An AI computing power resource dynamic allocation method and system

Abstract

The application relates to the technical field of computing resource allocation, and discloses an AI computing resource dynamic allocation method and system, which comprises the following steps: analyzing resource states according to pre-acquired computing resource state data, identifying discrete distributed fragment resources, calculating corresponding resource values, and constructing a fragment resource atlas; in response to a preset computing task, configuring a fragment combination mechanism to obtain fragment resource blocks; configuring a resource screening mechanism to dynamically screen computing resources that can meet the computing task after being combined with the fragment resource blocks, and obtaining a resource allocation scheme; and scheduling the computing resources based on the resource allocation scheme; by effectively utilizing the fragment resources, the application can improve the overall resource utilization rate of the computing resources, improve the task response speed, actively combine the fragment resources to complete corresponding computing tasks, reduce task queuing and execution time, and thus improve the dynamic scheduling efficiency and task execution efficiency of the computing resources.

Description

Technical Field

[0001] This application relates to the field of computing power resource allocation technology, and more specifically to a method and system for dynamic allocation of AI computing power resources. Background Technology

[0002] Currently, AI technology is developing rapidly and is widely used in fields such as large model training and deep learning, leading to a significant increase in the demand for large-scale, high-efficiency computing power. Existing computing resource management and allocation methods mostly adopt static or semi-dynamic strategies, which distribute computing tasks as a whole across one or several pre-defined, complete physical or virtual nodes to pursue centralized resource allocation and ease of management. However, these methods cannot meet the needs of highly dynamic, heterogeneous, and variable-load AI computing environments.

[0003] Existing technologies suffer from the following problems: scheduling computing resources based on the overall resource occupancy rate lacks awareness and allocation of fragmented resources distributed discretely in the spatiotemporal dimension, ignoring fragmented resources and resulting in their waste; passively waiting for sufficiently large idle resources to appear leads to long task waiting times, low resource utilization efficiency, and an inability to respond to computing tasks in real time; the use of unidirectional and static resource matching methods makes it impossible to perform bidirectional matching based on real-time task requirements and dynamically changing resource states, making it impossible to predict the actual effect of resource combinations and resulting in low computing resource allocation efficiency; to solve at least one of the above problems, this application proposes a dynamic allocation method and system for AI computing resources. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this application is to provide a method and system for dynamic allocation of AI computing resources, which can effectively solve the problems in the background technology. The specific technical solution of this application is as follows:

[0005] A method for dynamically allocating AI computing resources includes:

[0006] Based on the pre-acquired computing power resource status data, analyze the resource status, identify discretely distributed fragmented resources through a preset status analysis model, calculate the corresponding resource value, and construct a fragmented resource map;

[0007] In response to a preset computing task, a fragment combination mechanism is configured to select fragment resources from the fragment resource map whose fragment resource recombination benefits are greater than a preset benefit threshold, and combine them to obtain fragment resource blocks.

[0008] Configure a resource filtering mechanism to dynamically filter out computing resources that can meet the computing task after being combined with the fragmented resource blocks, and obtain a resource allocation scheme;

[0009] The computing resources are scheduled based on the resource allocation scheme to dynamically allocate AI computing resources.

[0010] Specifically, the step of analyzing resource status based on pre-acquired computing power resource status data, identifying discretely distributed fragmented resources through a preset status analysis model, calculating the corresponding resource value, and constructing a fragmented resource map includes:

[0011] Based on the pre-acquired computing resource status data, the usage status of the resources is analyzed, and discretely distributed fragmented resources are identified through a preset status analysis model to obtain a set of fragmented resources.

[0012] Calculate the resource value of each fragment resource in the fragment resource set, treat the fragment resources as resource layer entities, treat the calculation tasks as task layer entities, establish connections between the corresponding entities according to resource value, and construct a fragment resource graph.

[0013] Specifically, based on the pre-acquired computing resource status data, the usage status of the resources is analyzed, and discretely distributed fragmented resources are identified through a preset status analysis model to obtain a set of fragmented resources, including:

[0014] Based on the pre-acquired computing resource status data, the resource consumption of running tasks is analyzed through a preset resource status prediction model to predict resource idle time.

[0015] Extract the static and dynamic load characteristics of each computing resource to construct a resource feature set;

[0016] By combining the resource idle time and resource feature set, a set of fragmented resources discretely distributed within each preset time interval is identified through a preset state analysis model, thus obtaining a set of fragmented resources.

[0017] Specifically, the step of calculating the resource value of each fragment resource in the fragment resource set, treating the fragment resources as resource layer entities and the calculation tasks as task layer entities, establishing connections between corresponding entities according to resource value, and constructing a fragment resource graph includes:

[0018] For each fragment resource in the fragment resource set, analyze and combine static and dynamic attributes to calculate the corresponding resource value;

[0019] Each fragment resource is treated as a resource layer entity, and each computing task is treated as a task layer entity.

[0020] Based on the resource value, connections are established between the task-level entities and the resource-level entities required by the corresponding computing tasks to obtain a fragmented resource map.

[0021] Specifically, in response to a preset computing task, a fragment combination mechanism is configured to select fragment resources from the fragment resource map whose fragment resource recombination benefit is greater than a preset benefit threshold, and combine them to obtain fragment resource blocks, including:

[0022] In response to a preset computational task, task features are extracted using a preset task parsing model, and a task feature vector is constructed.

[0023] The fragment resources in the fragment resource map are encoded by a preset resource encoder to obtain fragment resource codes;

[0024] By combining the task feature vector and fragment resource encoding, fragment resources whose fragment resource recombination benefits are greater than a preset benefit threshold are selected and combined to obtain fragment resource blocks.

[0025] Specifically, by combining the task feature vector and fragment resource encoding, fragment resources whose fragment resource recombination benefits are greater than a preset benefit threshold are selected and combined to obtain fragment resource blocks, including:

[0026] Instantiate the corresponding task agent based on the task feature vector, and instantiate multiple corresponding resource agents based on the fragment resource encoding;

[0027] Multiple resource agents simulate a competition process. The task agent selects at least one resource agent based on the competition process, resulting in a set of resource agents. The resource agents compete based on the matching degree between the resource code and the task feature vector.

[0028] Based on the similarity between fragmented resources, calculate the recombination benefit between fragmented resources in the resource agent set;

[0029] Fragmented resources with recombination benefits greater than a preset benefit threshold are selected from the set of resource agents and combined to obtain fragmented resource blocks.

[0030] Specifically, the configuration resource filtering mechanism dynamically filters out computing resources that, when combined with the fragmented resource blocks, can satisfy the computing task, resulting in a resource allocation scheme, including:

[0031] Analyze the computing resource requirements of computing tasks, perform resource status analysis on non-fragmented resources, and generate corresponding resource status results;

[0032] Based on the resource status results, the combination process of non-fragmented resources and fragmented resource blocks is simulated through a preset sand table simulation model, and a set of non-fragmented resources that can meet the computing task after combination is selected.

[0033] By combining the non-fragmented resource set and the fragmented resource blocks, a resource allocation scheme is obtained.

[0034] Specifically, based on the resource status results, a pre-set sandbox simulation model is used to simulate the combination process of non-fragmented resources and fragmented resource blocks, and a set of non-fragmented resources that can satisfy the computing task after combination is selected, including:

[0035] Based on the resource status results, the process of simulating the logical combination of each non-fragmented resource and fragmented resource block through a preset sand table simulation model, and running the corresponding computing tasks, is used to obtain simulation results.

[0036] The performance metrics of the simulation results are analyzed, and the result performance metrics are constructed.

[0037] Select the set of non-fragmented resources that can meet the computing task and achieve the best performance indicators after combination.

[0038] Specifically, based on the resource allocation scheme, computing resources are scheduled to dynamically allocate AI computing resources, including:

[0039] Based on the resource allocation scheme, generate corresponding resource scheduling instructions;

[0040] The computing resources are scheduled according to the resource scheduling instructions to dynamically allocate AI computing resources.

[0041] A dynamic allocation system for AI computing resources, used to implement the aforementioned dynamic allocation method for AI computing resources, includes:

[0042] The fragment resource map construction module analyzes the resource status based on the pre-acquired computing power resource status data, identifies discretely distributed fragment resources through a preset status analysis model, calculates the corresponding resource value, and constructs a fragment resource map.

[0043] The fragment resource combination module, in response to a preset computing task, configures a fragment combination mechanism, selects fragment resources from the fragment resource map whose fragment resource recombination benefits are greater than a preset benefit threshold, and combines them to obtain fragment resource blocks.

[0044] The resource allocation module is configured with a resource filtering mechanism to dynamically filter out computing resources that can meet the computing task when combined with the fragmented resource blocks, and obtain a resource allocation scheme.

[0045] The resource scheduling module schedules computing resources based on the resource allocation scheme to dynamically allocate AI computing resources.

[0046] The beneficial effects of this application are as follows: Based on the identified fragmented resources, a fragmented resource map is constructed, tasks and resources are instantiated as intelligent agents, the optimal resource combination is selected by simulating the competition process of multiple intelligent agents, non-fragmented resources and recombined fragmented resource blocks are logically combined and the resource scheduling results are deduced to realize the dynamic allocation of computing resources; by effectively utilizing fragmented resources, the overall resource utilization rate of computing resources can be improved, the task response speed can be increased, fragmented resources can be actively combined to complete corresponding computing tasks, and task queuing and execution time can be reduced, thereby improving the dynamic scheduling efficiency of computing resources and the task execution efficiency. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a dynamic allocation method for AI computing resources in an embodiment of this application.

[0048] Figure 2 This is a schematic diagram of the fragment resource map in the embodiments of this application;

[0049] Figure 3 This is a flowchart illustrating the fragment resource block filtering process in the embodiments of this application;

[0050] Figure 4 This is a schematic diagram of the structure of an AI computing resource dynamic allocation system according to an embodiment of this application. Detailed Implementation

[0051] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0052] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0053] Hereinafter, the terms "first," "second," and other generic terms are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0054] refer to Figure 1 The image shows a specific implementation of a dynamic allocation method for AI computing resources according to this application, including:

[0055] S101. Analyze the resource status based on the pre-acquired computing power resource status data, identify discretely distributed fragmented resources through the preset status analysis model, calculate the corresponding resource value, and construct a fragmented resource map.

[0056] S102. In response to a preset computing task, configure a fragment combination mechanism, select fragment resources from the fragment resource map whose fragment resource recombination benefits are greater than a preset benefit threshold, and combine them to obtain fragment resource blocks.

[0057] S103. Configure a resource filtering mechanism to dynamically filter out computing resources that can meet the computing task after being combined with the fragmented resource blocks, and obtain a resource allocation scheme.

[0058] S104. Based on the resource allocation scheme, the computing resources are scheduled to dynamically allocate AI computing resources.

[0059] Large-scale computing tasks that run for a long time will occupy a lot of computing resources. If the computing core has a short period of idle time, a lot of scattered and underutilized computing resources will remain during the computing process. Because these fragmented resources are scattered, have short lifespans, and vary in form, existing computing resource scheduling systems have difficulty effectively identifying and utilizing them, resulting in wasted computing resources and low resource utilization.

[0060] This embodiment uses a state analysis model to perform time-series analysis and feature extraction on the pre-acquired computing resource state data, predicts resource idle time, identifies discretely distributed fragmented resources, calculates the value of each fragmented resource, abstracts resources and tasks into entities, and constructs a fragmented resource map. After responding to a computing task, a task feature vector is extracted using a task parsing model, and a resource encoder is used to semantically encode the map resources. A multi-agent competition mechanism is used to select resources with high matching degree, and the corresponding recombination benefits are calculated. Fragmented resources with recombination benefits exceeding a preset benefit threshold are combined into fragmented resource blocks. The state of non-fragmented resources is dynamically analyzed, and the process of combining them with fragmented resource blocks to run tasks is simulated using a sandbox simulation model. Based on performance index analysis, a set of non-fragmented resources that meets the computing task requirements and has the best overall efficiency is selected to obtain a resource allocation scheme. The resource allocation scheme is compiled into a scheduling instruction sequence and sent to the resource management platform through a standard interface to execute the actual resource allocation and task deployment, completing the dynamic allocation.

[0061] It is important to emphasize that by identifying, analyzing, combining, and scheduling fragmented computing resources, this embodiment transforms discrete resources that traditional resource scheduling systems cannot utilize into highly efficient and usable computing power supply, thereby improving the overall utilization rate of computing resources and task throughput. The sandbox simulation mechanism can avoid the blindness of allocation decisions, ensure the performance and stability of computing tasks, improve the scheduling efficiency and resource utilization of AI computing resources, and achieve real-time response to computing tasks.

[0062] In this embodiment, the resource status is analyzed based on the pre-acquired computing resource status data. The availability status of each computing resource in a specific future time period is predicted by a preset status analysis model, and discretely distributed fragmented resources are identified. The resource value corresponding to the fragmented resources is calculated by combining factors such as the computing performance, network location, energy efficiency, and remaining validity period of the fragments, and a fragmented resource map is constructed. The model can quickly and accurately identify and extract the corresponding fragmented resources, solving the problem of lack of identification and perception of distributed resources in the traditional resource scheduling process. By constructing a fragmented resource map, data support is provided for resource scheduling decisions, improving the accuracy and applicability of resource scheduling.

[0063] In response to preset computing tasks, a fragment combination mechanism is configured to analyze task requirements and encode resource features. The resource requirement feature vector of the computing task is extracted, and the corresponding resource vector is calculated. Based on the resource requirement feature vector and the resource vector, fragment resources with a fragment resource recombination benefit greater than a preset benefit threshold are selected from the fragment resource map and combined to obtain fragment resource blocks. By matching computing tasks and fragment resources and evaluating benefits, the value utilization of fragment resources is maximized, avoiding performance instability caused by random combination of fragment resources. Combining fragment resource map with fragment resource screening improves the accuracy and efficiency of resource screening.

[0064] Specifically, a resource screening mechanism is configured to simulate the resource combination process. During the simulation, the entire lifecycle of each candidate resource combination is simulated, and performance index data during task execution is recorded. The computing power resources that can meet the computing task after being combined with fragmented resource blocks are dynamically screened to obtain a resource allocation scheme. The simulation simulation enables the performance prediction of the resource allocation scheme, avoiding performance uncertainty in the actual resource allocation process. The sandbox simulation mechanism can quickly complete the performance evaluation of multiple resource allocation schemes and screen out the best performing resource allocation scheme, thereby improving the overall utilization rate of computing power resources and the overall computing performance.

[0065] Based on the resource allocation scheme, the resource allocation scheme is parsed into a sequence of scheduling instructions that can be executed by the underlying infrastructure, including but not limited to resource reservation instructions, network configuration instructions, and container deployment instructions, to schedule computing resources and dynamically allocate AI computing resources; through standardized instruction conversion, the accurate execution of the resource allocation process is ensured, avoiding deviations between the scheme and the actual deployment, and improving the efficiency and accuracy of resource allocation.

[0066] This application constructs a fragmented resource map based on identified fragmented resources, instantiates tasks and resources as agents, simulates a multi-agent competition process to select the optimal resource combination, logically combines non-fragmented resources with recombined fragmented resource blocks, and deduces resource scheduling results to achieve dynamic allocation of computing resources. By effectively utilizing fragmented resources, the overall resource utilization rate of computing resources can be improved, task response speed can be increased, fragmented resources can be actively combined to complete corresponding computing tasks, and task queuing and execution time can be reduced, thereby improving the dynamic scheduling efficiency and task execution efficiency of computing resources.

[0067] Furthermore, based on the pre-acquired computing power resource status data, the resource status is analyzed, discretely distributed fragmented resources are identified through a preset status analysis model, and the corresponding resource value is calculated to construct a fragmented resource map, including:

[0068] S201. Based on the pre-acquired computing power resource status data, analyze the resource usage status, identify discretely distributed fragmented resources through a preset status analysis model, and obtain a set of fragmented resources.

[0069] S202. Calculate the resource value of each fragment resource in the fragment resource set, treat the fragment resources as resource layer entities, the calculation tasks as task layer entities, establish connections between the corresponding entities according to the resource value, and construct a fragment resource graph.

[0070] In this embodiment, based on the pre-acquired computing resource status data, the usage status of resources is analyzed. A preset status analysis model identifies discretely distributed fragmented resources. Trend analysis is performed on time-series data such as resource utilization, memory usage, and network traffic in the computing resource status data to identify periodic patterns in resource usage and predict the probability of each computing resource being idle in a specific future time period. Based on the prediction results, a two-dimensional screening is performed: in the time dimension, idle periods shorter than the complete task execution cycle are selected; in the spatial dimension, unused resource capacity is identified, resulting in a set of fragmented resources. This dual mechanism of prediction and identification achieves accurate identification of fragmented resources, solving the problem of traditional resource monitoring and allocation focusing only on overall utilization while ignoring the local availability of fragmented resources. The combination of spatiotemporal dual-dimensional screening ensures the comprehensiveness and accuracy of fragment identification, providing comprehensive data support for resource integration and utilization, and expanding the system's available resource boundaries.

[0071] Specifically, the resource value of each fragment resource in the fragment resource set is calculated by combining the static and dynamic attributes of the fragments. Static attributes include, but are not limited to, hardware performance indicators, topology location, and reliability level, while dynamic attributes include, but are not limited to, remaining validity period, load volatility, and energy efficiency. Fragment resources are treated as resource layer entities, and computing tasks are treated as task layer entities. The matching degree between entities is analyzed based on the matching degree between task requirements and resource attributes. Connections are established between corresponding matching entities according to resource value to construct a fragment resource graph. By quantifying the resource value and constructing the resource graph, the practical value of fragment resources in different scenarios can be effectively reflected, the matching relationship between resources and tasks can be established, and decision support for computing resource scheduling can be provided, thereby improving the accuracy and efficiency of computing resource allocation.

[0072] Furthermore, based on the pre-acquired computing resource status data, the resource usage status is analyzed, and discretely distributed fragmented resources are identified through a preset status analysis model, resulting in a fragmented resource set, including:

[0073] S301. Based on the pre-acquired computing resource status data, analyze the resource consumption of running tasks through a preset resource status prediction model, and predict the resource idle time.

[0074] S302. Extract the static and dynamic load characteristics of each computing resource and construct a resource feature set;

[0075] S303. Combining the resource idle time and resource feature set, the fragmented resources discretely distributed within each preset time interval are identified through a preset state analysis model to obtain a fragmented resource set.

[0076] In this embodiment, time-series analysis is performed on the pre-acquired computing resource status data to predict resource availability at a specific future time point. A preset resource status prediction model is used to analyze the resource consumption of running tasks and predict resource idle time. The computing resource status data originates from a time-series data stream continuously collected by the cluster monitoring system, including but not limited to the computing power utilization rate, memory usage, process identifier, task queue status, and task start / stop timestamps of each computing unit. The computing resource status data undergoes data cleaning and normalization preprocessing to obtain preprocessed computing resource status data. This preprocessed time-series data is then input into the preset resource status prediction model for analysis. This includes, but is not limited to, Long Short-Term Memory (LSTM) network models. A large amount of historical computing resource status data is used to train the LTM network model, resulting in a pre-trained LTM network model. Preprocessed time-series data is then input into the pre-trained LTM network model, which analyzes whether the resource consumption curve of the currently running task is in an upward, stable, or downward phase. The model calculates predicted values ​​for the status of various indicators of the resource unit over a future period and outputs predicted resource idle time based on these values. By predicting the status of computing resources, idle periods can be predicted in advance and accurately, providing accurate data on computing resource status for the resource allocation process and enabling efficient dynamic allocation of resources.

[0077] Specifically, static and dynamic load characteristics of each computing resource are extracted. Static characteristics describe the inherent capabilities of the resource, while dynamic load characteristics describe the real-time status of the resource, directly affecting its immediate performance and behavior. A resource feature set is constructed. Static characteristics include, but are not limited to, hardware identification and model, and hardware capability indicators. Dynamic load characteristics include, but are not limited to, real-time utilization indicators and system environment indicators. The extracted static and dynamic load characteristics are integrated to obtain the resource feature set. By constructing a multi-dimensional resource feature set, the inherent attributes and real-time status of the resource are fused, providing a comprehensive and accurate data foundation for resource analysis and allocation, and improving the scientific nature and accuracy of decision-making.

[0078] By combining the calculated resource idle time and resource feature set, a pre-defined state analysis model is used to identify discretely distributed fragmented resources within each preset time interval, resulting in a fragmented resource set. The state analysis model includes, but is not limited to, a random forest model. A pre-trained random forest model is trained using a large amount of historical resource feature data. The resource idle time and resource feature set are input into the pre-trained random forest model, which identifies the corresponding fragmented resources in both time and space. Spatially, it extracts resources that are completely idle within the target time interval and identifies resources that are idle within the target time interval. The model outputs the identified fragmented resources, forming the fragmented resource set. Combining predictive information with resource features enables efficient and accurate filtering of potentially valuable, discretely distributed fragmented resources from the resource pool, expanding the available computing power resources.

[0079] Furthermore, the resource value of each fragment resource in the fragment resource set is calculated separately. Fragment resources are treated as resource-layer entities, and computation tasks as task-layer entities. Connections are established between corresponding entities according to resource value to construct a fragment resource graph, including:

[0080] S401. For each fragment resource in the fragment resource set, analyze and combine the static and dynamic attributes to calculate the corresponding resource value.

[0081] S402. Treat each fragment resource as a resource layer entity and each computing task as a task layer entity.

[0082] S403. Establish connections between the task layer entities and the resource layer entities required by the corresponding computing tasks according to the resource value, and obtain the fragment resource map.

[0083] In this embodiment, for each fragment resource in the fragment resource set, static and dynamic attributes are analyzed based on the fragment resource's inherent capabilities, current state, and environment to calculate its corresponding resource value, reflecting the overall utility of the fragment resource for computing tasks in the current system environment. For each fragment resource in the fragment resource set, corresponding static and dynamic attributes are extracted. Static attributes include, but are not limited to, hardware model, theoretical peak computing power, total memory capacity, network interface bandwidth, hardware generation, and supported instruction sets. Dynamic attributes include, but are not limited to, predicted idle start time, predicted idle duration, and overall load rate of the current node. A multi-factor weighted evaluation model is trained using a large amount of historical static and dynamic attribute data to obtain a pre-trained multi-factor weighted evaluation model. The static and dynamic attributes are input into the pre-trained multi-factor weighted evaluation model, which dynamically assigns corresponding weights to each attribute and weights and fuses the attribute values ​​according to the weights to output the resource value of each fragment resource. By combining static and dynamic attributes, the multi-dimensional attributes of fragment resources are transformed into numerical values, providing an accurate data foundation for resource allocation decisions. During resource scheduling and allocation, fragment resources with higher value can be prioritized, thereby significantly improving the overall efficiency of resource allocation.

[0084] like Figure 2 As shown, each fragment resource is treated as a resource layer entity, and each computing task as a task layer entity. Not all fragment resources can meet the needs of a specific task, nor are all resources that can meet the needs equally valuable. For each task layer entity, connections are established between it and all resource layer entities that can meet the basic needs. The resource value calculated for that task type is used as the weight of the connection edge, resulting in a fragment resource graph. For each task layer entity in the graph, the type of computing power required by the task is analyzed to obtain the resource requirement constraints corresponding to the task. Based on the resource requirement constraints, it is analyzed whether the resource layer entity meets the constraints, and resource layer entities that can meet the resource requirement constraints are selected. Connections are established between the selected resource layer entities and their corresponding task layer entities, and the resource value of the resource layer entity is used as the edge weight, constructing the fragment resource graph. By constructing the fragment resource graph, a reference is provided for the fragment resource selection and allocation process, enabling the rapid and efficient selection of high-value matching pairs, thereby achieving accurate resource matching and global optimization of resource utilization.

[0085] Furthermore, in response to a preset computational task, a fragment combination mechanism is configured to select fragment resources from the fragment resource map whose fragment resource recombination benefit is greater than a preset benefit threshold, and combine them to obtain fragment resource blocks, including:

[0086] S501. In response to a preset computation task, extract task features through a preset task parsing model and construct a task feature vector.

[0087] S502. Encode the fragment resources in the fragment resource map using a preset resource encoder to obtain fragment resource codes;

[0088] S503. Combining the task feature vector and fragment resource encoding, select fragment resources whose fragment resource recombination benefits are greater than a preset benefit threshold, and combine them to obtain fragment resource blocks.

[0089] In this embodiment, in response to a preset computing task, the description information of the computing task is obtained. This description information includes, but is not limited to, job configuration files, command-line parameters, and container image information. A preset task parsing model is used to analyze the description information and extract task features, constructing a task feature vector. The task parsing model includes, but is not limited to, a neural network model. The neural network model is trained using a large amount of historical task description information to obtain a pre-trained neural network model. The description information of the computing task is input into the pre-trained neural network model, which analyzes the information, extracts and quantizes the corresponding feature values, and outputs the task feature vector. Through the task parsing model, vague, text-based task requirements are transformed into precise, computable feature vectors, providing data support for resource encoding and resource matching.

[0090] Specifically, a pre-defined resource encoder encodes fragmented resources in the fragmented resource map, resulting in fragmented resource codes. The resource encoder includes, but is not limited to, a multilayer perceptron model. This model is trained using a large amount of fragmented resource map data to obtain a pre-trained multilayer perceptron model. This pre-trained model is then used to analyze the features of each resource layer entity in the fragmented resource map, performing nonlinear transformations and dimensionality reduction on the features, analyzing the interaction relationships between features, and outputting the fragmented resource codes. By transforming high-dimensional resource attributes into low-dimensional encoded information through the resource encoder, the complexity of resource matching calculations can be simplified and computational efficiency improved. This not only relies on simple rule matching but also incorporates the adaptation relationship between resources and tasks, improving the accuracy and efficiency of the fragmented resource matching process.

[0091] By combining task feature vectors and fragment resource codes, the matching degree between task feature vectors and each fragment resource code is analyzed. Considering topological affinity, hardware consistency, and time synchronization among resources, the recombination benefit between fragment resources is calculated. Fragment resources with recombination benefits exceeding a preset benefit threshold are selected and combined to obtain fragment resource blocks. This benefit-driven resource combination strategy maximizes the utilization of fragment resources, avoids performance instability and resource waste caused by random resource combinations, and improves resource integration effectiveness and overall system efficiency.

[0092] like Figure 3As shown, by combining the task feature vector and fragment resource encoding, fragment resources whose fragment resource recombination benefits are greater than a preset benefit threshold are selected and combined to obtain fragment resource blocks, including:

[0093] S601. Instantiate the corresponding task agent based on the task feature vector, and instantiate multiple corresponding resource agents based on the fragment resource encoding.

[0094] S602. Multiple resource agents simulate a competition process. The task agent selects at least one resource agent according to the competition process to obtain a set of resource agents. The resource agents compete based on the matching degree between the resource code and the task feature vector.

[0095] S603. Based on the similarity between fragmented resources, calculate the recombination benefit between fragmented resources in the resource agent set;

[0096] S604. Select fragment resources from the resource intelligence set whose recombination benefits are greater than a preset benefit threshold, and combine them to obtain fragment resource blocks.

[0097] In this embodiment, a corresponding task agent is instantiated based on the task feature vector, and multiple corresponding resource agents are instantiated based on the fragment resource encoding. The computational task and fragment resources are instantiated as agents with autonomous decision-making capabilities. The goal of the task agent is to efficiently acquire the required resources; the goal of the resource agents is to match resources to the most suitable task. The task agent instantiated based on the task feature vector includes all resource requirement information and decision-making capabilities described by the task feature vector; the resource agent instantiated based on the fragment resource encoding includes all information and decision-making capabilities related to the fragment resources. By constructing multiple agents, the complex resource scheduling problem is transformed into an easily manageable agent interaction problem, providing a new approach to achieving efficient and adaptive resource matching and enhancing resource matching efficiency.

[0098] Specifically, multiple resource agents compete based on the matching degree between resource codes and task feature vectors. The competition process is simulated, and the task agent selects at least one resource agent, resulting in a resource agent set. Each resource agent, based on its internal state, calculates the matching degree between the corresponding fragment resource and the task agent's needs by comparing the similarity between the fragment resource code and the task feature vector. All resource agents send their calculated matching degree scores as competition signals to the task agent. The task agent receives all competition signals, filters out a set of fragment resources that can jointly meet the needs, selects the resource agent with the highest matching degree, analyzes the remaining resource agents for resources that can complement the selected resources, and selects the next resource with the highest matching degree, thus obtaining the resource agent set. By simulating the agent competition process, efficient resource pre-screening is achieved, allowing for rapid selection of the corresponding fragment resource set and improving the efficiency of fragment resource screening and combination.

[0099] The selected resource agent set is analyzed. Based on the similarity between fragmented resources, the recombination benefit between fragmented resources in the resource agent set is calculated. For each fragmented resource in the resource agent set, its attribute features are extracted, and an attribute feature vector is constructed. For resource pairs in the set, the similarity between corresponding attribute feature vectors is calculated. The similarity calculation method includes, but is not limited to, calculating the cosine similarity between attribute feature vectors. A linear regression model is trained using a large amount of historical fragmented resource similarity data to obtain a pre-trained linear regression model. The calculated similarity between fragmented resources is input into the pre-trained linear regression model, which maps the similarity to the corresponding combined benefit and outputs the recombination benefit between fragmented resources in the resource agent set. The model matches the calculated similarity between fragmented resource pairs in the pre-trained linear regression curve. The corresponding combined return is calculated by summing the combined returns of every two pairs of fragment resources in the recombined combination. For example, a recombined combination includes fragment resource A, fragment resource B, and fragment resource C. The similarity between fragment resource A and fragment resource B is calculated, and their combined return is matched against a linear regression curve. Similarly, the similarity between fragment resource A and fragment resource C is calculated, and their combined return is matched against a linear regression curve. The combined returns of fragment resource A and fragment resource B, fragment resource A and fragment resource C, and fragment resource B and fragment resource C are then summed to obtain the recombined return of the combination. By calculating the recombined return, simple combinations of fragment resources are avoided; instead, a high-quality combination of fragment resources based on recombinant synergies is achieved, which can improve the actual execution performance and stability of the combined fragment resources.

[0100] From the resource agent set, fragmented resources whose recombination benefits exceed a preset benefit threshold are selected and combined to obtain fragmented resource blocks. Based on the minimum acceptable benefit requirement for resource recombination operations, a benefit threshold is set. The recombination benefit between each pair of fragmented resource combinations is compared with the benefit threshold, and fragmented resources whose recombination benefits exceed the threshold are selected and combined to obtain fragmented resource blocks. By selecting fragmented resources whose combined benefits meet the benefit requirements, ineffective resource scheduling and resource jitter can be avoided, improving the utilization efficiency and resource utilization rate of the combined fragmented resources.

[0101] Furthermore, a resource filtering mechanism is configured to dynamically filter out computing resources that, when combined with the fragmented resource blocks, can satisfy the computing task, thereby obtaining a resource allocation scheme, including:

[0102] S701. Analyze the computing resource requirements of the computing task, perform resource status analysis on non-fragmented resources, and generate corresponding resource status results.

[0103] S702. Based on the resource status results, simulate the combination process of non-fragmented resources and fragmented resource blocks through a preset sand table simulation model, and select a set of non-fragmented resources that can meet the computing task after combination.

[0104] S703. Combining the non-fragmented resource set and the fragmented resource block, a resource allocation scheme is obtained.

[0105] In this embodiment, the computational resource requirements of the computing task are analyzed, and the resource status of non-fragmented resources is analyzed to generate corresponding resource status results. Based on the task feature vector, the computational task's requirements for complete resource units are analyzed, including but not limited to the total GPU memory and computing power resources required by the task. Based on the real-time information of resource units not marked as fragmented resources, the availability status, resource capacity, and other status information of non-fragmented resources are analyzed to generate corresponding resource status results. By performing status analysis on non-fragmented resources, resource scheduling failures due to incomplete information or misjudgment of status can be avoided, providing accurate data for resource allocation and integration, and ensuring the successful implementation of the resource allocation scheme.

[0106] Specifically, based on the resource status results, the process of combining non-fragmented resources and fragmented resource blocks is simulated through a preset sandbox simulation model. The corresponding performance indicators are calculated, and the set of non-fragmented resources that can meet the computing task requirements and have the best performance after combination is selected. By simulating resource combination, different resource combination methods can be evaluated, the optimal combination method can be selected, and the quality of resource scheduling and computing performance can be improved.

[0107] By integrating non-fragmented resource sets and fragmented resource blocks, and setting corresponding logical relationships and collaborative working methods, a resource allocation scheme is obtained. By setting the resource allocation scheme, it can be ensured that the resource allocation scheme can be accurately and error-free transmitted to the execution layer, thereby improving resource allocation efficiency and resource utilization, and thus responding to computing tasks in real time.

[0108] Furthermore, based on the resource status results, a pre-set sandbox simulation model is used to simulate the combination process of non-fragmented resources and fragmented resource blocks, and a set of non-fragmented resources that can satisfy the computing task after combination is selected, including:

[0109] S801. Based on the resource status results, simulate the logical combination of each non-fragmented resource and fragmented resource block through a preset sand table simulation model, and run the corresponding computing task to obtain the simulation results.

[0110] S802. Analyze the performance indicators of the simulation results and construct the result performance indicators;

[0111] S803. Select the set of non-fragmented resources that can meet the computing task and achieve the best performance indicators after combination.

[0112] In this embodiment, based on the resource status results, a pre-set sandbox simulation model is used to simulate the logical combination of each non-fragmented resource and fragmented resource block, and to run the corresponding computing tasks. The resource allocation scheme of each combination is evaluated to obtain simulation results. The sandbox simulation model includes, but is not limited to, a nonlinear regression model. A large amount of historical resource status result data is used to train the nonlinear regression model to obtain a pre-trained nonlinear regression model. The resource status results are input into the pre-trained nonlinear regression model. The model logically associates and combines non-fragmented resources and fragmented resource blocks according to the resource status results and the attributes of fragmented resource blocks, constructs virtual computing units, uses virtual computing units to simulate task execution, and records performance data during the execution process to obtain simulation results. By simulating the resource combination and task execution process, performance indicators under different resource combination conditions can be analyzed. Based on the performance indicators of the simulated conditions, the optimal resource combination can be selected, which can reduce the risk of actual resource scheduling failure and improve the resource scheduling and allocation effect.

[0113] Specifically, the performance indicators of the simulation results are analyzed, and performance indicators are constructed. These indicators are extracted from the simulation results, including but not limited to time, efficiency, and stability indicators. The extracted performance indicators are then weighted and fused to obtain the final performance indicators. By analyzing and calculating the simulation results performance indicators, accurate data support is provided for the resource combination process.

[0114] Specifically, the process involves selecting a set of non-fragmented resources that, when combined, meet the computational task and yield the optimal performance metrics. It then iterates through the combinations of non-fragmented and fragmented resource blocks and their simulation results for each solution. The process selects combinations whose total resource capacity is greater than or equal to the task requirements, satisfies the task's hard constraints, and does not encounter any unsolvable errors during simulation. Finally, the combination with the optimal performance metrics is chosen from these selected combinations to obtain the corresponding set of non-fragmented resources. By filtering based on task requirements and performance, it is ensured that the selected solution is theoretically feasible and delivers optimal performance in practice, thereby improving the overall efficiency and effectiveness of computational resource execution.

[0115] Furthermore, based on the resource allocation scheme, computing resources are scheduled to dynamically allocate AI computing resources, including:

[0116] S901. Generate corresponding resource scheduling instructions based on the resource allocation scheme;

[0117] S902. The computing resources are scheduled according to the resource scheduling instructions to dynamically allocate AI computing resources.

[0118] In this embodiment, based on the resource allocation scheme, a corresponding resource scheduling instruction is generated; the resource allocation scheme is parsed to extract the resource identifiers, task deployment descriptions, network and storage configuration requirements, etc., from the non-fragmented resource set and fragmented resource blocks to obtain allocation elements; and the corresponding resource scheduling instruction is generated based on the allocation elements. By converting the scheme into specific scheduling instructions, the decision scheme can be executed accurately, thereby improving the accuracy and efficiency of the dynamic allocation process of computing resources.

[0119] Specifically, computing resources are scheduled according to resource scheduling instructions to dynamically allocate AI computing resources. Through client tools or API interfaces provided by the system, the sequence of resource scheduling instructions is sent to the cluster's master node to execute the corresponding resource allocation operations. By executing the corresponding resource scheduling instructions, the accurate execution of the resource allocation scheme can be ensured, thereby improving the resource utilization and task throughput of the AI ​​computing cluster.

[0120] like Figure 4 As shown, an AI computing resource dynamic allocation system is used to implement a method for dynamically allocating AI computing resources, including:

[0121] The fragment resource map construction module analyzes the resource status based on the pre-acquired computing power resource status data, identifies discretely distributed fragment resources through a preset status analysis model, calculates the corresponding resource value, and constructs a fragment resource map.

[0122] The fragment resource combination module, in response to a preset computing task, configures a fragment combination mechanism, selects fragment resources from the fragment resource map whose fragment resource recombination benefits are greater than a preset benefit threshold, and combines them to obtain fragment resource blocks.

[0123] The resource allocation module is configured with a resource filtering mechanism to dynamically filter out computing resources that can meet the computing task when combined with the fragmented resource blocks, and obtain a resource allocation scheme.

[0124] The resource scheduling module schedules computing resources based on the resource allocation scheme to dynamically allocate AI computing resources.

[0125] In this embodiment, the fragmented resource graph construction module identifies underutilized fragmented resource segments that are discretely distributed in time or space from computing resource status data through a state analysis model. It establishes a multi-dimensional resource value assessment system, calculates corresponding value indicators for each fragmented resource, and abstracts fragmented resources and computing tasks as entities and connections in a graph, constructing a fragmented resource graph. This transforms wasted fragmented resources from an unusable state into assessable and schedulable resources, providing more computing resources for resource scheduling and optimization, and improving resource scheduling efficiency and resource utilization. The fragmented resource combination module responds to computing task requests and dynamically combines fragmented resources. When a new task arrives, it performs intelligent search and matching in the fragmented resource graph according to the fragment combination mechanism, simulating a multi-agent competition process to select corresponding fragmented resources. These selected fragmented resources are then integrated to obtain fragmented resource blocks. Through fragmented resource integration, multiple low-value, scattered fragmented resources can be integrated into high-value, usable resource blocks, increasing the effective resource supply of the cluster and improving resource scheduling efficiency and resource utilization.

[0126] Specifically, the resource allocation module receives fragmented resource blocks and configures a resource filtering mechanism, including a sandbox simulation model, to dynamically simulate the computation process of running computing tasks after various combinations of non-fragmented resources and fragmented resource blocks. From all feasible solutions, it selects the set of non-fragmented resources that meets the task requirements and achieves optimal global performance, thus obtaining a resource allocation scheme. Through the sandbox simulation mechanism, the risk of substandard task performance or scheduling failures due to improper resource combinations can be reduced, improving the stability and reliability of the resource scheduling process. Through globally optimal decision-making, the overall efficiency of the resource allocation process is improved, maximizing the benefits of resource utilization. The resource scheduling module transforms the resource allocation scheme into a standard scheduling instruction sequence, ensuring that computing tasks are accurately deployed to the physical resources specified in the scheme, completing the dynamic allocation and scheduling of AI computing resources.

[0127] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.

Claims

1. A method for dynamically allocating AI computing resources, characterized in that, include: Based on the pre-acquired computing power resource status data, analyze the resource status, identify discretely distributed fragmented resources through a preset status analysis model, calculate the corresponding resource value, and construct a fragmented resource map; In response to a preset computational task, task features are extracted using a preset task parsing model, and a task feature vector is constructed. The fragment resources in the fragment resource map are encoded by a preset resource encoder to obtain fragment resource codes; Instantiate the corresponding task agent based on the task feature vector, and instantiate multiple corresponding resource agents based on the fragment resource encoding; Multiple resource agents simulate a competition process. The task agent selects at least one resource agent based on the competition process, resulting in a set of resource agents. The resource agents compete based on the matching degree between the resource code and the task feature vector. Based on the similarity between fragmented resources, calculate the recombination benefit between fragmented resources in the resource agent set; From the set of resource agents, select fragment resources whose recombination benefits are greater than a preset benefit threshold, and combine them to obtain fragment resource blocks; Configure a resource filtering mechanism to dynamically filter out computing resources that can meet the computing task after being combined with the fragmented resource blocks, and obtain a resource allocation scheme; The computing resources are scheduled based on the resource allocation scheme to dynamically allocate AI computing resources.

2. The AI ​​computing power resource dynamic allocation method according to claim 1, characterized in that, The step of analyzing resource status based on pre-acquired computing power resource status data, identifying discretely distributed fragmented resources through a preset status analysis model, calculating the corresponding resource value, and constructing a fragmented resource map includes: Based on the pre-acquired computing resource status data, the usage status of the resources is analyzed, and discretely distributed fragmented resources are identified through a preset status analysis model to obtain a set of fragmented resources. Calculate the resource value of each fragment resource in the fragment resource set, treat the fragment resources as resource layer entities, treat the calculation tasks as task layer entities, establish connections between the corresponding entities according to resource value, and construct a fragment resource graph.

3. The method for dynamic allocation of AI computing resources according to claim 2, characterized in that, The process involves analyzing the usage status of resources based on pre-acquired computing resource status data, identifying discretely distributed fragmented resources through a preset status analysis model, and obtaining a set of fragmented resources, including: Based on the pre-acquired computing resource status data, the resource consumption of running tasks is analyzed through a preset resource status prediction model to predict resource idle time. Extract the static and dynamic load characteristics of each computing resource to construct a resource feature set; By combining the resource idle time and resource feature set, a set of fragmented resources discretely distributed within each preset time interval is identified through a preset state analysis model, thus obtaining a set of fragmented resources.

4. The method for dynamic allocation of AI computing resources according to claim 2, characterized in that, The step of calculating the resource value of each fragment resource in the fragment resource set, treating fragment resources as resource layer entities and computation tasks as task layer entities, establishing connections between corresponding entities according to resource value, and constructing a fragment resource graph includes: For each fragment resource in the fragment resource set, analyze and combine static and dynamic attributes to calculate the corresponding resource value; Each fragment resource is treated as a resource layer entity, and each computing task is treated as a task layer entity. Based on the resource value, connections are established between the task-level entities and the resource-level entities required by the corresponding computing tasks to obtain a fragmented resource map.

5. The method for dynamic allocation of AI computing resources according to claim 1, characterized in that, The resource filtering mechanism dynamically filters out computing resources that, when combined with the fragmented resource blocks, can satisfy the computing task, resulting in a resource allocation scheme, including: Analyze the computing resource requirements of computing tasks, perform resource status analysis on non-fragmented resources, and generate corresponding resource status results; Based on the resource status results, the combination process of non-fragmented resources and fragmented resource blocks is simulated through a preset sand table simulation model, and a set of non-fragmented resources that can meet the computing task after combination is selected. By combining the non-fragmented resource set and the fragmented resource blocks, a resource allocation scheme is obtained.

6. The method for dynamic allocation of AI computing resources according to claim 5, characterized in that, Based on the resource status results, a pre-set sandbox simulation model is used to simulate the combination process of non-fragmented resources and fragmented resource blocks. A set of non-fragmented resources that, after combination, can satisfy the computational task is selected, including: Based on the resource status results, the process of simulating the logical combination of each non-fragmented resource and fragmented resource block through a preset sand table simulation model, and running the corresponding computing tasks, is used to obtain simulation results. The performance metrics of the simulation results are analyzed, and the result performance metrics are constructed. Select the set of non-fragmented resources that can meet the computing task and achieve the best performance indicators after combination.

7. The method for dynamic allocation of AI computing resources according to claim 1, characterized in that, Based on the resource allocation scheme, computing resources are scheduled to dynamically allocate AI computing resources, including: Based on the resource allocation scheme, generate corresponding resource scheduling instructions; The computing resources are scheduled according to the resource scheduling instructions to dynamically allocate AI computing resources.

8. A dynamic allocation system for AI computing resources, characterized in that, A method for dynamically allocating AI computing resources as described in any one of claims 1 to 7, comprising: The fragment resource map construction module analyzes the resource status based on the pre-acquired computing power resource status data, identifies discretely distributed fragment resources through a preset status analysis model, calculates the corresponding resource value, and constructs a fragment resource map. The fragment resource combination module, in response to a preset computing task, configures a fragment combination mechanism, selects fragment resources from the fragment resource map whose fragment resource recombination benefits are greater than a preset benefit threshold, and combines them to obtain fragment resource blocks. The resource allocation module is configured with a resource filtering mechanism to dynamically filter out computing resources that can meet the computing task when combined with the fragmented resource blocks, and obtain a resource allocation scheme. The resource scheduling module schedules computing resources based on the resource allocation scheme to dynamically allocate AI computing resources.

Images