area under the sequence curve using the trapezoidal method, representing the total amount of memory data transfer accumulated during the entire startup phase. The fluctuation frequency of the cache miss rate sequence is extracted using Fourier transform to extract the dominant frequency component, reflecting the spatial locality of change in the task access pattern.

[0081] S204. Based on the startup consumption feature vector of each historical task in the reference task set, the startup feature vector is obtained.

[0082] Specifically, the terminal aggregates the startup consumption feature vectors of each historical task in the reference task set to obtain a startup feature vector. The terminal then uses a weighted average algorithm to process the peak sequence of CPU utilization, the integral area of ​​memory bandwidth, and the fluctuation frequency of cache miss rate included in the startup consumption feature vector. The weights are dynamically adjusted based on the sample quality and time proximity of the historical tasks. After weighted averaging the startup consumption feature vectors of each historical task, the startup feature vector is obtained.

[0083] This embodiment provides a software-based virtualization allocation method for computing resources. It matches the task type parameters of the current task with historical tasks included in a historical performance database to obtain historical tasks with the same task type parameters, forming a reference task set. The computational resource consumption time series of each historical task in the reference task set is then transformed into a startup feature vector. This method solves the problem of inconsistent time series data lengths while ensuring the relevance of the analysis samples. It also quantifies resource consumption characteristics from multiple perspectives to extract feature patterns, effectively capturing the unique startup patterns of different task types. This provides reliable input data for subsequent resource prediction, improves the accuracy and reliability of hotspot distribution prediction, and thus enhances response speed and resource utilization efficiency in high-concurrency task startup scenarios.

[0084] In one embodiment, the computational hotspot distribution map of the startup task is calculated based on the startup feature vector and real-time status data, including:

[0085] S301, based on the initiation feature vector and real-time state data, constructs a spatiotemporal cross feature vector using the following formula:

[0086]

[0087] in, It is a spatiotemporal cross feature vector. It is the initiation feature vector. It is real-time status data. It is the symbol for element-wise multiplication.

[0088] Specifically, the terminal fuses the startup feature vector with real-time status data using a formula to construct a spatiotemporal cross-feature vector. Among these features, the startup feature vector... Real-time status data is used to characterize the inherent resource requirement patterns of a task. Resource availability of computing resources corresponding to resource tags used to reflect real-time status data is determined. A spatiotemporal cross-feature vector is obtained by element-wise multiplication of the startup feature vector and real-time status data to capture the interaction between task demand and computing resource status. For example, when the CPU core demand feature value in the startup feature vector is 0.8, and the availability of the corresponding core in the real-time status data of any computing resource is 0.3, the cross-feature will amplify this supply-demand imbalance, making resource scarcity more apparent.

[0089] S302, based on the startup feature vector, calculates the rate of change of computing resources to obtain the time-series derived feature vector.

[0090] Specifically, the terminal extracts the rate of change of each computing resource consumption index over time from the startup feature vector through differential operations, obtaining a time-series derived feature vector. This time-series derived feature vector reflects the dynamic characteristics of resource demand during task startup. For example, numerical differentiation methods can be used to calculate the rate of change of each computing resource consumption index over time.

[0091] S303 inputs the spatiotemporal cross feature vector, time-series derived feature vector, startup feature vector and real-time status data into the trained computing power hotspot distribution prediction model to obtain the computing power hotspot distribution map of the startup task.

[0092] Specifically, the terminal inputs the spatiotemporal cross feature vector, time-series derived feature vector, startup feature vector, and real-time status data into the trained computing power hotspot distribution prediction model to obtain the computing power hotspot distribution map of the startup task. The computing power hotspot distribution prediction model is constructed based on historical startup data using a gradient boosting decision tree framework. This model integrates multiple weak decision trees to form a strong predictor, capable of handling complex nonlinear relationships between high-dimensional features and outputting the hotspot probability distribution of each time slice and resource unit during the task startup process.

[0093] The computing power hotspot distribution prediction model was obtained through the following method:

[0094] S3031, Obtain preset historical startup data, which includes the startup feature vector of historical tasks, real-time status data, and corresponding computing power hotspot distribution map.

[0095] Specifically, historical startup data is acquired, including startup feature vectors, real-time status data, and corresponding computing power hotspot distribution maps of historical tasks. The preset historical startup data can be collected based on actual work requirements.

[0096] S3032 uses the startup feature vector of historical tasks and real-time status data as input features, and the corresponding computing power hotspot distribution map as the prediction target. It divides the historical startup data into a model training set and a model validation set according to a preset ratio.

[0097] Specifically, the terminal uses the startup feature vector and real-time status data from historical task initiation as input features, and the corresponding computing power hotspot distribution map as the prediction target, forming a complete sample set of input feature-prediction target mapping relationships. The historical startup data is then divided into a model training set and a model validation set according to a preset ratio. The model training set is used for model parameter learning, while the model validation set is used for hyperparameter tuning and model performance evaluation. Illustratively, the preset ratio can be dynamically adjusted based on the model recognition accuracy in actual work; the default setting is model training set: model validation set = 7:3.

[0098] S3033, construct an initial computing power hotspot distribution prediction model, calculate the pseudo residual between the input features in the model training set and the prediction target, and iteratively construct a decision tree based on the pseudo residual to optimize the initial computing power hotspot distribution prediction model until the preset iteration termination condition is met, and obtain the preliminary computing power hotspot distribution prediction model.

[0099] Specifically, the terminal employs a gradient boosting decision tree algorithm to construct an initial prediction model for computing power hotspot distribution. It calculates the pseudo-residuals between the input features of each sample in the model training set and the prediction target, iteratively constructing new decision trees. Each decision tree specifically learns residual patterns that the preceding model failed to capture, fitting the most relevant patterns from the input features using feature splitting rules. In each iteration, the newly constructed decision tree is added to the ensemble model, and its output is combined with the existing model with appropriate weights to gradually reduce the overall prediction error until a preset iteration termination condition is met, resulting in the initial prediction model for computing power hotspot distribution. Gradient boosting decision tree is an ensemble learning method that gradually reduces prediction error by sequentially constructing a sequence of decision trees. Each new decision tree learns the negative gradient direction of the residuals of the preceding model, i.e., the pseudo-residuals. For example, the preset iteration termination condition is typically set to reach the maximum number of trees or when the performance on the validation set no longer significantly improves, ensuring that the model learns sufficiently while avoiding overfitting.

[0100] S3034, the preliminary computing power hotspot distribution prediction model is tested through the model validation set until the preliminary computing power hotspot distribution prediction model meets the preset accuracy requirements, and the trained computing power hotspot distribution prediction model is obtained.

[0101] Specifically, the terminal uses a model validation set to test the preliminary computing power hotspot distribution prediction model. When the preliminary computing power hotspot distribution prediction model meets preset conditions, the trained preliminary computing power hotspot distribution prediction model is tested to input the spatiotemporal cross feature vector, time-series derived feature vector, startup feature vector, and real-time status data of the current task, and output a computing power hotspot distribution map of the current task. Illustratively, the preset accuracy requirement is an accuracy rate of 85%, but this can be adjusted according to actual work requirements.

[0102] This embodiment provides a software-based virtualized allocation method for computing resources. By extracting the spatiotemporal cross-features of the startup feature vector and real-time state data, and the temporal derived features of the startup feature vector, spatiotemporal cross-feature vectors and temporal derived feature vectors are obtained. Furthermore, an ensemble learning algorithm is used to mine resource usage patterns from historical data, generating an accurate computing power hotspot distribution map for startup tasks. This achieves precise capture of resource demand patterns during task startup, overcoming the blindness of traditional experience-based allocation. It provides reliable input data for subsequent resource competition analysis and pre-allocation decisions, improving the accuracy and reliability of hotspot distribution prediction, thereby enhancing response speed and resource utilization efficiency in high-concurrency task startup scenarios.

[0103] In one embodiment, the loss function of the computing power hotspot distribution prediction model is:

[0104]

[0105] in, It is the model loss. It is the total number of historical startup data. Any historical task , It is a historical mission The result of the loss It is the total number of decision trees. Any decision tree The total number of leaf nodes, It is a decision tree The output vector, It is the quantity penalty coefficient. It is the weighted penalty coefficient;

[0106] The first loss term is obtained using the following formula:

[0107]

[0108] in, It is a historical mission The result of the loss It is a computing power hotspot distribution prediction model for historical tasks The predicted value, It is a historical mission The distribution map of computing power hotspots It is the threshold for result loss.

[0109] Specifically, the terminal calculates the loss predicted by the model according to a formula, and adjusts the model parameters based on this loss to improve the accuracy and generalization ability of the prediction model for hotspot distribution of computing power. For example, the total number of historical startup data... The number of historical tasks included in the preset historical launch count can be determined. Optionally, historical tasks... The result of the loss Used to measure the model's performance on historical tasks Predicted computing power hotspot distribution map and historical tasks The differences in the actual distribution of computing power hotspots, among which It is a computing power hotspot distribution prediction model for historical tasks The predicted value is the model's prediction of the historical task input. Obtained through processing; It is a historical mission The distribution map of computing power hotspots can be obtained based on historical startup data; result loss threshold Used to measure the convergence speed of prediction error, when the absolute value of the prediction error is less than or equal to the result loss threshold. Squared loss is used to improve convergence efficiency when the error exceeds the result loss threshold. Linear loss is used to reduce the impact of outliers, resulting in a loss threshold. The model's prediction accuracy can be set according to actual work requirements. Optionally, the total number of decision trees... This can be obtained by traversing a trained computing power hotspot distribution prediction model. For example, any decision tree... Total number of leaf nodes The decision tree is determined based on the pre-defined iteration termination conditions for each decision tree during model construction. Optionally, the decision tree... The output vector It is a decision tree It consists of the predicted values ​​of all leaf nodes. For example, the quantity penalty coefficient... Used to control the growth rate of the tree, weighted penalty coefficient The quantity penalty coefficient is used to adjust the fluctuation range of the output vector. and weighted penalty coefficient It can be configured according to the actual work requirements.

[0110] This embodiment provides a software-based virtualization allocation method for computing resources. By designing a robust composite loss function, it provides a comprehensive and balanced optimization objective for training the computing hotspot distribution prediction model. It utilizes Huber loss to reduce the interference of abnormal samples on the training process and effectively controls the model complexity through regularization terms. This enables the computing hotspot distribution prediction model to avoid overfitting while maintaining high accuracy, achieving stable and reliable prediction of resource hotspot distribution during task startup. This improves the response speed and resource utilization efficiency when dealing with high-concurrency task startup scenarios.

[0111] In one embodiment, based on the computing power hotspot distribution map, task description vector, and real-time status data, the competition situation at task startup is simulated to obtain quantitative data on resource competition intensity, including:

[0112] S501 adjusts real-time state data based on preset simulation scenario coefficients to obtain scenario state data.

[0113] Specifically, the terminal adjusts real-time status data tagged with resources based on preset simulated scenario coefficients to generate scenario status data reflecting different computing resources under different load conditions. The preset simulated scenario coefficients are derived from historical load statistics and predictive uncertainty settings, and include optimistic, baseline, and pessimistic coefficients, corresponding to light, normal, and heavy load states, respectively. The terminal multiplies the CPU utilization, available memory capacity, and cache status parameters at each level from the real-time status data with these coefficients to obtain the scenario status data for each simulated scenario coefficient.

[0114] S502, for each simulated scenario coefficient, based on the computing power hotspot distribution map and scenario status data, calculates the competition intensity value of the simulated scenario coefficient; the competition intensity value is used to characterize the computing resource usage intensity corresponding to the start of the task.

[0115] Specifically, for each resource tag in each simulated scenario coefficient, the terminal compares the expected usage intensity of each computing resource indicator value corresponding to each time node in the computing power hotspot distribution map with the scenario status data item by item. It then quantifies the resource demand difference based on the comparison of each resource indicator parameter at each time node, and integrates these resource demand differences into a competition intensity value for that resource tag in that simulated scenario coefficient. For example, the terminal can employ a resource competition index algorithm. This algorithm generates a standardized competition intensity between 0 and 1 by weighted aggregation of the relative oversupply of each resource type, accurately reflecting the overall resource pressure level when starting a task in a specific scenario. Optionally, the computing resource indicator values ​​can be CPU utilization, memory capacity, and cache status parameters at various levels.

[0116] S503, based on the competition intensity values ​​of each simulated scenario coefficient, detects the local extrema of the competition intensity values ​​and obtains the local extrema data of each competition intensity value.

[0117] Specifically, the extreme value detection algorithm in the terminal application signal processing scans the curves of the competition intensity values ​​of each simulated scene coefficient under each resource label over time, identifies local maxima points where the values ​​are higher than or equal to those of neighboring time nodes, and records the time position, peak height, and peak width of these extreme points to obtain local extreme value data for each competition intensity value. This local extreme value data characterizes the critical time nodes and competition situations where computational resource competition is most intense during task initiation. For example, the terminal can employ a peak detection algorithm based on a sliding window. This algorithm identifies extreme values ​​by comparing the numerical relationship between the current point and points within the preceding and following windows, effectively filtering noise interference and capturing meaningful competition burst points.

[0118] S504: Sort the competition intensity values ​​of each simulated scenario coefficient according to the local extreme value data of each competition intensity value, and combine the local extreme value data corresponding to the competition intensity values ​​to obtain the quantitative data of resource competition intensity.

[0119] Specifically, the terminal sorts the competition intensity values ​​of each simulated scenario coefficient under each resource tag according to the size of the local extreme value data, and integrates them with the local extreme value data of each simulated scenario coefficient to form resource competition intensity quantification data. Among them, the resource competition intensity quantification data includes the competition intensity ranking and competition extreme value statistical characteristics of computing resources under each resource tag under each simulated scenario coefficient.

[0120] This embodiment provides a software-based virtualization allocation method for computing resources. It simulates the load states of different computing resources using different scenario coefficients, obtaining scenario state data for different computing resources under different simulated scenario coefficients. The method then quantifies and evaluates the resource contention intensity caused by task startup in each scenario, obtaining a contention intensity value. Furthermore, it detects the contention intensity value, identifying the critical moments and extent of contention outbreaks, obtaining local extreme value data for the contention intensity value. Finally, it integrates the local extreme value data of all computing resources under all simulated scenario parameters with the contention intensity value to obtain quantified contention intensity data. This method enables a proactive assessment of the resource impact of task startup, identifying potential resource bottlenecks and contention risks, thereby improving response speed and resource utilization efficiency in high-concurrency task startup scenarios.

[0121] In one embodiment, a pre-allocation scheme for tasks is generated based on resource contention intensity quantification data and real-time status data, including:

[0122] S601 constructs allocation constraints between tasks and computing resources based on quantified data of resource contention intensity, preset startup time thresholds, and preset computing resource capacity thresholds.

[0123] Specifically, the terminal constructs allocation constraint formulas between tasks and computing resources based on quantified resource contention intensity data, preset startup time thresholds, and preset computing resource capacity thresholds, thus obtaining the allocation constraints between tasks and each computing resource. The quantified resource contention intensity data provides a contention pressure index for each resource unit, used to identify highly contentious computing resources that need to be avoided; the startup time threshold ensures that the task completes the startup process within a specified time range; the computing resource capacity threshold limits the maximum allocatable amount for each resource unit. The allocation constraints include, but are not limited to, time constraints, capacity constraints, and avoidance constraints. Time constraints indicate that the task startup completion time does not exceed the preset startup time threshold; capacity constraints indicate that the allocated resources do not exceed the preset computing resource capacity threshold; and avoidance constraints indicate that computing resources with excessively high contention intensity values ​​should be avoided. For example, the preset startup time threshold can be set according to the actual requirements for task startup time in work; the preset computing resource capacity threshold can be set according to the requirements for available computing resource space at task startup in work.

[0124] S602 uses tasks and computing resources as graph nodes, and fills in the connection edges between graph nodes based on the computing power hotspot distribution map and real-time status data to obtain a complete allocation bipartite graph.

[0125] Specifically, the terminal constructs a complete bipartite graph of task and computing resource allocation. In this graph, the left set of nodes represents tasks to be allocated, and the right set of nodes represents available computing resources. The graph nodes can be partitioned based on resource labels, and the edges connecting the left and right sets of nodes represent possible allocation relationships. Based on the computing power hotspot distribution map and real-time status data, each edge is assigned a weight value. The weight value is a weighted sum of competition avoidance benefits and performance benefits. The computing power hotspot distribution map provides information on task demand preferences for each resource, and the real-time status data reflects the current availability of each resource. For example, demand-capacity matching degree can be used as the basis for edge weights, adjusted using competition intensity data to ensure that the allocation weight of highly competitive resources is correspondingly reduced. For example, edge weights can be calculated based on the computing power hotspot distribution map and real-time status data. The computing power hotspot distribution map extracts the consumption of each computing resource by tasks, and combined with the available computing resource capacity in the real-time status data, the basic performance benefits are calculated. Competition avoidance benefits are quantified based on the competition intensity values ​​of each computing resource in the resource competition intensity quantification data. The weighted combination of the basic performance benefits and competition avoidance benefits yields the weight value of the edge connecting two nodes.

[0126] S603. Based on the allocation constraints, solve the matching relationship between tasks and computing resources to obtain candidate pre-allocation schemes; each candidate pre-allocation scheme forms a candidate pre-allocation scheme set.

[0127] Specifically, the terminal uses the Hungarian algorithm to find the task-resource matching relationship that maximizes the overall allocation efficiency while satisfying the allocation constraints, resulting in several candidate pre-allocation schemes. These candidate pre-allocation schemes form a candidate pre-allocation scheme set. The Hungarian algorithm is a polynomial-time algorithm that solves the maximum weight matching problem in a bipartite graph through matrix transformations and augmented path finding. It progressively optimizes the matching scheme through steps such as row reduction, column reduction, and covering labels, ultimately finding the optimal allocation correspondence.

[0128] S604 selects a candidate pre-allocation scheme from the candidate pre-allocation scheme set based on the completion time limit parameters, computing power requirement parameters, and real-time status data to obtain the task pre-allocation scheme.

[0129] Specifically, the terminal comprehensively evaluates the task based on completion time limits, computing power requirements, and real-time status data, combined with preset optimization objectives. It then selects the candidate pre-allocation scheme with the highest comprehensive evaluation as the pre-allocation scheme for the task. The completion time limit determines the time feasibility requirement of the scheme, the computing power requirement ensures that allocated resources meet the task execution needs, and the real-time status data provides the real-world environmental constraints for scheme execution. For example, the TOPSIS multi-attribute decision-making method can be used to rank the schemes by calculating their relative closeness to the ideal solution, selecting the scheme with the best overall performance as the final pre-allocation scheme. For example, the preset optimization objectives include time satisfaction, resource suitability, and contention avoidance. Time satisfaction can be obtained by calculating the minimum startup time target value using each computing resource based on the completion time limit parameter; resource suitability can be obtained by calculating the variance of the resource utilization equilibrium target value based on the computing power requirement parameter and real-time status data; and contention avoidance can be obtained by calculating the contention avoidance target value through quantitative data on resource contention intensity.

[0130] This embodiment provides a software-based virtualized allocation method for computing resources. By clearly defining various constraints on resource allocation, allocation constraints are obtained. The complex resource allocation problem is transformed into a computable graph matching model, resulting in a complete allocation bipartite graph. An optimization algorithm is used to solve for high-quality candidate solutions, and the most suitable pre-allocation scheme is selected based on task characteristics and computing resource status. This effectively balances task requirements and computing resource conditions, optimizing resource allocation while satisfying various constraints. It significantly improves resource utilization efficiency while reducing resource contention risks, thereby enhancing response speed and resource utilization efficiency in high-concurrency task startup scenarios.

[0131] In one embodiment, the method further includes:

[0132] S701, Obtain execution feedback data of the pre-allocation scheme; the execution feedback data is used to characterize the changes in the real-time status data of computing resources when the pre-allocation scheme is executed.

[0133] Specifically, the terminal obtains execution feedback data of the pre-allocation scheme. This execution feedback data includes, but is not limited to, time series data of actual CPU core utilization, memory bandwidth usage fluctuations, prediction accuracy data of the pre-allocation scheme, and cache hit rate change rate prediction accuracy data of the pre-allocation scheme.

[0134] S702, based on execution feedback data, computing power hotspot distribution map and resource competition intensity quantification data, quantifies the prediction accuracy of computing resource usage when the task starts, and obtains the prediction accuracy data of the pre-allocation scheme.

[0135] Specifically, the terminal performs spatiotemporal alignment comparisons based on the number of execution feedbacks, the distribution map of computing power hotspots, and the quantitative data of resource contention intensity. The terminal uses root mean square error to measure the overall deviation between the predicted and actual values ​​of CPU utilization, available memory capacity, and cache status parameters at all levels, and uses correlation coefficients to evaluate the consistency of the changing trends, thereby obtaining the prediction accuracy data of the pre-allocation scheme.

[0136] S703 updates the historical performance database based on prediction accuracy data, task description vectors, and execution feedback data, resulting in an updated historical performance database.

[0137] Specifically, the terminal adds the task description vector, prediction feature data, actual execution feedback data, and accuracy evaluation results of the current task to the historical database as a new data sample. Simultaneously, the weights of existing historical records are adjusted based on prediction accuracy data, with historical samples that accurately predict the data receiving higher weights, while older samples with weaker correlations are gradually de-weighted, resulting in an updated historical performance database.

[0138] This embodiment provides a software-based virtualized allocation method for computing resources. By acquiring execution feedback data from a pre-allocation scheme, and combining this with a computing power hotspot distribution map and resource contention intensity quantification data, the accuracy of the pre-allocation scheme is quantified to obtain prediction accuracy data. Based on the evaluation results, the quality and representativeness of the historical performance database are dynamically optimized. The prediction accuracy data, task description vectors, and execution feedback data are then added to the historical performance database to obtain an updated historical performance database. This method achieves gradual adaptation to changes in computing resource load and task characteristics, continuously improving prediction accuracy and optimizing resource allocation strategies, thereby enhancing response speed and resource utilization efficiency when dealing with high-concurrency task initiation scenarios.

[0139] The aforementioned software-based virtualization allocation method for computing resources for workloads involves acquiring the task description vector and real-time status data of computing resources. The task description vector includes task type parameters, completion time limit parameters, and computing power requirement parameters. Based on the task description vector and a pre-set historical performance database, the computing resources required to start the task are analyzed to obtain a startup feature vector. The historical performance database includes task type parameters and computing resource consumption time series of historical tasks. Based on the startup feature vector and real-time status data, a computing power hotspot distribution map of the startup task is calculated. This map characterizes the usage of each computing resource at different time points during task startup. Based on the computing power hotspot distribution map, task description vector, and real-time status data, the competition situation during task startup is simulated to obtain quantified data on resource competition intensity. Based on the quantified data on resource competition intensity and real-time status data, a pre-allocation scheme for the task is generated. This pre-allocation scheme characterizes the allocation relationship between the task and each computing resource. This method effectively solves the core problems of traditional resource allocation methods, such as long startup delays and intense resource competition due to a lack of forward-looking perspective. It significantly improves the response speed and resource utilization efficiency when dealing with high-concurrency task startup scenarios, while also enhancing the determinism and reliability of task execution under complex load environments.

[0140] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0141] Based on the same inventive concept, this application also provides a software-based payload computing resource virtualization allocation device for implementing the software-based payload computing resource virtualization allocation method described above. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more software-based payload computing resource virtualization allocation device embodiments provided below can be found in the limitations of the software-based payload computing resource virtualization allocation method described above, and will not be repeated here.

[0142] In one exemplary embodiment, such as Figure 2 As shown, a software-based payload computing resource virtualization allocation device 200 is provided, comprising:

[0143] The data acquisition module 201 is used to acquire the task description vector of the task and the real-time status data of the computing resources; the task description vector includes the task type parameters, completion time limit parameters and computing power requirement parameters of the task.

[0144] The consumption fitting module 202 is used to analyze the computing resources required to start a task based on the task description vector and a preset historical performance database, and obtain the startup feature vector; the historical performance database includes the task type parameters and computing resource consumption time series of historical tasks.

[0145] The resource hotspot calculation module 203 is used to calculate the computing power hotspot distribution map of the startup task based on the startup feature vector and real-time status data; the computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points when the startup task is started.

[0146] The competition calculation module 204 is used to simulate the competition situation when a task starts based on the computing power hotspot distribution map, task description vector and real-time status data, and obtain quantitative data of resource competition intensity.

[0147] The scheme generation module 205 is used to generate a pre-allocation scheme for tasks based on the resource competition intensity quantification data and real-time status data; the pre-allocation scheme is used to characterize the allocation relationship between tasks and each computing resource.

[0148] Furthermore, the consumption fitting module is also used for:

[0149] Select historical tasks from the historical performance database that have the same task type parameters as the task to form a reference task set;

[0150] For each historical task in the reference task set, the computational resource consumption time series of the historical task is time-aligned to obtain the time-aligned computational resource consumption sequence of each historical task.

[0151] For each historical task in the reference task set, the consumption of computing resources when starting the historical task is quantified based on the time-aligned computing resource consumption sequence, and the start-up consumption feature vector of each historical task is obtained.

[0152] The startup feature vector is obtained based on the startup consumption feature vector of each historical task in the reference task set.

[0153] Furthermore, the resource hotspot calculation module is also used for:

[0154] Based on the initiation feature vector and real-time state data, the spatiotemporal cross feature vector is constructed using the following formula:

[0155]

[0156] in, It is a spatiotemporal cross feature vector. It is the initiation feature vector. It is real-time status data. It is the element-wise multiplication symbol;

[0157] Based on the initiation feature vector, the rate of change of computing resources is calculated to obtain the time-series derived feature vector;

[0158] The spatiotemporal cross feature vector, time-series derived feature vector, startup feature vector, and real-time status data are input into the trained computing power hotspot distribution prediction model to obtain the computing power hotspot distribution map of the startup task.

[0159] The computing power hotspot distribution prediction model was obtained through the following method:

[0160] Obtain preset historical startup data, which includes startup feature vectors of historical tasks, real-time status data, and corresponding computing power hotspot distribution maps;

[0161] Using the historical task startup feature vector and real-time status data as input features, and the corresponding computing power hotspot distribution map as the prediction target, the historical startup data is divided into a model training set and a model validation set according to a preset ratio.

[0162] An initial computing power hotspot distribution prediction model is constructed. The pseudo residuals between the input features in the model training set and the prediction target are calculated. A decision tree is iteratively constructed based on the pseudo residuals to optimize the initial computing power hotspot distribution prediction model until the preset iteration termination condition is met, thus obtaining the preliminary computing power hotspot distribution prediction model.

[0163] The preliminary computing power hotspot distribution prediction model is tested using a model validation set until it meets the preset accuracy requirements, thus obtaining a trained computing power hotspot distribution prediction model.

[0164] Furthermore, the loss function of the computing power hotspot distribution prediction model is:

[0165]

[0166] in, It is the model loss. It is the total number of historical startup data. Any historical task , It is a historical mission The result of the loss It is the total number of decision trees. Any decision tree The total number of leaf nodes, It is a decision tree The output vector, It is the quantity penalty coefficient. It is the weighted penalty coefficient;

[0167] The first loss term is obtained using the following formula:

[0168]

[0169] in, It is a historical mission The result of the loss It is a computing power hotspot distribution prediction model for historical tasks The predicted value, It is a historical mission The distribution map of computing power hotspots It is the threshold for result loss.

[0170] Furthermore, the contention calculation module is also used for:

[0171] Based on the computing power hotspot distribution map, task description vectors, and real-time status data, the competition situation at task startup is simulated to obtain quantitative data on resource competition intensity, including:

[0172] Based on preset simulation scenario coefficients, adjust real-time status data to obtain scenario status data;

[0173] For each simulated scenario coefficient, based on the computing power hotspot distribution map and the scenario status data, the competition intensity value of the simulated scenario coefficient is calculated; the competition intensity value is used to characterize the computing resource usage intensity corresponding to the start of the task.

[0174] Based on the competition intensity values ​​of each simulated scenario coefficient, the local extrema of the competition intensity values ​​are detected to obtain the local extrema data of each competition intensity value;

[0175] The competition intensity values ​​of each simulated scenario coefficient are sorted based on the local extreme value data of each competition intensity value. Combined with the local extreme value data corresponding to the competition intensity values, the quantitative data of resource competition intensity is obtained.

[0176] Furthermore, the solution generation module is also used for:

[0177] Based on the quantitative data of resource competition intensity, the preset start time threshold, and the preset computing resource capacity threshold, the allocation constraints between tasks and each computing resource are constructed.

[0178] Using tasks and computing resources as graph nodes, and based on the computing power hotspot distribution map and real-time status data, fill the connection edges between graph nodes to obtain a complete allocation bipartite graph;

[0179] Based on the allocation constraints, the matching relationship between tasks and computing resources is solved to obtain candidate pre-allocation schemes; the candidate pre-allocation schemes form a candidate pre-allocation scheme set.

[0180] Based on the completion time limit parameters, computing power requirement parameters, and real-time status data, a candidate pre-allocation scheme is selected from the candidate pre-allocation scheme set to obtain the task pre-allocation scheme.

[0181] Furthermore, the device also includes a feedback module, which is used for:

[0182] Obtain execution feedback data for the pre-allocation scheme; the execution feedback data is used to characterize the changes in the real-time status data of computing resources when the pre-allocation scheme is executed;

[0183] Based on execution feedback data, computing power hotspot distribution map and resource competition intensity quantification data, the prediction accuracy of computing resource usage at task startup is quantified to obtain prediction accuracy data of pre-allocation scheme;

[0184] Based on prediction accuracy data, task description vectors, and execution feedback data, the historical performance database is updated to obtain the updated historical performance database.

[0185] In one embodiment, such as Figure 3 A computer device is provided, comprising:

[0186] At least one processor 301, and a memory 302 communicatively connected to at least one of the processors 301: the memory stores application code executable by at least one of the processors, the application code being executed by at least one of the processors to enable at least one of the processors to perform the software-based load computing resource virtualization allocation method as described above.

[0187] Computer equipment may also include: sensor 303.

[0188] The processor 301, memory 302 and sensor 303 can be connected via a bus or other means, with the bus being an example in the figure.

[0189] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0190] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0191] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A software-based virtualization allocation method for computing resources, characterized in that, The method includes: Obtain the task description vector and real-time status data of computing resources for the task; the task description vector includes the task type parameter, completion time limit parameter, and computing power requirement parameter for the task. Based on the task description vector and a preset historical performance database, the computing resources required to start the task are analyzed to obtain a startup feature vector; the historical performance database includes the task type parameters and computing resource consumption time series of historical tasks. Based on the startup feature vector and the real-time status data, a computing power hotspot distribution map is calculated to start the task; the computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points when the task is started; Based on the computing power hotspot distribution map, the task description vector, and the real-time status data, the competition situation when the task starts is simulated to obtain quantitative data on resource competition intensity. Based on the quantified data of resource competition intensity and the real-time status data, a pre-allocation scheme for the task is generated; the pre-allocation scheme is used to characterize the allocation relationship between the task and each of the computing resources.

2. The method according to claim 1, characterized in that, The step involves analyzing the computing resources required to start the task based on the task description vector and a preset historical performance database to obtain a startup feature vector, including: Select historical tasks from the historical performance database that have the same task type parameter as the task to form a reference task set; For each historical task in the reference task set, the computational resource consumption time series of the historical task is time-aligned to obtain the time-aligned computational resource consumption sequence of each historical task. For each historical task in the reference task set, the consumption of computing resources when starting the historical task is quantified based on the time-aligned computing resource consumption sequence to obtain the startup consumption feature vector of each historical task. The startup feature vector is obtained based on the startup consumption feature vector of each historical task in the reference task set.

3. The method according to claim 1, characterized in that, The step of calculating the computing power hotspot distribution map for launching the task based on the launch feature vector and the real-time status data includes: Based on the startup feature vector and the real-time state data, a spatiotemporal cross feature vector is constructed using the following formula: in, It is a spatiotemporal cross feature vector. It is the initiation feature vector. It is real-time status data. It is the element-wise multiplication symbol; Based on the startup feature vector, the rate of change of the computing resources is calculated to obtain the time-series derived feature vector; The spatiotemporal cross feature vector, the time-series derived feature vector, the startup feature vector, and the real-time status data are input into the trained computing power hotspot distribution prediction model to obtain the computing power hotspot distribution map for starting the task. The computing power hotspot distribution prediction model is obtained through the following method: Obtain preset historical startup data, which includes the startup feature vector of the historical task, the real-time status data, and the corresponding computing power hotspot distribution map; Using the startup feature vector and real-time status data of the historical tasks as input features, and the corresponding computing power hotspot distribution map as the prediction target, the historical startup data is divided into a model training set and a model validation set according to a preset ratio. An initial computing power hotspot distribution prediction model is constructed. The pseudo residual between the input features and the prediction target in the model training set is calculated. A decision tree is iteratively constructed based on the pseudo residual to optimize the initial computing power hotspot distribution prediction model until the preset iteration termination condition is met, thus obtaining the preliminary computing power hotspot distribution prediction model. The preliminary computing power hotspot distribution prediction model is tested using the model validation set until it meets the preset accuracy requirements, thus obtaining the trained computing power hotspot distribution prediction model.

4. The method according to claim 3, characterized in that, The loss function of the computing power hotspot distribution prediction model is: in, It is the model loss. It is the total number of historical startup data. Any historical task , It is a historical mission The result of the loss It is the total number of decision trees. Any decision tree The total number of leaf nodes, It is a decision tree The output vector, It is the quantity penalty coefficient. It is the weighted penalty coefficient; The first loss term is obtained using the following formula: in, It is a historical mission The result of the loss It is a computing power hotspot distribution prediction model for historical tasks The predicted value, It is a historical mission The distribution map of computing power hotspots It is the threshold for result loss.

5. The method according to claim 1, characterized in that, Based on the computing power hotspot distribution map, the task description vector, and the real-time status data, the competition situation at the start of the task is simulated to obtain quantitative data on resource competition intensity, including: Based on preset simulation scenario coefficients, the real-time state data is adjusted to obtain scenario state data; For each simulated scenario coefficient, based on the computing power hotspot distribution map and the scenario state data, the competition intensity value of the simulated scenario coefficient is calculated; the competition intensity value is used to characterize the computing resource usage intensity corresponding to the start of the task. Based on the competition intensity value of each of the simulated scenario coefficients, the local extrema of the competition intensity value are detected to obtain the local extrema data of each of the competition intensity values; The competition intensity values ​​of each simulated scenario coefficient are sorted according to the local extreme value data of each competition intensity value, and the resource competition intensity quantification data is obtained by combining the local extreme value data corresponding to the competition intensity values.

6. The method according to claim 1, characterized in that, The step of generating a pre-allocation scheme for the task based on the quantified resource competition intensity data and the real-time status data includes: Based on the quantified data of resource competition intensity, the preset start time threshold, and the preset computing resource capacity threshold, the allocation constraints between the task and each computing resource are constructed. Using the task and the computing resources as graph nodes, and based on the computing power hotspot distribution map and the real-time status data, fill the connection edges between the graph nodes to obtain a complete allocation bipartite graph; Based on the allocation constraints, the matching relationship between the task and the computing resources is solved to obtain candidate pre-allocation schemes; each candidate pre-allocation scheme constitutes a candidate pre-allocation scheme set. Based on the completion time limit parameter, computing power requirement parameter and real-time status data, the candidate pre-allocation scheme is selected from the candidate pre-allocation scheme set to obtain the pre-allocation scheme for the task.

7. The method according to claim 1, characterized in that, The method further includes: Obtain execution feedback data of the pre-allocation scheme; the execution feedback data is used to characterize the changes in the real-time status data of the computing resources when the pre-allocation scheme is executed; Based on the execution feedback data, the computing power hotspot distribution map, and the resource competition intensity quantification data, the prediction accuracy of the computing resource usage when the task starts is quantified to obtain the prediction accuracy data of the pre-allocation scheme. Based on the prediction accuracy data, the task description vector, and the execution feedback data, the historical performance database is updated to obtain the updated historical performance database.

8. A software-based virtualization allocation device for computing resources, characterized in that, The device includes: The data acquisition module is used to acquire the task description vector of the task and the real-time status data of the computing resources; the task description vector includes the task type parameter, completion time limit parameter and computing power requirement parameter of the task. The consumption fitting module is used to analyze the computing resources required to start the task based on the task description vector and a preset historical performance database, and obtain a startup feature vector; the historical performance database includes the task type parameters and computing resource consumption time series of historical tasks; The resource hotspot calculation module is used to calculate the computing power hotspot distribution map of the task startup based on the startup feature vector and the real-time status data; the computing power hotspot distribution map is used to characterize the usage of each computing resource at different time points when the task is started; The competition calculation module is used to simulate the competition situation when the task starts based on the computing power hotspot distribution map, the task description vector and the real-time status data, and obtain resource competition intensity quantification data. The scheme generation module is used to generate a pre-allocation scheme for the task based on the resource competition intensity quantification data and the real-time status data; the pre-allocation scheme is used to characterize the allocation relationship between the task and each of the computing resources.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images