A virtual aggregation management method for computing cluster idle computing power fragments

Abstract

The application discloses a kind of virtual aggregation management methods for calculating cluster idle computing power fragments, and it is related to data processing technical field.The method comprises the following steps: inputting the first consumption sequence of target resource type of current computing task into consumption prediction model, to determine the reserved resource amount of target resource type;Determine the resource idle index of first resource unit in target resource type based on the distance between first resource unit and second resource unit;Determine the idle resource unit of target resource type based on the reserved resource amount and the resource idle index of first resource unit in target resource type;Determine the virtualization cost between idle resource units based on the occupied resource size of idle resource unit in target resource type, resource total quantity, task survival duration and the distance between each idle resource unit;Virtualize and aggregate between idle resource units based on the virtualization cost between idle resource units.The application can improve the reliability of resource management.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a virtual aggregation management method for idle computing power fragments in a computing cluster. Background Technology

[0002] With the development of artificial intelligence and high-performance computing, the demand for computing power, such as graphics processing units (GPUs), has surged, and large-scale computing clusters have become the core infrastructure for technological innovation. Traditional resource management uses computing nodes as the allocation unit. In multi-task and multi-user scenarios, idle resources are scattered. Even if the total amount of idle resources is sufficient, tasks may not be able to run due to the lack of suitable computing nodes, resulting in wasted computing power.

[0003] Currently, virtualization technology is used to abstract and aggregate the idle computing power fragments of different computing nodes into virtual computing nodes at the logical layer, forming a unified resource pool to solve the problem of resource fragmentation.

[0004] However, existing aggregation methods do not fully consider the communication costs and resource stability between resource units, aggregating only based on the proximity principle. This can easily lead to communication delays within resource units and poor reliability of resource management. Summary of the Invention

[0005] This invention provides a virtual aggregation management method for idle computing power fragments in a computing cluster, which can improve the reliability of resource management.

[0006] A first aspect of this invention provides a virtual aggregation management method for idle computing power fragments in a computing cluster, comprising:

[0007] The first consumption sequence of the target resource type at the current moment of the current computing task is input into the consumption prediction model to predict the future demand of the target resource type and determine the reserved resource amount of the target resource type.

[0008] Based on the distance between each first resource unit and each second resource unit, the resource availability index of each first resource unit in the target resource type is determined; the first resource unit is the resource unit with available resources of the target resource type, and the second resource unit is the resource unit whose current computing task has used resources of the target resource type.

[0009] Based on the reserved resource amount of the target resource type and the resource idle index of each first resource unit in the target resource type, determine the idle resource units of the target resource type;

[0010] Based on the occupied resource size, total resource quantity, task duration, and distance between each idle resource unit in the target resource type, the virtualization cost between each idle resource unit is determined; the virtualization cost is used to characterize the communication cost of merging idle resource units into the same virtual computing node.

[0011] Based on the virtualization cost between each idle resource unit, virtualization aggregation is performed between idle resource units to obtain the optimal aggregation scheme for the target resource type.

[0012] Furthermore, this application also proposes that, before inputting the first consumption sequence of the target resource type for the current computing task at the current moment into the consumption prediction model to predict the future demand of the target resource type and determine the reserved resource amount of the target resource type, the following steps are also included:

[0013] Obtain a reference computing task that belongs to the same computing type as the current computing task;

[0014] Based on the second consumption sequence of the current computing task in the target resource type and the third consumption sequence of each reference computing task in the target resource type, the task similarity between the current computing task and each reference computing task is determined.

[0015] Based on the task similarity between the current computing task and each reference computing task, the prediction reference weights of each reference computing task are determined.

[0016] Based on the predicted reference weights of each reference computing task and the complete consumption sequence of each reference computing task, the target model is trained to obtain the consumption prediction model.

[0017] Furthermore, this application also proposes obtaining a reference computing task of the same computing type as the current computing task, including:

[0018] Based on the resource consumption differences among the fourth consumption sequence of target resource types for each historical computing task, each historical computing task is clustered to obtain at least one computing type cluster.

[0019] Based on the resource consumption difference between the current computing task and the target resource type second consumption sequence and the historical computing tasks in the computing type cluster in the target resource type fourth consumption sequence, the task consistency between the current computing task and the computing type cluster is determined.

[0020] Each historical computing task in the computing type cluster with the highest task consistency is identified as a reference computing task belonging to the same computing type as the current computing task.

[0021] Furthermore, this application proposes determining the task consistency between the current computing task and the computing type cluster based on the resource consumption difference between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type within the computing type cluster, including:

[0022] The resource consumption differences between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type in the computing type cluster are negatively correlated and normalized to obtain at least one consistency evaluation value.

[0023] The consistency evaluation values ​​are averaged to obtain the task consistency degree between the current computing task and the computing type cluster.

[0024] Furthermore, this application also proposes determining the task similarity between the current computing task and each reference computing task based on a second consumption sequence of the current computing task in the target resource type and a third consumption sequence of each reference computing task in the target resource type, including:

[0025] The second consumption sequence of the current computing task in the target resource type is compared with the third consumption sequence of each second reference computing task in the target resource type to construct the first computing preference vector of the current computing task in the target computing type; the target computing type is the computing type to which the current computing task belongs.

[0026] The third consumption sequence of the first reference computing task in the target resource type is compared with the third consumption sequence of each second reference computing task in the target resource type to construct a second computing preference vector of the first reference computing task in the target computing type; the first reference computing task is any reference computing task, and the second reference computing tasks are reference computing tasks other than the first reference computing task.

[0027] The cosine similarity between the first computational preference vector and the second computational preference vector is determined as the task similarity between the current computational task and the first reference computational task.

[0028] Furthermore, this application also proposes determining the prediction reference weights for each reference computing task based on the task similarity between the current computing task and each reference computing task, including:

[0029] The difference in resource consumption between the current computing task in the second consumption sequence of the target resource type and the reference computing task in the third consumption sequence of the target resource type is divided by the task similarity between the current computing task and the reference computing task to determine the reference evaluation value of the reference computing task.

[0030] The reference evaluation values ​​are negatively correlated and normalized to obtain the predicted reference weights for the reference calculation task.

[0031] Furthermore, this application proposes inputting the first consumption sequence of the target resource type for the current computing task at the current moment into the consumption prediction model to predict the future demand of the target resource type and determine the reserved resource amount for the target resource type, including:

[0032] Input the first consumption sequence of the target resource type of the current computing task at the current time into the consumption prediction model to obtain the predicted consumption sequence of the target resource type of the current computing task;

[0033] Based on the maximum predicted consumption value in the predicted consumption sequence, determine the reserved resource amount for the target resource type.

[0034] Furthermore, this application also proposes determining the resource idle index of each first resource unit in the target resource type based on the distance between each first resource unit and each second resource unit, including:

[0035] The average distance between the first resource unit and each of the second resource units is taken as the average distance.

[0036] By performing positive correlation normalization on the distance mean, the resource availability index of the first resource unit in the target resource type is obtained.

[0037] Furthermore, this application proposes determining the virtualization cost between idle resource units based on the occupied resource size, total resource quantity, task duration, and distance between each idle resource unit in the target resource type, including:

[0038] The resource call cost of the idle resource unit is determined based on the occupied resource size, total resource quantity, and task duration of the idle resource unit in the target resource type.

[0039] The virtualization cost between the first idle resource unit and the second idle resource unit is determined based on the first resource call cost of the first idle resource unit, the second resource call cost of the second idle resource unit, and the distance between the first idle resource unit and the second idle resource unit.

[0040] Furthermore, this application also proposes a method for virtualizing idle resource units based on their virtualization costs, thereby obtaining the optimal aggregation scheme for the target resource type, including:

[0041] Each virtual computing node to be built is allocated an idle resource unit to obtain an initialized virtual computing node;

[0042] Using the initialized virtual computing node as the first node, the remaining idle resource units as the second node, and the negative correlation normalized value of the virtualization cost between the first node and the second node as the corresponding edge weight, the optimal aggregation scheme for the target resource type is determined by the Hungarian algorithm.

[0043] The present invention has the following beneficial effects:

[0044] The virtual aggregation management method for idle computing power fragments in a computing cluster provided in this embodiment of the invention first determines the reserved resource amount for the target resource type through a consumption prediction model, which can reserve appropriate resources for tasks in advance and avoid the impact of insufficient resources on task operation. Then, it determines the resource idle index based on distance, which can more reasonably evaluate the availability of resource units. Next, it determines idle resource units based on the reserved resource amount and the resource idle index, ensuring the rationality of resource selection. Then, it considers the size of occupied resources, the total number of resources, the duration of task existence, and distance to determine the virtualization cost, comprehensively measuring the communication cost of merging resource units into virtual computing nodes. Finally, it performs virtualization aggregation based on the virtualization cost to obtain the optimal aggregation scheme. By comprehensively considering multiple factors, it avoids the problem of relying solely on the proximity principle, reduces the generation of communication delays, and thus improves the reliability of resource management. Attached Figure Description

[0045] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating a virtual aggregation management method for idle computing power fragments in a computing cluster, provided in one embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram illustrating the construction process of a consumption prediction model provided in one embodiment of the present invention. Detailed Implementation

[0048] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a virtual aggregation management method for idle computing power fragments in a computing cluster proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0050] In traditional computing cluster resource management, the communication cost and resource stability between resource units are not fully considered in virtual aggregation operations. Aggregation is based solely on the principle of proximity in physical location, which leads to increased internal communication latency after the formation of virtual computing nodes and reduced reliability of resource management.

[0051] If the above problems are not resolved, the continued existence of communication delays will hinder the execution of computing tasks, while insufficient reliability of resource management will lead to task failures or system anomalies, seriously affecting the ability of computing clusters to support technological innovation infrastructure.

[0052] In this regard, such as Figure 1 As shown, this invention proposes a virtual aggregation management method for idle computing power fragments in a computing cluster. This method can be applied to electronic devices and includes the following steps S110 to S150:

[0053] S110, Input the first consumption sequence of the target resource type of the current computing task at the current time into the consumption prediction model, predict the future demand of the target resource type, and determine the reserved resource amount of the target resource type;

[0054] S120, based on the distance between each first resource unit and each second resource unit, determine the resource idle index of each first resource unit in the target resource type; the first resource unit is the resource unit with idle resources of the target resource type, and the second resource unit is the resource unit that has used the target resource type resources in the current computing task;

[0055] S130, based on the reserved resource amount of the target resource type and the resource idle index of each first resource unit in the target resource type, determine the idle resource unit of the target resource type;

[0056] S140, based on the occupied resource size, total resource quantity, task duration, and distance between each idle resource unit in the target resource type, determine the virtualization cost between each idle resource unit; the virtualization cost is used to characterize the communication cost of merging idle resource units into the same virtual computing node;

[0057] S150: Based on the virtualization cost between each idle resource unit, virtualization aggregation is performed between idle resource units to obtain the optimal aggregation scheme for the target resource type.

[0058] For ease of understanding, the following explains some key terms in this embodiment:

[0059] A computing cluster is a system consisting of multiple computer nodes interconnected by a network, working together to provide powerful computing capabilities. This cluster can handle large-scale data and complex computational tasks, and is the infrastructure for modern high-performance computing and artificial intelligence applications.

[0060] Idle computing fragments refer to the remaining, underutilized, scattered, and small pieces of computing resources on a single computing node in a computing cluster, resulting from factors such as task scheduling, resource allocation strategies, or task completion. Although the total amount of these fragmented resources may be considerable, their dispersed nature makes them difficult to utilize effectively by a single large task.

[0061] Virtual aggregation management is a management strategy that uses logical abstraction and reorganization techniques to aggregate scattered idle computing power fragments in a computing cluster to form one or more virtual computing nodes, thereby improving resource utilization and scheduling flexibility.

[0062] A current computing task refers to a computing workload that is running or about to run at a specific point in time. This task has a specific demand pattern and lifecycle for computing resources.

[0063] The target resource type refers to a specific type of computing resource required by the current computing task, such as video memory resources, streaming multiprocessor resources, video memory bandwidth, etc.

[0064] The first consumption sequence refers to the sequence of resource consumption data for the target resource type, arranged chronologically from a past period up to the current time. This sequence reflects the dynamic changes in the current computing task's demand for the target resource type.

[0065] A consumption prediction model is a mathematical model based on historical data and specific algorithms used to predict future resource consumption trends. This model can output the expected demand for a target resource type over a future period based on input data.

[0066] Reserved resources refer to the amount of resources planned and reserved in advance to ensure the smooth execution of current computing tasks, based on the prediction results of the future demand for target resource types using consumption prediction models.

[0067] The first resource unit refers to a physical or logical unit in a computing cluster that has idle resources of the target resource type, such as a compute node, a virtual machine, or a container.

[0068] The second resource unit refers to a physical or logical unit of resources of the target resource type that the current computing task has already used. These units may be located on the same physical node as the first resource unit, or they may be located on a different node.

[0069] The resource availability index is a numerical value used to quantify the degree of idleness or availability of a first resource unit in a target resource type. This index comprehensively considers factors such as the physical location of the resource unit and its distance from already used resource units.

[0070] Idle resource units refer to resource units in a computing cluster that are selected for virtual aggregation based on reserved resource quantity and resource idleness indicators.

[0071] Occupied resource size refers to the amount of target resource type resources in a free resource unit that have been occupied by other tasks but have not yet been released.

[0072] The total amount of resources refers to the total amount of resources of the target resource type that a single idle resource unit possesses.

[0073] Task duration refers to the expected duration for a currently running task on an idle resource unit. This parameter reflects the stability of the idle resource unit.

[0074] Virtualization cost refers to the cost incurred when merging two or more idle resource units into a single virtual computing node. This cost primarily characterizes factors such as communication latency and resource scheduling complexity within the merged virtual computing node.

[0075] A virtual computing node is a node that logically aggregates the idle computing power fragments of multiple physical resource units through virtualization technology, and presents itself to the outside world as a unified computing entity.

[0076] The optimal aggregation scheme refers to the combination of idle resource units that maximizes resource utilization, minimizes communication latency, and ensures resource stability after considering factors such as virtualization costs.

[0077] This embodiment provides a virtual aggregation management method for idle computing power fragments in a computing cluster. This method aims to solve the problems of resource fragmentation, high communication costs, and poor resource stability in existing technologies. Through refined resource prediction, idle resource identification, and virtualization cost assessment, it achieves optimal aggregation of idle computing power fragments.

[0078] This method first inputs the first consumption sequence of the target resource type at the current moment into the consumption prediction model to predict the future demand for that target resource type and determine the reserved resource amount for that target resource type. For example, a linear regression model can be used to fit the first consumption sequence, and a fixed proportion of reserved resources can be determined based on the fitting result. This approach can provide a preliminary basis for subsequent resource allocation.

[0079] Furthermore, based on the distance between each first resource unit and each second resource unit, a resource idle index for each first resource unit in the target resource type is determined. Specifically, the physical distance between each first resource unit and all second resource units can be calculated, and these distances can be accumulated and normalized to obtain the resource idle index for that first resource unit. The smaller the distance, the closer the resource unit is to the resources currently used by the task, and the more likely its resources are to be subsequently used; therefore, its resource idle index may be lower.

[0080] Based on this, and considering the reserved resource quantity for the target resource type and the resource availability index of each first resource unit within that target resource type, the available resource units for that target resource type are determined. For example, the first resource units are sorted in ascending order of their resource availability index, and the units with the highest resource availability index are selected. The corresponding first resource unit (wherein) This represents the amount of reserved resources for the predicted target resource type. The remaining portion represents the amount of resources currently in use, while the remaining portion represents the free resource units available for the current computing task.

[0081] Subsequently, based on the occupied resource size, total resource quantity, task duration, and distance between each idle resource unit of the target resource type, the virtualization cost between each idle resource unit is determined. Specifically, a base cost can be set for each idle resource unit, which is related to the occupied resource size, total resource quantity, and task duration of that unit. For example, a unit with less occupied resources, a larger total resource quantity, and a shorter task duration may have a lower base cost. Then, the base costs of two idle resource units are simply weighted and summed with the distance between them to obtain the virtualization cost between them. This virtualization cost is used to characterize the communication cost of merging the idle resource units into the same virtual computing node.

[0082] Finally, based on the virtualization cost between each idle resource unit, virtualization aggregation is performed among the idle resource units to obtain the optimal aggregation scheme for the target resource type.

[0083] The following example will provide a more detailed explanation of the above technical solution:

[0084] Suppose a large data analysis task (the current computing task) is running in a computing cluster, which primarily relies on Graphics Processing Unit (GPU) resources (the target resource type). The cluster contains multiple computing nodes, some of which have idle GPU computing power fragments.

[0085] First, the system collects a sequence of GPU resource consumption over a past period for the data analysis task. For example, the GPU consumption per minute for the task over the past 10 minutes might be [80%, 85%, 90%, 82%, 88%, 92%, 85%, 80%, 75%, 70%]. This sequence is then fed into a consumption prediction model. This model could be a simple moving average model, for example, calculating the average consumption over the past 5 minutes as a prediction of future demand. Assuming the prediction shows that the task's average GPU demand over the next 10 minutes is 85%, the system determines that the data analysis task's reserved GPU resources are 85%.

[0086] Next, the system identifies all first resource units (e.g., nodes A, B, C, and D) with available idle GPU resources in the cluster, as well as second resource units (e.g., nodes X and Y) where the GPU resources used by the data analysis task reside. The system calculates the network latency (distance) between each first resource unit and all second resource units. For example, the average latency between node A and nodes X and Y is 5ms, node B is 10ms, node C is 3ms, and node D is 12ms. The system directly uses these average latency values ​​as the resource availability index for each first resource unit. In this example, node D has the highest availability index (highest latency), followed by node B.

[0087] Then, based on the 85% GPU reserved resource amount and the resource idle index of each first resource unit, the system determines the final idle resource units. From the relationship between the resource idle indexes, it can be seen that the system will prioritize using node C, which has a lower idle index. Ultimately, nodes A, B, and D are determined as the idle resource units for this data analysis task.

[0088] The system then determines the virtualization cost between these idle resource units. For nodes A, B, and D, the system considers their respective occupied GPU size, total number of GPUs, task duration, and the distance between each other. Based on these parameters, the system calculates the virtualization cost between nodes A and B, A and D, and B and D.

[0089] Finally, based on these virtualization costs, the system performs virtualization aggregation on nodes A, B, and D to obtain the optimal aggregation scheme. For example, the system can use a heuristic algorithm to prioritize merging units with the lowest virtualization costs. If the virtualization cost between nodes A and D is the lowest, the system will aggregate nodes A and D into a single virtual computing node. If there are still remaining idle resource units (such as node B), and their virtualization costs with the already aggregated virtual nodes or new idle units meet the requirements, aggregation continues until the optimal aggregation scheme that satisfies the task requirements is formed.

[0090] Based on the above examples, this method, by introducing predictions of future demand for the target resource type, can more accurately determine the amount of reserved resources, avoiding resource waste caused by insufficient or excessive resource reservation in traditional methods. Unlike existing methods that aggregate solely based on proximity, this method, when determining idle resource units, considers not only the distance between the resource unit and the currently used resource units, but also comprehensively evaluates resource availability indicators, making the selected idle resource units more usable. More importantly, when determining the virtualization cost between idle resource units, this method incorporates factors such as the size of occupied resources, the total number of resources, and the duration of the task, in addition to considering the distance between units. This allows the aggregation decision to more comprehensively reflect the stability and communication efficiency of resource units, thereby avoiding the problem of high internal communication latency. Through a comprehensive evaluation of virtualization costs, this method can obtain a better aggregation scheme, effectively reducing the communication overhead within virtual computing nodes and improving the reliability of resource management.

[0091] This embodiment first determines the reserved resource quantity for the target resource type through a consumption prediction model, enabling the timely reservation of appropriate resources for tasks and preventing task operation from being affected by insufficient resources. Next, resource idleness indicators are determined based on distance, allowing for a more reasonable assessment of resource unit availability. Then, idle resource units are determined based on the reserved resource quantity and resource idleness indicators, ensuring the rationality of resource selection. Next, the virtualization cost is determined by considering the size of occupied resources, the total number of resources, task duration, and distance, comprehensively measuring the communication cost of merging resource units into virtual computing nodes. Finally, virtualization aggregation is performed based on the virtualization cost to obtain the optimal aggregation scheme. By comprehensively considering multiple factors, this approach avoids problems arising from relying solely on the proximity principle, reduces communication latency, and thus improves the reliability of resource management.

[0092] In some of the embodiments described above in this application, the accuracy of the consumption prediction model directly affects the rationality of resource reservation and the efficiency of subsequent idle computing power aggregation. If the consumption prediction model fails to fully learn and adapt to the resource consumption patterns of different computing tasks, its prediction results may have significant deviations, leading to excessive resource reservation and waste, or insufficient reservation affecting task execution, thereby affecting the resource utilization and performance of the entire computing cluster.

[0093] In this regard, such as Figure 2 As shown, this application further proposes that S210 to S240 are included before S110:

[0094] S210, Obtain a reference computing task that belongs to the same computing type as the current computing task;

[0095] S220, Based on the second consumption sequence of the current computing task in the target resource type and the third consumption sequence of each reference computing task in the target resource type, determine the task similarity between the current computing task and each reference computing task;

[0096] S230, Based on the task similarity between the current computing task and each reference computing task, determine the prediction reference weight of each reference computing task;

[0097] S240, Based on the predicted reference weights of each reference computing task and the complete consumption sequence of each reference computing task, the target model is trained to obtain the consumption prediction model.

[0098] In this embodiment, reference computing tasks of the same computing type as the current computing task are obtained. This aims to identify historical computing tasks with similar resource demand patterns or business logic to the current task. This ensures that the training data used to predict the future resource consumption of the current task is highly correlated with the characteristics of the current task, thereby improving the accuracy of the prediction. This process can be achieved by analyzing task metadata (e.g., task tags, business types, application IDs, etc.) for matching, or by performing cluster analysis on the resource consumption patterns of historical computing tasks to categorize the current task into a certain cluster. The historical computing tasks within this cluster are the reference computing tasks.

[0099] Based on the second consumption sequence of the current computing task in the target resource type, and the third consumption sequence of each reference computing task in the target resource type, the task similarity between the current computing task and each reference computing task is determined. This serves to quantify the degree of similarity between the current computing task and the reference computing tasks in terms of resource consumption behavior. This provides a basis for subsequently determining the prediction reference weights of the reference computing tasks, ensuring that reference computing tasks with high similarity play a greater role in model training. This similarity can be measured by calculating the Euclidean distance, Manhattan distance, Chebyshev distance, or Dynamic Time Warping (DTW) distance between two consumption sequences; the smaller the distance, the higher the similarity. Alternatively, it can be measured by calculating cosine similarity or Pearson correlation coefficient; the larger the value, the higher the similarity.

[0100] It should be noted that the second and third consumption sequences mentioned above are consistent with the first consumption sequence in terms of data structure, collection dimensions, calculation methods, and time-series representation. They are all resource consumption sequences constructed according to unified rules. The core difference between the various consumption sequences lies only in the different corresponding task objects and / or task times. The first consumption sequence is the consumption sequence of the target resource type of the current computing task at the current moment, the second consumption sequence is the consumption sequence of the target resource type of the current computing task at a historical moment, and the third consumption sequence is the consumption sequence of the target resource type of the reference computing task at the same historical moment as the second consumption sequence.

[0101] Based on the task similarity between the current computation task and each reference computation task, prediction reference weights are determined for each reference computation task. The purpose is to assign a weight value to each reference computation task, representing its importance in training the predictive model. This gives reference computation tasks that are more similar to the current computation task a greater influence in model training, thereby improving the model's prediction accuracy. These weights can be directly derived from the task similarity value, converted into weights using a normalization function (e.g., the Softmax function), or determined comprehensively based on the similarity value combined with other factors (e.g., the completion status of the reference computation task, data quality, etc.).

[0102] Based on the predicted reference weights of each reference computing task and the complete consumption sequence of each reference computing task, a consumption prediction model is trained on the target model. Its purpose is to use weighted historical consumption data to train a model capable of predicting future resource consumption. This generates a consumption prediction model that accurately predicts the future resource consumption of the current computing task, providing a reliable basis for subsequent resource reservation. This training process can employ machine learning algorithms such as weighted linear regression, weighted support vector machines, and weighted neural networks, or utilize time series prediction models, adjusting the loss function or sample contribution during the training process according to the weights.

[0103] The proposed solution improves prediction accuracy by specifically training the consumption prediction model before performing resource demand prediction. Specifically, the method first obtains reference computing tasks of the same type as the current computing task, ensuring a high correlation between the historical data used for model training and the characteristics of the current task. Then, by comparing the second consumption sequence of the current task with the third consumption sequences of each reference computing task, the task similarity between the current and reference computing tasks is quantified and determined. This step aims to identify historical computing tasks with resource consumption patterns most similar to the current task. Based on this, prediction reference weights are assigned to each reference computing task according to the determined task similarity, giving reference computing tasks with higher similarity to the current task a greater influence in model training. Finally, the complete consumption sequences of these reference computing tasks with prediction reference weights are used to train the target model, resulting in a highly customized consumption prediction model with stronger predictive capabilities. In this way, the resulting consumption prediction model can more accurately capture the resource consumption patterns of the current computing task, providing a more reliable basis for determining the amount of reserved resources for the target resource type, thereby optimizing the virtual aggregation effect of idle computing power fragments and avoiding resource waste or shortage due to inaccurate prediction.

[0104] Through the above technical solution, this application effectively addresses the problem of insufficient accuracy in consumption prediction models. By acquiring reference computing tasks of the same type as the current computing task and determining prediction reference weights based on task similarity, the consumption prediction model can be trained using historical data that is highly relevant to the current computing task and has undergone weight optimization. This significantly improves the prediction accuracy and robustness of the consumption prediction model, thereby making the reserved resource amount for the target resource type more reasonable.

[0105] In some embodiments described above in this application, a reference computing task of the same computing type as the current computing task is proposed for training a consumption prediction model. However, in actual implementation, how to efficiently and accurately identify a reference computing task of the same computing type as the current computing task to ensure the effectiveness of the selected reference computing task and the accuracy of the prediction model is a problem that needs to be solved.

[0106] In this regard, this application further proposes that S210 includes:

[0107] Based on the resource consumption differences among the fourth consumption sequence of target resource types for each historical computing task, each historical computing task is clustered to obtain at least one computing type cluster.

[0108] Based on the resource consumption difference between the current computing task and the target resource type second consumption sequence and the historical computing tasks in the computing type cluster in the target resource type fourth consumption sequence, the task consistency between the current computing task and the computing type cluster is determined.

[0109] Each historical computing task in the computing type cluster with the highest task consistency is identified as a reference computing task belonging to the same computing type as the current computing task.

[0110] In this embodiment, based on the resource consumption differences among the fourth consumption sequences of different historical computing tasks for the target resource type, the historical computing tasks are clustered to obtain at least one computing type cluster. This step aims to group tasks with similar consumption characteristics together by analyzing the resource consumption patterns of historical computing tasks on a specific target resource type, forming different computing type clusters. The purpose of this is to more effectively find reference computing tasks that match the current computing task's computing type, avoiding blind searching in massive amounts of historical data and improving matching efficiency and accuracy. A K-means clustering algorithm can be used. First, a distance metric (e.g., DTW distance) is defined to quantify the resource consumption differences among the fourth consumption sequences of different historical tasks. Then, an appropriate number of clusters K is selected; for example, the elbow method can be used to determine the value of the number of clusters K. Through iterative optimization, historical tasks are assigned to K clusters, resulting in high similarity among tasks within clusters and low similarity among tasks between clusters.

[0111] Based on the resource consumption difference between the current computing task's second consumption sequence in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type within the computing type cluster (e.g., the DTW distance between sequences can be used to measure the corresponding resource consumption difference), the task consistency between the current computing task and the computing type cluster is determined. This step aims to quantify the degree of matching between the current computing task and the established computing type clusters. By comparing the resource consumption sequence of the current computing task with the resource consumption sequences of historical computing tasks within the cluster, the most likely computing type to which the current computing task belongs can be assessed, thus providing a basis for subsequently selecting the most relevant reference computing task. The resource consumption difference between the second consumption sequence of the current computing task and the fourth consumption sequence of each historical computing task in the computing type cluster can be calculated. Then, all difference values ​​within the cluster are statistically processed, and this statistical result is used as the task consistency between the current computing task and the computing type cluster. The smaller the difference, the higher the consistency.

[0112] It should be noted that the fourth consumption sequence mentioned above is consistent with the second consumption sequence in terms of data structure, collection dimension, calculation method, and time series representation. Both are resource consumption sequences constructed according to unified rules. The core difference between the various consumption sequences lies only in the different task objects they correspond to. The second consumption sequence is the consumption sequence of the target resource type of the current computing task at a historical moment, while the fourth consumption sequence is the consumption sequence of the target resource type of a historical computing task at the same historical moment as the current computing task.

[0113] Each historical computational task in the computational type cluster with the highest task consistency is identified as a reference computational task belonging to the same computational type as the current computational task. This step selects the cluster that best matches the current computational task from all computational type clusters based on the task consistency determined in the previous step. All historical computational tasks in this cluster are considered to have the same computational type as the current task, thus serving as reference computational tasks for subsequent prediction model training. This ensures a high degree of relevance of the reference computational tasks and improves the accuracy of the prediction model. After calculating the task consistency between the current computational task and each computational type cluster, the cluster with the highest consistency value is directly selected. All historical computational tasks within this cluster are then selected as reference computational tasks.

[0114] In the above method, to accurately obtain reference computing tasks belonging to the same computing type as the current computing task, this application first analyzes a large number of historical computing tasks. Specifically, by evaluating the resource consumption differences between the fourth consumption sequences of each historical computing task on the target resource type, these historical computing tasks are systematically clustered to form several computing type clusters with similar resource consumption patterns. This process effectively structures and classifies the complex historical task data, laying the foundation for subsequent matching. Subsequently, for the current computing task to be processed, the system uses its second consumption sequence on the target resource type to compare with the previously formed computing type clusters. This comparison is not a simple one-to-one matching, but rather quantifies the task consistency between the current computing task and each cluster by calculating the resource consumption differences between the current computing task and the historical computing tasks in each computing type cluster. The higher the task consistency, the more closely the current computing task matches the computing type represented by that cluster. Finally, the system identifies the computing type cluster with the highest task consistency with the current computing task. All historical computing tasks in this cluster are precisely identified as reference computing tasks belonging to the same computing type as the current task due to their high similarity to the current task in resource consumption patterns. These selected reference tasks, compared to tasks randomly or coarsely selected from all historical tasks, more accurately reflect the potential behavioral patterns of the current task, thus providing high-quality input data for the subsequent training of the consumption prediction model. This refined reference task selection mechanism significantly improves the training effect and prediction accuracy of the consumption prediction model, thereby optimizing the efficiency of virtual aggregation management of idle computing power fragments in the computing cluster.

[0115] Through the above technical solution, this application effectively addresses the problem of efficiently and accurately identifying reference computing tasks with the same computing type as the current computing task when acquiring reference computing tasks. By clustering historical computing tasks and determining task consistency based on resource consumption differences, this application can accurately filter out reference computing tasks highly related to the current computing task from massive historical data. This refined reference task selection mechanism avoids prediction biases that may result from blindly or coarsely selecting reference computing tasks, significantly improving the training quality and prediction accuracy of the consumption prediction model. Therefore, it can provide a more reliable basis for resource reservation in computing clusters, thereby optimizing the virtual aggregation management of idle computing power fragments and improving resource utilization and system stability.

[0116] In some embodiments described above in this application, in order to train the resource consumption prediction model, it is necessary to obtain reference computing tasks belonging to the same computing type as the current computing task. This typically involves clustering historical computing tasks and determining the task consistency based on the resource consumption differences between the current computing task and each historical computing task in the computing type cluster. However, directly using the original resource consumption differences for comparison may not accurately reflect the true consistency between tasks, especially when there are significant differences in the numerical range or fluctuation range of the resource consumption sequences of different tasks. This may lead to bias in the task consistency assessment, thereby affecting the accuracy of the selection of reference computing tasks.

[0117] In response, this application further proposes determining the task consistency between the current computing task and the computing type cluster based on the resource consumption difference between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type within the computing type cluster. This includes:

[0118] The resource consumption differences between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type in the computing type cluster are negatively correlated and normalized to obtain at least one consistency evaluation value.

[0119] The consistency evaluation values ​​are averaged to obtain the task consistency degree between the current computing task and the computing type cluster.

[0120] In this embodiment, the goal is to convert the original resource consumption differences into a standardized and comparable indicator by performing negative correlation normalization on the resource consumption differences between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type within the computing type cluster, respectively. The purpose of negative correlation normalization is to ensure that the smaller the difference (i.e., the higher the similarity), the larger the obtained consistency evaluation value. For example, the difference value can be mapped to the interval between 0 and 1 by taking the reciprocal of the difference value and then linearly scaling it (e.g., Min-Max normalization), where 0 represents the largest difference and 1 represents the smallest difference.

[0121] The goal of "averaging all consistency evaluation values ​​to obtain the task consistency degree between the current computation task and the computation type cluster" is to comprehensively evaluate the consistency evaluation values ​​between all historical computation tasks and the current computation task within a computation type cluster, thereby obtaining a task consistency degree representative of the entire cluster. Averaging is a commonly used aggregation method that effectively smooths out individual differences, providing a stable and representative overall evaluation. For example, the arithmetic mean of all consistency evaluation values ​​can be calculated.

[0122] This application's solution performs negative correlation normalization on the resource consumption differences between the current computing task and historical computing tasks within the computing type cluster, transforming the original difference values, which may have different dimensions and numerical ranges, into consistency evaluation values ​​on a uniform scale. This transformation ensures that the smaller the difference, the higher the consistency evaluation value, thus accurately reflecting the similarity between tasks. Subsequently, by averaging these standardized consistency evaluation values, this solution can comprehensively evaluate the matching degree between the current computing task and the entire computing type cluster, obtaining a stable and representative task consistency degree. This approach avoids evaluation bias caused by single comparisons or excessive fluctuations in the original difference values, making the calculation of task consistency degree more accurate and robust. In this way, the computing type cluster that best matches the current computing task's computing type can be identified more accurately, thereby selecting high-quality reference computing tasks and providing a more reliable data foundation for the training of subsequent consumption prediction models.

[0123] As an example, the task consistency between the current computation task and the computation type cluster can be determined using the following formula:

[0124]

[0125] In the formula, Used to characterize the task consistency between the current computation task d and the o-th computation type cluster. This is used to characterize the difference in resource consumption between the current computing task d and the r-th historical computing task in the o-th computing type cluster, which is also the DTW distance between the second consumption sequence corresponding to the current computing task d and the fourth consumption sequence corresponding to the r-th historical computing task in the o-th computing type cluster. Used to characterize negative correlation normalization processing Used to characterize the total number of historical computing tasks in the o-th computing type cluster.

[0126] Here, negative correlation normalization refers to mapping the input resource consumption differences to standardized values ​​that meet the weight distribution requirements. Specifically, it involves first processing the input items... Taking the reciprocal or negative value allows historical computation tasks with smaller differences in resource consumption to obtain larger intermediate results. Then, using methods such as min-max normalization or the Softmax function, the intermediate results are scaled to the [0,1] interval to finally obtain the task consistency. It should be understood that the minimum and maximum values ​​of this minimum-maximum normalization method are determined based on the corresponding calculation processes of all historical computation tasks.

[0127] Through the aforementioned technical solution, this application can more accurately and robustly assess the task consistency between the current computing task and different computing type clusters. This precise assessment helps to identify reference computing tasks that are most similar to the current task's behavior pattern, thereby providing high-quality input data for training the consumption prediction model. Ultimately, this enables the consumption prediction model to more accurately predict the future demand for target resource types, thereby optimizing the determination of reserved resource amounts and improving the overall efficiency and resource utilization of virtual aggregation management of idle computing power fragments in the computing cluster.

[0128] In some of the embodiments described above in this application, when determining the task similarity between the current computing task and each reference computing task, simply comparing the resource consumption sequence may not be sufficient to capture the deep-seated correlation between tasks, resulting in inaccurate task similarity assessment, which in turn affects the training effect of the prediction model and the final aggregation scheme.

[0129] In this regard, this application further proposes that S220 includes:

[0130] The second consumption sequence of the current computing task in the target resource type is compared with the third consumption sequence of each second reference computing task in the target resource type to construct the first computing preference vector of the current computing task in the target computing type; the target computing type is the computing type to which the current computing task belongs.

[0131] The third consumption sequence of the first reference computing task in the target resource type is compared with the third consumption sequence of each second reference computing task in the target resource type to construct a second computing preference vector of the first reference computing task in the target computing type; the first reference computing task is any reference computing task, and the second reference computing tasks are reference computing tasks other than the first reference computing task.

[0132] The cosine similarity between the first computational preference vector and the second computational preference vector is determined as the task similarity between the current computational task and the first reference computational task.

[0133] In this embodiment, the step of constructing the first computational preference vector aims to quantify the resource consumption characteristics of the current computational task under a specific target computational type. By comparing the second consumption sequence of the current computational task in the target resource type with the third consumption sequences of each second reference computational task in the target resource type, the resource usage tendency of the current task relative to other reference tasks can be identified. For example, using the second consumption sequence of the current computational task as a benchmark, the DTW distance between it and the third consumption sequence of each second reference computational task is calculated, and the reciprocal of these distances or the values ​​transformed by a specific function are used as elements of the preference vector; the smaller the distance, the larger the preference value.

[0134] The step of constructing the second computational preference vector is used to characterize the relationships between reference computational tasks, providing a reference benchmark for subsequent task similarity calculations. By comparing the third consumption sequence of the first reference computational task in the target resource type with the third consumption sequences of all other second reference computational tasks in the target resource type, the resource usage preference of the first reference computational task relative to the other reference tasks can be constructed. For example, a method similar to constructing the first computational preference vector can be used to calculate the resource consumption differences or similarities between the first reference task and all other second reference tasks, forming a vector.

[0135] Cosine similarity is a metric that measures the cosine of the angle between two non-zero vectors, commonly used to evaluate text similarity, user preference similarity, and other scenarios. Its value ranges from -1 to 1; a higher value indicates that the two vectors are closer in direction, i.e., the higher the similarity. In this application, by calculating the cosine similarity between the first and second computational preference vectors, the task similarity between the current computational task and the first reference computational task can be effectively quantified. This method focuses on the direction of the vectors rather than their absolute magnitude, thereby capturing the similarity of tasks in terms of resource consumption patterns, even if their absolute consumption amounts differ.

[0136] This application's solution accurately assesses task similarity by introducing a computational preference vector and cosine similarity. First, for the current computational task, its second consumption sequence in the target resource type is compared one-to-one with the third consumption sequences in the target resource type of all second reference computational tasks (i.e., all other reference computational tasks besides the first reference computational task), thereby constructing a first computational preference vector for the current computational task. This vector reflects the relative tendency of the current computational task in resource consumption patterns compared to each of the reference computational tasks. Subsequently, to establish an objective reference frame, this application further compares the third consumption sequence in the target resource type of each first reference computational task with the third consumption sequences of all second reference computational tasks except itself, thereby constructing a second computational preference vector for that first reference computational task. This vector characterizes the relative relationship between the first reference computational task and other reference computational tasks in resource consumption patterns. Finally, by calculating the cosine similarity between the first computational preference vector of the current computational task and the second computational preference vector of a specific first reference computational task, the task similarity between the current computational task and that first reference computational task can be obtained. This method effectively captures the deep-seated similarities in resource consumption patterns between tasks, rather than simply comparing numerical values, thus providing a more accurate basis for determining subsequent prediction reference weights. This refined similarity assessment ensures that the consumption prediction model can more accurately learn historical patterns related to the current computing task during training, thereby improving the prediction accuracy of future demand for target resource types and providing more reliable decision support for the virtual aggregation management of idle computing power fragments in computing clusters.

[0137] The following is a concrete example to illustrate this. Assume the current computational task is "image rendering task A," the target resource type is "GPU computing power," and a set of reference computational tasks has been obtained, such as "image rendering task B," "video encoding task C," and "scientific computing task D." When determining the task similarity between "image rendering task A" and "image rendering task B," the system first obtains the second consumption sequence of "image rendering task A" on GPU computing power, and the third consumption sequence of the reference computational tasks such as "image rendering task B," "video encoding task C," and "scientific computing task D" on GPU computing power. Next, the second consumption sequence of "image rendering task A" is compared with the third consumption sequences of "video encoding task C" and "scientific computing task D" (excluding "image rendering task B"). For example, the DTW distance between them can be calculated, and the reciprocals of these distances are used as elements of the first computational preference vector of "image rendering task A." For example, if the distance between "image rendering task A" and "video encoding task C" is d2, and the distance between "image rendering task A" and "scientific computing task D" is d3, then the first computational preference vector can be represented as [1 / d2, 1 / d3]. Simultaneously, to construct the second computational preference vector for "image rendering task B," the system compares the third consumption sequence of "image rendering task B" with the third consumption sequences of reference computing tasks other than "image rendering task B" (i.e., "video encoding task C" and "scientific computing task D"). For example, calculating the distance d4 between "image rendering task B" and "video encoding task C," and the distance d5 between "image rendering task B" and "scientific computing task D," then the second computational preference vector for "image rendering task B" can be represented as [1 / d4, 1 / d5]. Finally, the cosine similarity between the first computational preference vector [1 / d2, 1 / d3] of "image rendering task A" and the second computational preference vector [1 / d4, 1 / d5] of "image rendering task B" is calculated.

[0138] Through the above technical solution, this application overcomes the limitations of traditional methods in assessing task similarity. By constructing a first computational preference vector and a second computational preference vector, and using cosine similarity for calculation, the inherent correlation and similarity in resource consumption patterns between the current computational task and the reference computational task can be more deeply explored. This preference vector-based similarity assessment method can effectively avoid the problem of inaccurate similarity judgment caused by large differences in the absolute resource consumption of tasks, thus providing a more accurate measure of task similarity. This enables a more accurate identification of the reference computational task most valuable for predicting the current computational task when determining the prediction reference weights, thereby improving the training quality and prediction accuracy of the consumption prediction model, providing more reliable data support for the virtual aggregation management of idle computing power fragments in the computing cluster, and ultimately optimizing resource utilization efficiency and system performance.

[0139] In some embodiments described above in this application, a consumption prediction model is trained by determining task similarity. However, in its implementation, how to accurately quantify the reference value of the reference computing task for predicting the future resource consumption of the current computing task, so as to ensure the accuracy and robustness of the prediction model, is a problem that needs to be solved.

[0140] In this regard, this application further proposes that S230 includes:

[0141] The difference in resource consumption between the current computing task in the second consumption sequence of the target resource type and the reference computing task in the third consumption sequence of the target resource type is divided by the task similarity between the current computing task and the reference computing task to determine the reference evaluation value of the reference computing task.

[0142] The reference evaluation values ​​are negatively correlated and normalized to obtain the predicted reference weights for the reference calculation task.

[0143] In this embodiment, the step of "dividing the resource consumption difference between the current computing task's second consumption sequence in the target resource type and the reference computing task's third consumption sequence in the target resource type by the task similarity between the current computing task and the reference computing task to determine the reference evaluation value of the reference computing task" aims to comprehensively consider the differences in resource consumption between the reference computing task and the current computing task, as well as their degree of similarity, to quantify the potential reference value of the reference computing task. The resource consumption difference can be calculated in various ways, such as the DTW distance between the two consumption sequences over time. Task similarity characterizes the degree of closeness between two tasks in terms of behavioral patterns or features, and can be obtained, for example, through methods such as cosine similarity. By combining the resource consumption difference with task similarity, the predictive reference value of the reference task can be evaluated more comprehensively.

[0144] The step of "performing negative correlation normalization on the reference evaluation values ​​to obtain the predicted reference weights for the reference computation task" aims to convert the original reference evaluation values ​​into predicted reference weights with specific ranges and meanings, facilitating unified processing in subsequent model training. Negative correlation normalization means that the larger the reference evaluation value (i.e., the greater the difference in resource consumption or the lower the task similarity), the smaller the predicted reference weight, and vice versa. This approach ensures that reference tasks more similar to the current task and with smaller differences in resource consumption receive higher weights, thus playing a greater role in predictive model training.

[0145] This application's solution addresses the problem of accurately quantifying the predictive reference value of a reference computing task by comprehensively considering resource consumption differences and task similarity. First, by calculating the resource consumption difference between the second consumption sequence of the current computing task and the third consumption sequence of the reference computing task, the degree of deviation in their resource usage patterns can be intuitively reflected. Simultaneously, combined with the established task similarity, the reference task can be evaluated from two dimensions: behavioral patterns and resource consumption. Dividing the resource consumption difference by the task similarity aims to construct a comprehensive reference evaluation value, which effectively reflects the overall quality of the reference task as a predictive basis: the smaller the difference and the higher the similarity, the lower the evaluation value, indicating a higher predictive value for the reference task. Subsequently, this reference evaluation value is negatively correlated and normalized, transforming the original evaluation value into a predictive reference weight with a uniform scale. This negative correlation ensures that reference tasks with lower evaluation values ​​(i.e., reference tasks with higher predictive value) receive higher predictive reference weights. In this way, when training the consumption prediction model, reference tasks that are closer to the current task in terms of resource consumption patterns and have more similar behavior patterns will have a greater influence. This will enable the trained consumption prediction model to more accurately capture the future resource requirements of the current computing task, thereby improving the accuracy of the amount of reserved resources for the target resource type and optimizing the resource management efficiency of the computing cluster.

[0146] As an example, the prediction reference weights for a computational task can be determined using the following formula:

[0147]

[0148] In the formula, Used to characterize the prediction reference weight of the reference computation task p in the target resource type l. Used to characterize the difference in resource consumption between the current computing task d and the reference computing task p in target resource type l. Used to characterize the task similarity between the current computation task d and the reference computation task p. Used to characterize negative correlation normalization.

[0149] It should be noted that, to ensure the calculation results are meaningful, in this embodiment of the invention, when performing fractional operations, if the denominator is 0, a parameter adjustment factor greater than 0 needs to be added to the denominator before summing to prevent the denominator from being 0. The value of the parameter adjustment factor can be set by the implementer according to the actual situation, and this invention does not impose any special restrictions. For example, the value of the parameter adjustment factor can be set to 0.01.

[0150] Here, negative correlation normalization refers to mapping the ratio of input resource consumption differences to task similarity to a standardized value that meets the weight distribution requirements. Specifically, this is achieved by first processing the input items... Taking the reciprocal or negative value results in a larger intermediate result for a reference computation task with smaller resource consumption differences and higher task similarity. Then, the intermediate result is scaled to the [0,1] interval using methods such as min-max normalization or the Softmax function to finally obtain the prediction reference weight. It should be understood that the minimum and maximum values ​​of this minimum-maximum normalization method are determined based on the corresponding computational processes of all reference computational tasks.

[0151] Through the aforementioned technical solution, this application can more precisely evaluate the reference value of reference computing tasks in predicting the future resource consumption of the current computing task. By comprehensively considering resource consumption differences and task similarity, it avoids the bias that may result from evaluating based on a single dimension, allowing reference computing tasks that are more closely matched to the current task in terms of resource usage patterns and behavioral characteristics to receive higher predictive reference weights. This weighting mechanism enables the consumption prediction model to more effectively utilize high-quality reference data during training, thereby significantly improving the accuracy of predicting future demand for the target resource type. Ultimately, this helps to more rationally determine the amount of reserved resources, reduce the risk of resource waste or shortage, and thus optimize the overall resource utilization and task scheduling efficiency of the computing cluster.

[0152] In some embodiments described above, this application proposes inputting the first consumption sequence of the target resource type for the current computing task at the current moment into a consumption prediction model to predict the future demand for the target resource type and determine the amount of reserved resources for that target resource type. However, in practice, how to accurately and effectively determine the amount of reserved resources from the predicted future demand to avoid resource waste or shortage is a technical problem that needs further clarification and optimization.

[0153] In this regard, this application further proposes that S110 includes:

[0154] Input the first consumption sequence of the target resource type of the current computing task at the current time into the consumption prediction model to obtain the predicted consumption sequence of the target resource type of the current computing task;

[0155] Based on the maximum predicted consumption value in the predicted consumption sequence, determine the reserved resource amount for the target resource type.

[0156] In this embodiment, the first consumption sequence of the current computing task for the target resource type at the current moment is input into the consumption prediction model to obtain the predicted consumption sequence of the current computing task for the target resource type. This step aims to infer the changing trend of the task's demand for the target resource type over a future period based on the real-time consumption data of the current task using a trained consumption prediction model. The consumption prediction model can be a time series analysis-based model, such as an autoregressive moving average model, which receives the first consumption sequence as input and outputs a sequence containing predicted consumption values ​​for multiple future time points.

[0157] Based on the maximum predicted consumption value in the predicted consumption sequence, the reserved resource amount for the target resource type is determined. This step aims to extract a representative value from the predicted future consumption sequence as the reserved resource amount to ensure sufficient resources are available during future task execution. Choosing the maximum value is to address potential peak demands at some future time, effectively avoiding resource shortages. The system can iterate through all predicted values ​​in the predicted consumption sequence and select the largest value as the reserved resource amount. For example, if the predicted sequence is [10GB, 12GB, 8GB, 15GB, 11GB], the reserved resource amount is determined to be 15GB. In some scenarios, to increase robustness, an additional safety margin, such as 20% of the maximum predicted consumption value, can be added to handle prediction errors or unexpected situations, ultimately determining the reserved resource amount.

[0158] The proposed solution first obtains a predicted consumption sequence of the target resource type for the current computing task at the current moment by inputting it into a consumption prediction model. This predicted consumption sequence reflects the dynamic changes in the task's future resource requirements. Subsequently, to ensure that the task's resource requirements are met at any point during execution, the proposed solution further determines the reserved resource amount for the target resource type based on the maximum predicted consumption value in the predicted consumption sequence. This approach effectively addresses the volatility of task resource requirements, especially when a peak in resource demand is predicted. By reserving the maximum value, task interruptions or performance degradation due to insufficient resources can be effectively avoided. Compared to reserving based solely on average or current values, using the maximum predicted consumption value as the reservation basis provides a more robust guarantee of resource supply for the task, thereby improving the stability of the computing cluster and the reliability of task execution.

[0159] As an example, the amount of reserved resources for a target resource type can be determined using the following formula:

[0160]

[0161] In the formula, The amount of reserved resources used to characterize target resource type l; This is used to characterize the safety margin coefficient, which is used to prevent resource overruns. The default value is 0.2. Used to characterize the predicted consumption of the current computation task d at time t. This is used to characterize the maximum value of the predicted consumption of the current computation task d at each time step.

[0162] Through the above technical solution, this application can determine the amount of reserved resources for a target resource type in a more robust and reliable manner from the predicted future resource demand sequence. By selecting the maximum value in the predicted consumption sequence as the reservation basis, it can effectively cope with potential peak resource demand in the future computing task, thereby avoiding task performance degradation or interruption due to insufficient resource reservation. At the same time, this method also avoids blindly over-reserving resources, improving resource utilization efficiency while ensuring stable task operation.

[0163] In some embodiments described above in this application, a resource availability index for each first resource unit in the target resource type is determined based on the distance between each first resource unit and each second resource unit. However, in practice, how to comprehensively consider the distance relationship between a first resource unit and multiple second resource units and transform it into a resource availability index that can effectively guide subsequent aggregation decisions is a problem that needs to be solved.

[0164] In this regard, this application further proposes that S120 includes:

[0165] The average distance is obtained by averaging the distances between the first resource unit and each of the second resource units.

[0166] By performing positive correlation normalization on the distance mean, the resource availability index of the first resource unit in the target resource type is obtained.

[0167] In this embodiment, the distances between the first resource unit and each of the second resource units are summed to obtain a cumulative distance value. This aims to quantify the overall "proximity" or "isolation" between a potential idle resource unit (the first resource unit) and all occupied resource units (the second resource units). The distance between the first resource unit and each of the second resource units can be understood as the absolute value of the hop count or hierarchical distance between them in the physical topology tree. Furthermore, the cumulative distance value between the first resource unit and all the second resource units is divided by the total number of all second resource units to obtain the average distance. By calculating the average distance, a comprehensive statistical index can be obtained, reflecting the overall spatial relationship between the first resource unit and the currently occupied resource units.

[0168] Positive correlation normalization is applied to the accumulated distance values ​​to obtain the resource availability index of the first resource unit within the target resource type. The purpose is to convert values ​​of different dimensions or ranges to a unified scale for easier comparison and subsequent calculations. As a concrete example, the ratio between the average distance corresponding to the first resource unit and the maximum possible hop count within the cluster (a preset system constant) can be used as the normalization result. Positive correlation normalization means that the larger the accumulated value, the larger the normalized index value, reflecting that the greater the overall distance between the first resource unit and occupied resource units, and the higher its availability.

[0169] In the virtual aggregation management of idle computing power fragments in a computing cluster, to accurately assess the resource idleness of each first resource unit, this application first calculates the average distance between the first resource unit and each second resource unit, thus obtaining the average distance value. This average distance value comprehensively reflects the average spatial distribution relationship between a potential idle resource unit (first resource unit) and all resource units (second resource units) currently occupied by computing tasks. A larger average distance value usually means that the first resource unit is physically or logically far from the occupied resource units, is less likely to be interfered with by the current task, and is more suitable for aggregation as an idle resource. Subsequently, to make the average distance value comparable and effectively guide subsequent resource selection, this application performs positive correlation normalization on the average distance value. Through this processing, the original average distance value is mapped to a unified, standardized interval, so that the larger the value, the higher the resource idle index, that is, the more "idle" or more suitable the first resource unit is for aggregation. This approach ensures that when determining idle resource units based on reserved resource quantity and resource idleness indicators, it can more accurately screen out those resource units that have the least conflict with the existing tasks' resource consumption and the greatest aggregation potential, thereby optimizing the identification process of idle resource units and laying the foundation for subsequent virtualization aggregation.

[0170] As an example, the resource availability index of the first resource unit for the target resource type can be determined using the following formula:

[0171]

[0172] In the formula, Used to characterize the The resource availability index of the first resource unit in target resource type l. Used to characterize the The topology number / hierarchical identifier of the second resource unit in the physical topology tree. Used to characterize the The topology number / hierarchical identifier of the first resource unit in the physical topology tree. Used to characterize the The second resource unit and the first The absolute value of the topological hop count or hierarchical distance of each first resource unit in the physical topology tree. M is used to characterize the total number of second resource units. To characterize positive correlation normalization, for example, a minimax normalization method can be used. It should be understood that the maximum and minimum values ​​of this normalization method are determined based on the same calculation process for all first resource units.

[0173] In order to improve computational efficiency, computational tasks generally need to schedule the nearest resource unit first, which can minimize data communication latency. Therefore, the smaller the distance between the first resource unit and each second resource unit, the smaller the resource idle index of the first resource unit in the target resource type.

[0174] Through the above technical solution, this application effectively addresses the problem of how to comprehensively consider the distance relationship between a first resource unit and multiple second resource units, and transform it into a resource idleness index that can effectively guide subsequent aggregation decisions. By accumulating the distances between the first resource unit and each second resource unit, and performing positive correlation normalization on the accumulated value, a comprehensive and standardized resource idleness index can be obtained. This index accurately reflects the overall spatial isolation degree between the first resource unit and the resource units currently occupied by the task. This allows for the priority selection of resource units with the least conflict with existing task resource occupation and the greatest aggregation potential when subsequently determining idle resource units. This improves the accuracy and efficiency of idle computing power fragment identification, provides a more reliable basis for subsequent virtualization aggregation, and ultimately optimizes the resource utilization and task scheduling performance of the computing cluster.

[0175] In some of the embodiments described above in this application, when determining the virtualization cost between each idle resource unit, directly considering multiple heterogeneous factors such as the size of occupied resources, the total number of resources, the duration of the task, and the distance between each idle resource unit may result in an inaccurate calculation of the virtualization cost, making it difficult to effectively distinguish between the internal management cost of a single resource unit and the communication cost between units, thereby affecting the optimization effect of the final aggregation scheme.

[0176] In this regard, this application further proposes that S140 includes:

[0177] The resource call cost of the idle resource unit is determined based on the occupied resource size, total resource quantity, and task duration of the idle resource unit in the target resource type.

[0178] The virtualization cost between the first idle resource unit and the second idle resource unit is determined based on the first resource call cost of the first idle resource unit, the second resource call cost of the second idle resource unit, and the distance between the first idle resource unit and the second idle resource unit.

[0179] In this embodiment, the resource call cost of an idle resource unit refers to the internal cost or overhead incurred by a single idle resource unit when it is virtualized and aggregated, due to its own resource characteristics (e.g., occupied resource size, total resource quantity, task duration). This resource call cost can be calculated using a preset function or model. Specifically, the resource call cost of an idle resource unit can be determined using the following formula:

[0180]

[0181] In the formula, Used to characterize the resource allocation cost of the a-th idle resource unit. Used to characterize the occupied resource size of the a-th free resource unit. Used to represent the total amount of resources in the a-th free resource unit. Used to characterize the duration of the task in the a-th idle resource unit. To characterize positive correlation normalization, for example, a minimax normalization method can be used. It should be understood that the maximum and minimum values ​​of this normalization method are determined based on the corresponding calculation process of all idle resource units.

[0182] For a virtual computing node, if some idle resource units are allocated to the virtual computing node, the current computing task will be completed soon afterward. At this time, these will form new computing power fragments, which will result in more resource fragments when building the virtual computing node. This will increase the cost of calling these idle resource units.

[0183] The first resource call cost of the first idle resource unit and the second resource call cost of the second idle resource unit refer to the resource call costs of each of the two idle resource units when calculating the virtualization cost between them. They are determined in the same way as the resource call costs of the idle resource units mentioned above, i.e., based on their respective occupied resource size, total resource quantity, and task duration.

[0184] The distance between the first and second idle resource units represents their proximity in terms of physical topology, network latency, or logical packets. Closer distances generally mean lower communication overhead and higher aggregation efficiency. Specifically, the distance between the first and second idle resource units can be considered as the number of topological hops between them in the physical topology tree.

[0185] The step of determining the virtualization cost between the first and second idle resource units aims to comprehensively consider the internal call costs of each of the two idle resource units and the communication costs between them, thereby obtaining a comprehensive virtualization cost for evaluating the feasibility of aggregation. This virtualization cost can be determined using the following formula:

[0186]

[0187] In the formula, Used to characterize the virtualization cost between the i-th first idle resource unit and the j-th second idle resource unit. The first resource call cost is used to characterize the i-th first idle resource unit. The second resource call cost is used to characterize the j-th second idle resource unit. Used to characterize the topological hop count in the physical topology tree between the i-th first free resource unit and the j-th second free resource unit. To characterize positive correlation normalization, for example, a minimax normalization method can be used. It should be understood that the maximum and minimum values ​​of this normalization method are determined based on the calculation process corresponding to the virtualization cost between all first and second idle resource units.

[0188] Among them, when multiple scattered idle resource units are merged to form a virtual computing node, if the resource call cost of two idle resource units is low and the communication cost between the two idle resource units is also low enough (the distance between them is small), it means that the two idle resource units can be merged into the same virtual computing node.

[0189] The proposed solution, when determining the virtualization cost between idle resource units, first abstracts the characteristics of each idle resource unit (such as the size of occupied resources, the total number of resources, and the duration of tasks) into a resource call cost for that unit. This step quantifies the internal complexity of a single resource unit into a unified index, allowing subsequent calculations to focus more on the interaction between units. Subsequently, when evaluating the virtualization cost between two idle resource units (e.g., the first idle resource unit and the second idle resource unit), the solution considers not only the individual resource call costs of these two units but also the crucial factor of the distance between them. By comprehensively considering the first resource call cost of the first idle resource unit, the second resource call cost of the second idle resource unit, and the distance between them, this application can more comprehensively and accurately assess the total cost required to merge these two idle resource units into a single virtual computing node. This step-by-step calculation method decouples the internal cost of a single unit from the communication cost between units, making the calculation of virtualization costs more refined and rational, thus providing a more reliable basis for subsequent virtualization aggregation decisions.

[0190] The above technical solution decomposes the virtualization cost of idle resource units into two parts: internal resource call cost and external communication cost, making the evaluation of virtualization costs more refined and accurate. First, by comprehensively considering the size of the idle resource unit's occupied resources, the total number of resources, and the duration of the task, the resource call cost is determined, more accurately reflecting the internal cost of virtualizing and maintaining a single unit. Second, when calculating the virtualization cost between two idle resource units, not only are their individual resource call costs considered, but the distance between them is also introduced, thus enabling a comprehensive evaluation of the total overhead required for aggregation, including internal management costs and external communication costs. This layered and refined cost calculation method avoids simply confusing heterogeneous factors, making the cost evaluation of virtualization aggregation between idle resource units more scientific and reasonable. This helps to select truly low-cost, high-efficiency aggregation schemes in subsequent aggregation processes, thereby improving the utilization rate of idle computing power fragments and overall performance of the computing cluster.

[0191] In some embodiments described above, virtualization aggregation based on the virtualization cost between idle resource units is proposed to obtain the optimal aggregation scheme for the target resource type. However, in practice, how to efficiently and accurately find a truly "optimal" aggregation scheme from numerous idle resource units based on complex virtualization cost relationships to minimize overall virtualization cost or maximize aggregation efficiency is a challenging problem. Simply calculating the virtualization cost does not directly provide the optimal aggregation strategy; a structured approach is needed to solve the many-to-many matching optimization problem.

[0192] In this regard, this application further proposes that S150 includes:

[0193] Each virtual computing node to be built is allocated an idle resource unit to obtain an initialized virtual computing node;

[0194] Using the initialized virtual computing node as the first node, the remaining idle resource units as the second node, and the negative correlation normalized value of the virtualization cost between the first node and the second node as the corresponding edge weight, the optimal aggregation scheme for the target resource type is determined by the Hungarian algorithm.

[0195] In this embodiment, the virtual computing nodes to be built refer to virtualized computing entities that need to be created to support future computing tasks. These nodes typically need to integrate multiple physically dispersed idle resource units to form a logically unified virtual resource pool with sufficient computing power. Their number can be preset according to the expected task load, resource requirements, or system configuration strategy. For example, the number of virtual computing nodes to be built can be dynamically determined based on the number or type of tasks currently pending.

[0196] Each virtual computing node to be built is allocated a free resource unit, aiming to provide an initial, basic physical resource support for each node. This allocation can employ various strategies; for example, a free resource unit can be randomly selected, or it can be initially screened and allocated based on factors such as the geographical location and resource quantity of the free resource unit. In this way, each virtual computing node to be built possesses a "seed" resource unit, serving as the starting point for its subsequent aggregation of other free resource units. After this allocation, each virtual computing node to be built has an initial free resource unit, thus forming an initialized virtual computing node. These initialized virtual computing nodes represent the initial state of the virtual aggregation process; they will serve as the "first node" in subsequent aggregation algorithms, matching and aggregating with other remaining free resource units.

[0197] In the application scenario of the Hungarian algorithm, the first node represents the party that needs to be matched. Here, the initial virtual computing nodes are designated as the first node, meaning they are the active or target party in the aggregation process, needing to select suitable partners from the remaining idle resource units for aggregation. The second node represents the other party that can be matched. Here, the remaining idle resource units are designated as the second node; they are the resource pool available for the initial virtual computing nodes to select and aggregate. This partitioning transforms the aggregation problem into a bipartite graph matching problem, where the first and second nodes constitute the two parts of the bipartite graph.

[0198] Virtualization cost represents the communication cost of merging idle resource units into the same virtual computing node, and it is generally desirable to minimize this cost. Negative correlation normalization of the original virtualization cost transforms the problem of "minimizing cost" into a problem of "maximizing weight." Negative correlation normalization ensures that lower costs correspond to higher weights, thus prioritizing low-cost aggregation paths in the algorithm. Normalization unifies costs of different dimensions or ranges to the same scale, preventing excessively high or low costs from unfairly impacting the matching results. In a bipartite graph, edge weights represent the "cost" or "benefit" of connecting two nodes (i.e., the first node and the second node). Here, it specifically quantifies the merits of initializing the virtual computing node and aggregating the remaining idle resource units. By converting the virtualization cost into edge weights, the Hungarian algorithm can find the optimal matching combination based on these weights.

[0199] This application's solution systematically determines the optimal aggregation scheme for a target resource type by transforming the complex problem of aggregating idle resource units into a bipartite graph matching problem solvable by the Hungarian algorithm. First, an initial idle resource unit is allocated to each virtual computing node to be constructed; these allocated idle resource units constitute the initial virtual computing nodes. These initial virtual computing nodes are considered "first nodes" in the subsequent matching process, representing virtual entities that need to be filled or expanded. Simultaneously, all unallocated idle resource units are considered "second nodes," which are physical resource fragments that can be selected and aggregated by the first nodes. To quantify the merits of aggregation between the first and second nodes, their virtualization costs are negatively correlated and normalized, and the resulting value is used as the weight of the edge connecting the first and second nodes. This negative correlation normalization ensures that the smaller the original virtualization cost, the larger the corresponding edge weight, thus transforming the cost minimization problem into a maximum weight matching problem that the Hungarian algorithm excels at solving. Subsequently, the constructed bipartite graph (containing the first node, the second node, and weighted edges) is input into the Hungarian algorithm. The Hungarian algorithm iteratively searches for augmenting paths to optimize the current match until an optimal combination of matches is found. This optimal combination of matches is the best aggregation scheme for the target resource type, ensuring that the overall virtualization cost is minimized among all possible aggregation methods, thus achieving efficient and economical virtualization aggregation of idle computing power fragments. In this way, this scheme overcomes the limitation of traditional aggregation methods in systematically finding the global optimum, ensuring the optimization of the aggregation result.

[0200] Through the above technical solution, this application effectively addresses the problem of systematically determining the optimal aggregation scheme from the complex virtualization cost relationships when virtually aggregating idle computing power fragments in a computing cluster. By treating the initial virtual computing nodes and remaining idle resource units as two parts of a bipartite graph, and using the negatively correlated normalized value of the virtualization cost as the edge weight, this scheme cleverly transforms the resource aggregation problem into a standard bipartite graph maximum weight matching problem. The application of the Hungarian algorithm ensures that among all possible aggregation combinations, an optimal match that minimizes the overall virtualization cost can be found precisely. This not only avoids the defects of traditional heuristic or greedy algorithms that may get trapped in local optima, but also significantly improves the efficiency of virtual computing node construction and resource utilization, thereby enabling the most optimized and economical integration and utilization of idle computing power fragments in the computing cluster.

[0201] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0202] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A virtual aggregation management method for idle computing power fragments in a computing cluster, characterized in that, The method includes: The first consumption sequence of the target resource type at the current moment of the current computing task is input into the consumption prediction model to predict the future demand of the target resource type and determine the reserved resource amount of the target resource type. Based on the distance between each first resource unit and each second resource unit, a resource idle index for each first resource unit in the target resource type is determined; the first resource unit is a resource unit with idle resources of the target resource type, and the second resource unit is a resource unit whose current computing task has used resources of the target resource type; Based on the reserved resource amount of the target resource type and the resource idle index of each first resource unit in the target resource type, the idle resource units of the target resource type are determined; Based on the occupied resource size, total resource quantity, task duration, and distance between each idle resource unit in the target resource type, the virtualization cost between each idle resource unit is determined; the virtualization cost is used to characterize the communication cost of merging the idle resource units into the same virtual computing node; Based on the virtualization cost between each of the idle resource units, virtualization aggregation is performed among the idle resource units to obtain the optimal aggregation scheme for the target resource type.

2. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 1, characterized in that, Before inputting the first consumption sequence of the target resource type of the current computing task at the current moment into the consumption prediction model to predict the future demand of the target resource type and determine the reserved resource amount of the target resource type, the method further includes: Obtain a reference computing task that belongs to the same computing type as the current computing task; Based on the second consumption sequence of the current computing task in the target resource type and the third consumption sequence of each of the reference computing tasks in the target resource type, the task similarity between the current computing task and each of the reference computing tasks is determined; Based on the task similarity between the current computing task and each of the reference computing tasks, the prediction reference weights of each of the reference computing tasks are determined. The consumption prediction model is obtained by training the target model based on the predicted reference weights of each of the reference computing tasks and the complete consumption sequence of each of the reference computing tasks.

3. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 2, characterized in that, The step of obtaining a reference computing task of the same computing type as the current computing task includes: Based on the resource consumption differences among the fourth consumption sequence of the target resource type for each historical computing task, each historical computing task is clustered to obtain at least one computing type cluster. Based on the resource consumption difference between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type in the computing type cluster, the task consistency between the current computing task and the computing type cluster is determined. Each historical computing task in the computing type cluster with the highest task consistency is identified as a reference computing task belonging to the same computing type as the current computing task.

4. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 3, characterized in that, The step of determining the task consistency between the current computing task and the computing type cluster based on the resource consumption difference between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type within the computing type cluster includes: The resource consumption differences between the second consumption sequence of the current computing task in the target resource type and the fourth consumption sequence of each historical computing task in the target resource type in the computing type cluster are respectively subjected to negative correlation normalization processing to obtain at least one consistency evaluation value. The consistency evaluation values ​​are averaged to obtain the task consistency degree between the current computing task and the computing type cluster.

5. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 2, characterized in that, The step of determining the task similarity between the current computing task and each of the reference computing tasks based on the second consumption sequence of the current computing task in the target resource type and the third consumption sequence of each of the reference computing tasks in the target resource type includes: The second consumption sequence of the current computing task in the target resource type is compared with the third consumption sequence of each second reference computing task in the target resource type to construct a first computing preference vector of the current computing task in the target computing type; the target computing type is the computing type to which the current computing task belongs. The third consumption sequence of the first reference computing task in the target resource type is compared with the third consumption sequence of each second reference computing task in the target resource type to construct a second computing preference vector of the first reference computing task in the target computing type; the first reference computing task is any one of the reference computing tasks, and the second reference computing tasks are the reference computing tasks other than the first reference computing task. The cosine similarity between the first computational preference vector and the second computational preference vector is determined as the task similarity between the current computational task and the first reference computational task.

6. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 2, characterized in that, The step of determining the prediction reference weights for each of the reference computing tasks based on the task similarity between the current computing task and each of the reference computing tasks includes: The resource consumption difference between the current computing task in the second consumption sequence of the target resource type and the reference computing task in the third consumption sequence of the target resource type is divided by the task similarity between the current computing task and the reference computing task to determine the reference evaluation value of the reference computing task. The reference evaluation value is subjected to negative correlation normalization to obtain the predicted reference weight of the reference calculation task.

7. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 1, characterized in that, The step of inputting the first consumption sequence of the target resource type of the current computing task at the current moment into the consumption prediction model, predicting the future demand of the target resource type, and determining the reserved resource amount of the target resource type includes: The first consumption sequence of the current computing task for the target resource type at the current time is input into the consumption prediction model to obtain the predicted consumption sequence of the current computing task for the target resource type. Based on the maximum predicted consumption value in the predicted consumption sequence, the reserved resource amount for the target resource type is determined.

8. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 1, characterized in that, The determination of the resource idle index of each first resource unit in the target resource type based on the distance between each first resource unit and each second resource unit includes: The average distance is obtained by averaging the distances between the first resource unit and each of the second resource units; The mean distance is positively correlated and normalized to obtain the resource availability index of the first resource unit in the target resource type.

9. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 1, characterized in that, The determination of the virtualization cost between each idle resource unit based on the occupied resource size, total resource quantity, task duration, and distance between each idle resource unit of the target resource type includes: Based on the size of the occupied resources, the total number of resources, and the duration of the task in the target resource type, the resource call cost of the idle resource unit is determined. Based on the first resource call cost of the first idle resource unit, the second resource call cost of the second idle resource unit, and the distance between the first idle resource unit and the second idle resource unit, the virtualization cost between the first idle resource unit and the second idle resource unit is determined.

10. The virtual aggregation management method for idle computing power fragments in a computing cluster according to claim 1, characterized in that, The step of virtualizing and aggregating the idle resource units based on the virtualization cost between each idle resource unit to obtain the optimal aggregation scheme for the target resource type includes: Each virtual computing node to be built is allocated one of the aforementioned idle resource units to obtain an initialized virtual computing node; Using the initialized virtual computing node as the first node, the remaining idle resource units as the second node, and the negative correlation normalized value of the virtualization cost between the first node and the second node as the corresponding edge weight, the optimal aggregation scheme for the target resource type is determined by the Hungarian algorithm.

Images