Distributed heterogeneous computing cluster scheduling method and device and electronic equipment

Abstract

The application provides a scheduling method and device of a distributed heterogeneous computing power cluster and electronic equipment, and relates to the technical field of cloud computing. The scheduling method comprises the following steps: acquiring task demand computing power through a trained computing power prediction model according to task request information carried by a task request; screening a plurality of computing power clusters in the universe according to the task request information and the task demand computing power to obtain one or more target computing power clusters; scoring the target computing power clusters according to a preset cluster scoring rule and obtaining a scoring result of the target computing power clusters; wherein the cluster scoring rule at least comprises one of scheduling priority scoring, network scoring or cost scoring; and screening an ultimate computing power cluster according to the scoring result of the one or more target computing power clusters and sending the task request to the ultimate computing power cluster. The application predicts the task demand computing power through the computing power prediction model and multi-dimensionally scores, thereby improving the matching accuracy of the ultimate computing power cluster and improving the resource utilization rate.

Description

Technical Field

[0001] This application relates to the field of cloud computing technology, and in particular to a scheduling method, apparatus and electronic device for a distributed heterogeneous computing power cluster. Background Technology

[0002] With the development of artificial intelligence, generative artificial intelligence, and large-scale pre-trained models, computing tasks are facing higher demands in terms of scale, complexity, and real-time performance. To meet this demand, more and more cloud service providers are building heterogeneous computing clusters with various computing architectures distributed across different geographical regions, driving the formation of a multi-cloud integrated computing service system. In this context, the scheduling of computing clusters not only relates to the efficient utilization of heterogeneous resources but also directly affects the execution performance of tasks and the service capabilities and response efficiency of the overall computing service system.

[0003] In existing technologies, the scheduling of computing power clusters typically employs a static matching strategy based on user-specified resource specifications. Specifically, when submitting a task request, the user specifies the required computing power type, computing power resource scale, or instance specifications. Based on the selected computing power specifications, a set of clusters with matching resource types is selected from the existing computing power clusters. Subsequently, based on information such as the computing power resource inventory and operating status of each cluster, the candidate clusters are further filtered to ultimately determine the target computing power cluster and schedule the task to that cluster for execution. This scheduling method can achieve a basic match between tasks and resources to some extent, but it relies heavily on resource parameters actively provided by the user and lacks analysis and dynamic prediction of the actual computing power requirements of the task itself.

[0004] Because the scheduling process often fails to accurately identify the actual computing power consumption of tasks, there is a mismatch between task resource allocation and actual needs. Therefore, there is an urgent need for a scheduling method for distributed heterogeneous computing power clusters to address the problem of low matching between scheduled computing resources and tasks. Summary of the Invention

[0005] This application provides a scheduling method, apparatus, and electronic device for a distributed heterogeneous computing power cluster, which solves the technical problem of low matching degree between scheduled computing power resources and tasks, resulting in resource waste or task execution failure.

[0006] Firstly, this application provides a scheduling method for a distributed heterogeneous computing power cluster, comprising:

[0007] Receive a task request, and based on the task request information carried in the task request, obtain the task requirement computing power corresponding to the task request through a trained computing power prediction model.

[0008] Based on the task request information and the computing power required by the task, multiple computing power clusters across the entire domain are filtered to obtain one or more target computing power clusters;

[0009] According to the preset cluster scoring rules, one or more target computing power clusters are scored, and the scoring results of one or more target computing power clusters are obtained; wherein, the cluster scoring rules include at least one of scheduling priority scoring, network scoring, or cost scoring;

[0010] Based on the scoring results of one or more of the target computing power clusters, the final computing power cluster is selected and the task request is sent to the final computing power cluster.

[0011] Optionally, in the method described above, the step of filtering multiple computing power clusters across the entire domain based on the task request information and the computing power required by the task to obtain one or more target computing power clusters includes:

[0012] Based on the task request information, obtain the request computing power type of the task request, and based on the constructed computing power cluster profile, obtain the cluster computing power type of multiple computing power clusters in the entire domain.

[0013] Based on the cluster computing power type of multiple computing power clusters in the entire domain, a computing power cluster that matches the requested computing power type is selected from the multiple computing power clusters in the entire domain as the first target computing power cluster, so as to obtain multiple first target computing power clusters;

[0014] Monitor the status information of multiple first target computing power clusters and obtain the current available computing power of multiple first target computing power clusters;

[0015] Based on the current available computing power of multiple first target computing power clusters, select a first target computing power cluster whose current available computing power is greater than or equal to the computing power required by the task from the multiple first target computing power clusters, and use it as a second target computing power cluster, so as to obtain one or more second target computing power clusters.

[0016] Optionally, in the method described above, scoring one or more target computing power clusters according to preset cluster scoring rules includes:

[0017] Based on the constructed computing power cluster profile, the scheduling weight of the second target computing power cluster is obtained, and the scheduling weight is recorded as the scheduling priority score of the second target computing power cluster, so as to obtain the scheduling priority score of one or more second target computing power clusters.

[0018] Optionally, in the method described above, scoring one or more target computing power clusters according to preset cluster scoring rules includes:

[0019] Based on the task request information, extract the first location information corresponding to the task request, and based on the constructed computing power cluster profile, obtain the second location information of one or more second target computing power clusters;

[0020] Based on the first location information and one or more second location information, the actual distance between the task request and each second target computing power cluster is calculated and obtained, and the extreme value method is used to normalize one or more of the actual distances to obtain the normalized distance value between the task request and each second target computing power cluster.

[0021] Monitor and acquire network latency information of one or more second target computing power clusters, and use the extreme value method to normalize one or more network latency information to obtain the normalized network latency value of each second target computing power cluster;

[0022] Based on the normalized distance value and normalized network latency value belonging to the same second target computing power cluster, the network score of one or more second target computing power clusters is calculated and obtained by weighted summation through a preset weight threshold.

[0023] Optionally, in the method described above, scoring one or more target computing power clusters according to preset cluster scoring rules includes:

[0024] Monitor and obtain the current available computing power of one or more second target computing power clusters, and calculate and obtain the resource occupancy ratio corresponding to the task request being scheduled to each second target computing power cluster based on the task requirement computing power and the current available computing power;

[0025] Based on the constructed computing power cluster profile, obtain the unit computing power price information of one or more second target computing power clusters, and calculate the execution cost corresponding to scheduling the task request to each second target computing power cluster based on the task requirement computing power and the unit computing power price.

[0026] Obtain the preset cost of the task request, and calculate the cost matching degree corresponding to scheduling the task request to each of the second target computing power clusters based on the preset cost and the execution cost;

[0027] Based on the resource utilization ratio and cost matching degree of the same second target computing power cluster, the cost score of one or more second target computing power clusters is calculated by weighted summation through preset weight thresholds.

[0028] Optionally, training the computing power prediction model using the method described above includes:

[0029] Based on historical task requests, obtain the scheduling execution logs of the computing power cluster, and based on the scheduling execution logs, obtain the historical scheduling data of the computing power cluster; wherein, the historical scheduling data of the computing power cluster includes at least one of computing power type, parameter quantity, data size, execution cost, or computing power usage.

[0030] The historical scheduling data is cleaned and structured to obtain a feature dataset, and the feature dataset is used to train the computing power prediction model to obtain the trained computing power prediction model.

[0031] The step of obtaining the task requirement computing power corresponding to the task request based on the task request information carried in the task request, through a trained computing power prediction model, includes:

[0032] Extract the task request information contained in the task request; wherein, the task request information includes at least one of computing power type, parameter quantity, data scale, or expected cost;

[0033] Input at least one of the computing power type, parameter quantity, data scale, or expected cost into the trained computing power prediction model to obtain the computing power required for the task request.

[0034] Optionally, in the method described above, obtaining the scoring results of one or more of the target computing power clusters includes:

[0035] Based on one or more of the target computing power clusters, according to at least one of the scheduling priority score, the network score, or the cost score, a weighted summation is performed using a preset weight threshold to calculate and obtain the score result of each target computing power cluster.

[0036] Based on the scoring results of one or more of the target computing power clusters, a final computing power cluster is selected and the task request is sent to the final computing power cluster, including:

[0037] Based on the scoring results of one or more target computing power clusters, the target computing power cluster with the highest score is selected as the final computing power cluster, and the task request is sent to the final computing power cluster.

[0038] Optionally, in the method described above, the scheduling method further includes:

[0039] Monitor the status information of the final computing power cluster, and based on the status information, when the final computing power cluster experiences performance fluctuations, perform scheduling priority scoring, network scoring, and cost scoring on one or more second target computing power clusters respectively; wherein, the performance fluctuations include computing power resource utilization exceeding a preset threshold or task execution response time exceeding a preset time threshold;

[0040] Based on the scheduling priority score, the network score, and the cost score, a weighted summation is performed using a preset weight threshold to calculate the score result for each second target computing power cluster.

[0041] Based on the scoring results of one or more second target computing power clusters, the second target computing power cluster with the highest score is selected as the new final computing power cluster, and the task request is sent to the new final computing power cluster.

[0042] Secondly, this application provides a scheduling device, comprising:

[0043] The computational power demand acquisition module is used to receive task requests and, based on the task request information carried in the task request, obtain the computational power demand corresponding to the task request through a trained computational power prediction model.

[0044] The computing power cluster filtering module is used to filter multiple computing power clusters across the entire domain based on the task request information and the computing power required by the task, and obtain one or more target computing power clusters.

[0045] The computing power cluster scoring module is used to score one or more target computing power clusters according to preset cluster scoring rules, and to obtain the scoring results of one or more target computing power clusters; wherein, the cluster scoring rules include at least one of scheduling priority scoring, network scoring or cost scoring;

[0046] The task request sending module is used to filter and obtain the final computing power cluster based on the scoring results of one or more target computing power clusters, and send the task request to the final computing power cluster.

[0047] Thirdly, this application also provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0048] The memory stores computer-executed instructions;

[0049] The processor executes the computer execution instructions stored in the memory to implement the above scheduling method.

[0050] Fourthly, this application also provides a computer-readable storage medium storing a computer program, wherein the computer program implements the above-described scheduling method when executed by a processor.

[0051] Fifthly, this application also provides a computer program product, including a computer program that implements the above-described scheduling method when executed by a processor.

[0052] This application provides a scheduling method, apparatus, and electronic device for distributed heterogeneous computing power clusters. By receiving task requests and extracting the task request information contained within them, and combining this with a trained computing power prediction model to predict the computing power required for the corresponding task, accurate analysis of the task's computing power load is achieved. By using the predicted task demand computing power and task request information as filtering conditions, multiple computing power clusters across the entire domain are filtered to obtain target computing power clusters that meet the requirements of structural compatibility and resource capabilities, effectively enhancing the adaptability between tasks and resources and improving the accuracy of resource selection. Based on this, multi-dimensional scoring is performed on the target computing power clusters using pre-set cluster scoring rules, and corresponding scoring results are obtained. Candidate clusters are quantified from aspects such as scheduling priority, network performance, and resource cost, improving the comprehensiveness and flexibility of scheduling decisions. Finally, the final computing power cluster is selected based on the scoring results, and task requests are allocated to this cluster for execution, achieving a high degree of matching between scheduling decisions and actual task requirements, improving the success rate of task execution, resource utilization, and overall service performance. Attached Figure Description

[0053] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0054] Figure 1 This is a schematic diagram of the cluster scheduling architecture provided in an embodiment of this application;

[0055] Figure 2 A flowchart illustrating the scheduling method for a distributed heterogeneous computing power cluster provided in this application embodiment;

[0056] Figure 3 This is a schematic diagram of the method for obtaining a target computing power cluster provided in an embodiment of this application;

[0057] Figure 4 This is a schematic flowchart of a network scoring method provided in an embodiment of this application;

[0058] Figure 5 A schematic flowchart of a cost scoring method provided in this application embodiment;

[0059] Figure 6 A flowchart illustrating the training of the computing power prediction model provided in this application embodiment;

[0060] Figure 7 This is a schematic flowchart of a scheduling execution monitoring method provided in an embodiment of this application;

[0061] Figure 8 A schematic diagram of the scheduling device provided in the embodiments of this application;

[0062] Figure 9 A schematic diagram of the structure of the electronic device provided in this application.

[0063] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0064] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0065] It should be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0066] In existing technologies, task scheduling in computing clusters typically relies on users manually specifying the required resource specifications in task requests, such as computing power type, resource scale (e.g., memory, number of cores), or specific instance specifications. Upon receiving a request, the system often only performs type matching and resource availability comparison among available clusters based on the user-specified resource configuration parameters to select a suitable computing cluster for task distribution. However, the resource specifications provided by users often lack details about the actual computing power requirements of the task, easily leading to resource redundancy or insufficient resources, resulting in low resource utilization or task execution failure. For example, when users overestimate the resources required for a task, it results in wasted computing resources; conversely, when users underestimate the task requirements, the task may fail to complete or frequently fail due to insufficient resources, affecting overall computing efficiency. On the other hand, existing scheduling strategies lack comprehensive analysis of multi-dimensional factors such as network performance, the location relationship between tasks and clusters, and execution costs. When facing cluster scheduling decisions across different regions, cloud service providers, and architectures, it is difficult to achieve globally optimal resource allocation and cost control. It is evident that existing technologies are insufficient to meet the actual needs of complex tasks for adaptability to computing environments in scenarios involving computing power scheduling in distributed heterogeneous environments.

[0067] Based on the aforementioned technical problems and needs, the inventive concept of this application is to propose a scheduling method for distributed heterogeneous computing power clusters, addressing issues such as inaccurate matching between task requirements and resource supply, and the lack of multi-dimensional decision-making capabilities in computing power scheduling. Specifically, the scheduling process no longer relies on user-specified resource configuration parameters. Instead, it receives task requests, extracts task request information, and automatically obtains the corresponding task's required computing power using a trained computing power prediction model, thereby accurately determining the scale of resources required by the task. After obtaining the task's required computing power, based on the prediction result and task request information, multiple heterogeneous computing power clusters across the entire domain are initially screened to obtain target computing power clusters that meet the requirements of type matching and resource capabilities. To further optimize scheduling decisions, a multi-dimensional scoring mechanism is established to comprehensively score the target computing power clusters from aspects such as scheduling priority, network performance, and cost adaptability. Based on the scoring results, the optimal final computing power cluster is selected, and the task request is scheduled to be executed on that cluster. This application starts with task characteristics, takes computing power matching as the core, and combines multi-dimensional scoring to select computing power clusters that match task requests. This effectively improves the utilization efficiency of computing power resources, reduces computing failures caused by improper resource allocation, and enhances the adaptability to diverse task requirements.

[0068] Figure 1 This is a schematic diagram of the cluster scheduling architecture provided in an embodiment of this application. Figure 1As shown, user terminal 101 is used to initiate task requests. The task request is first sent to control device 102, which is used to receive task requests, parse task information, execute scheduling strategies, and select resources. Upon receiving the task request, control device 102 extracts the task request information and calls the computing power prediction model module deployed in server 103 to predict the task's computing power requirements. Server 103 supports the operation of the computing power prediction model and stores configuration information required for scheduling, such as scoring rules and computing power cluster profiles, providing model execution and data support for the entire scheduling process. After obtaining the required computing power for the task, control device 102, in conjunction with the task request information, executes a resource filtering and scoring process in the distributed heterogeneous computing power cluster 104 to select a matching computing power cluster for task scheduling. This distributed heterogeneous computing power cluster 104 consists of multiple resource clusters of different types, geographical locations, and service providers, reflecting the diversity and heterogeneity of computing power resources. The computing power clusters I and II shown in the example represent two types of heterogeneous resource pools, respectively. For example, computing cluster I can be a high-performance cluster suitable for high-throughput tasks such as image processing and deep neural network execution; computing cluster II may be a hybrid resource cluster including edge nodes, more suitable for low-cost, latency-sensitive, or batch task processing scenarios. The control device 102 calculates the comprehensive score of each target cluster by performing scheduling priority scoring, network scoring, and cost scoring on these clusters, and then performing weighted summation using preset weights. Finally, the cluster with the highest score is selected as the final computing cluster, and task requests are sent to that cluster for execution.

[0069] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0070] Figure 2 This is a flowchart illustrating the scheduling method for a distributed heterogeneous computing power cluster provided in an embodiment of this application. Figure 2 As shown, it includes:

[0071] S21: Receive the task request and, based on the task request information carried in the task request, obtain the task requirement computing power corresponding to the task request through the trained computing power prediction model.

[0072] In this embodiment, task requests are typically issued by user terminals or upper-level scheduling platforms, carrying task request information that includes computational characteristics and scheduling constraints. This task request information characterizes the task's basic requirements and preferences for computing resources. To reduce the uncertainty caused by relying on users manually configuring resource requirement parameters, a trained computing power prediction model is used to analyze the task request information and predict the required computing power. For example, the computing power prediction model employs supervised learning, regression analysis, neural networks, or other prediction methods. During the model training phase, it learns the mapping relationship between task characteristics and resource consumption. Finally, when the task request information is input at runtime, the corresponding required computing power is output. The required computing power is used to quantitatively describe the scale of computing resources needed to complete the current task, for example, expressed using floating-point operations per second (FLOPS), video memory capacity, or comprehensive computing power metrics. This embodiment achieves accurate matching between task requirements and resource supply by using a computing power prediction model to dynamically analyze task requests, providing basic data support for resource selection and scoring in subsequent scheduling. It effectively reduces resource waste or computational failures caused by users misjudging resource needs, while improving the overall automation level of scheduling.

[0073] S22, based on task request information and task-required computing power, filters multiple computing power clusters across the entire domain to obtain one or more target computing power clusters.

[0074] In this embodiment, based on the actual needs of the current task, resources from multiple distributed heterogeneous computing power clusters are screened across the entire domain to obtain a set of target computing power clusters with scheduling potential, providing a candidate resource pool for subsequent cluster scoring and final scheduling. Since computing infrastructure typically consists of multiple computing power clusters, and these clusters may be distributed across different geographical regions, belong to different cloud service providers, and possess different types of computing resources, these heterogeneous resources are standardized for matching and screening based on task request information and task-required computing power. Task request information includes settings for computing power type, resource scale, computing performance, or execution mode, while the task-required computing power provides a clear lower limit for resource requirements; both together form the basis of the screening criteria. For example, during the screening process, preliminary matching of resource types can first be performed among all computing power clusters across the entire domain based on the resource type or structural preferences expressed in the task request, excluding computing power architectures that do not meet the task requirements, thereby constructing a candidate cluster set compatible with resource types. Subsequently, by further combining real-time status information of each cluster, including dynamic indicators such as current available computing power, resource utilization, and service reachability, candidate clusters are filtered to eliminate unavailable or resource-insufficient cluster nodes. This embodiment achieves the screening and filtering of computing power clusters across the entire domain by comprehensively considering static attributes and dynamic operating status, ensuring that the selected target computing power clusters are capable of undertaking the current task in terms of type, scale, and availability. By constructing a set of target clusters, a foundation is laid for subsequent multi-dimensional scoring and scheduling decisions, improving scheduling accuracy while enhancing adaptability to complex tasks and scheduling capabilities for heterogeneous resources.

[0075] S23, score one or more target computing power clusters according to the preset cluster scoring rules, and obtain the scoring results of one or more target computing power clusters; wherein, the cluster scoring rules include at least one of scheduling priority scoring, network scoring or cost scoring.

[0076] In this embodiment, to achieve multi-dimensional scheduling decisions, a pre-defined cluster scoring rule is used to score each target computing power cluster. This scoring rule is constructed based on multiple representative dimensions, comprehensively measuring the scheduling value of each computing power cluster from aspects such as scheduling priority, network performance, and resource cost. The scoring results provide a quantitative basis for the subsequent selection of the final computing power cluster, ensuring that task scheduling has policy transparency and selection rationality. The purpose of scheduling priority scoring is to quantify the usage priority of each computing power cluster according to the pre-defined scheduling strategy, enabling scheduling decisions to reflect the global tendency and policy orientation of resource scheduling. Network scoring focuses on analyzing the network connection performance between the computing power cluster and the task data source, selecting clusters with better network conditions and more efficient data transmission to improve the response speed and stability of task execution. For example, for tasks that are sensitive to data transmission speed, such as large-scale model execution requests and real-time edge data processing, network scoring can improve task response efficiency. The cost scoring is set to measure the economics of task resource usage, ensuring that the selected cluster has a good cost-effectiveness ratio while meeting performance requirements, thereby improving the overall resource utilization efficiency and cost control capabilities of the scheduling process. This embodiment evaluates each target computing power cluster according to scoring rules and outputs corresponding scoring results, achieving a comprehensive analysis of resource schedulability from multiple perspectives and indicators. This scoring mechanism makes the scheduling process highly controllable and flexible in terms of strategy.

[0077] S24. Based on the scoring results of one or more target computing power clusters, filter and obtain the final computing power cluster, and send the task request to the final computing power cluster.

[0078] In this embodiment, each target computing power cluster obtains a comprehensive score based on preset cluster scoring rules. The score result can be a result of a single scoring dimension or a total score of multiple scoring dimensions (such as scheduling priority, network performance, cost, etc.). During scheduling decisions, the scores of each target computing power cluster are compared, and the target computing power cluster with the highest score or optimal configuration is selected as the final computing power cluster. After the final computing power cluster is determined, a scheduling instruction containing task request information is sent to the final computing power cluster, triggering dynamic resource allocation and actual task deployment. The transmission of task requests can be completed through a unified cluster communication interface or a heterogeneous platform adaptation layer, ensuring coordination and consistency between scheduling and cluster management. Through this embodiment, an executable scheduling mapping relationship is established between task requests and computing power resources, realizing a closed loop in the resource scheduling process. In summary, by using final selection and task distribution based on scoring results, not only can the task execution success rate and performance response level be improved, but the problem of excessive or insufficient resource configuration can also be effectively reduced, improving the overall utilization efficiency of computing power resources in a distributed heterogeneous environment.

[0079] In one embodiment, Figure 3 This is a schematic flowchart illustrating a method for obtaining a target computing power cluster according to an embodiment of this application. It specifically describes one implementation of step S22 above, which involves obtaining one or more target computing power clusters. Based on the above embodiment, as... Figure 3 As shown, it includes:

[0080] S31. Based on the task request information, obtain the request computing power type of the task request, and based on the constructed computing power cluster profile, obtain the cluster computing power type of multiple computing power clusters in the entire domain.

[0081] S32, based on the cluster computing power type of multiple computing power clusters in the whole domain, select and obtain the computing power cluster that matches the requested computing power type from the multiple computing power clusters in the whole domain, and use it as the first target computing power cluster, so as to obtain multiple first target computing power clusters;

[0082] S33, monitor the status information of multiple first target computing power clusters and obtain the current available computing power of multiple first target computing power clusters;

[0083] S34, based on the current available computing power of multiple first target computing power clusters, select from the multiple first target computing power clusters the first target computing power clusters to obtain the first target computing power clusters whose current available computing power is greater than or equal to the computing power required by the task, and use them as the second target computing power clusters, so as to obtain one or more second target computing power clusters.

[0084] In this embodiment, the screening process includes two stages: the first stage obtains the first target computing power cluster based on static resource attribute matching, and the second stage obtains the second target computing power cluster based on real-time resource capability judgment. This ensures the rationality of resource selection in terms of both type adaptability and capacity feasibility, laying the foundation for subsequent scoring and scheduling. First, based on the task request information, the requested computing power type required by the task is extracted. The requested computing power type is a basic constraint description of the computing resource structure, reflecting the task's adaptation requirements for the resource architecture. For example, a task request may specify a Graphics Processing Unit (GPU) as the required resource type, or it may more specifically require the use of heterogeneous computing resources such as "a certain model of GPU". This information can be set directly by the user or automatically extracted through request templates or historical task tags. Then, based on the constructed computing power cluster profile, the cluster computing power type of the global computing power cluster is obtained. The computing power cluster profile is a structured description of each computing power cluster, used to characterize the basic attributes of the cluster, including but not limited to computing power type, resource scale, geographical location, network topology, cost information, scheduling weight, etc. The cluster computing power type identifies the category of computing resources hosted by the cluster and serves as the basis for static resource screening. Based on this, by comparing the requested computing power type of the task request with the cluster computing power type of each cluster, a matching computing power cluster is selected from multiple computing power clusters across the entire domain as the primary target computing power cluster. This stage primarily achieves preliminary filtering at the type level, excluding clusters with incompatible computing architectures or resource categories, ensuring that clusters participating in subsequent scoring and scheduling all have structural scheduling feasibility.

[0085] After initial screening, the status of multiple first-target computing power clusters is monitored, and their current available computing power information is collected. Current available computing power typically refers to the total amount of resources in a cluster that is currently idle or available for service, and can be described using multi-dimensional resource indicators such as GPU memory, number of cores, and number of tensor processing units. The computing power required for the task is compared with the current available computing power of each first-target computing power cluster, and clusters whose current available computing power is greater than or equal to the task's required computing power are selected as second-target computing power clusters. This selection ensures that the task can be successfully scheduled and executed in the target cluster after it is issued, avoiding failures due to insufficient resources. One or more second-target computing power clusters obtained in this way constitute a set of scheduling candidate resources with structural compatibility and capability satisfaction. This embodiment, through a hierarchical selection mechanism, realizes the process of building a task-adaptive resource pool from the global computing power clusters, which not only improves selection efficiency and reduces the computational complexity of scoring, but also ensures the resource foundation stability of the scheduling targets, further improving resource utilization and task scheduling reliability in a distributed heterogeneous environment.

[0086] In one specific embodiment, a detailed explanation is provided regarding the scheduling priority scoring of the target computing power cluster in step S23 above. Based on the above embodiment, it includes:

[0087] S231. Based on the constructed computing power cluster profile, obtain the scheduling weight of the second target computing power cluster and record the scheduling weight as the scheduling priority score of the second target computing power cluster, so as to obtain the scheduling priority score of one or more second target computing power clusters.

[0088] In this embodiment, the computing cluster profile is a structured representation of the static characteristics of each computing cluster. It is typically built and maintained uniformly during the initialization or resource update phase, and is used to describe the attributes of the computing cluster, including but not limited to the computing architecture deployed in the cluster (such as GPU, CPU, TPU), computing power scale, cloud provider, geographical location, unit computing power price, pre-set average computing power cost baseline for each task type, and scheduling weight. The scheduling weight is a parameter preset by factors such as operation and maintenance strategies, business needs, resource usage plans, or service level agreements, representing the priority level of each cluster in the scheduling strategy. Specifically, firstly, the scheduling weight corresponding to each second target computing cluster in its profile is read as a static scoring factor. This scheduling weight can be a preset scalar value, such as an integer from 0 to 100; a larger value indicates a higher priority for the cluster, and tasks are more likely to be scheduled to that cluster first. Then, this scheduling weight is directly used as the cluster's scheduling priority score (i.e., score1). Using scheduling weights for priority scoring enables resource allocation based on strategic objectives, such as prioritizing high-efficiency clusters, low-cost clusters, or clusters with better geographical locations. In resource-constrained or high-concurrency scenarios, it can automatically prioritize resource scheduling according to preset scheduling strategies, reducing response latency or resource conflicts caused by resource scheduling relying entirely on real-time computing. It should be noted that the computing cluster profile is a database storing static characteristics; the specific construction steps and forms can adopt existing construction methods, and no specific restrictions are imposed here.

[0089] In one specific embodiment, Figure 4 This is a schematic flowchart of a network scoring method provided in an embodiment of this application. It provides a detailed explanation of step S23 above, which involves performing network scoring on the target computing power cluster. Based on the above embodiment, as... Figure 4 As shown, it includes:

[0090] S41, based on the task request information, extract the first location information corresponding to the task request, and based on the constructed computing power cluster profile, obtain the second location information of one or more second target computing power clusters;

[0091] S42, based on the first location information and one or more second location information, calculate the actual distance between the task request and each second target computing power cluster, and use the extreme value method to normalize one or more actual distances to obtain the normalized distance value between the task request and each second target computing power cluster.

[0092] S43, monitor and acquire network latency information of one or more second target computing power clusters, and use the extreme value method to normalize one or more network latency information to obtain the normalized network latency value of each second target computing power cluster.

[0093] S44. Based on the normalized distance value and normalized network latency value belonging to the same second target computing power cluster, the network score of one or more second target computing power clusters is calculated and obtained by weighted summation through a preset weight threshold.

[0094] In this embodiment, network scoring is performed on each second target computing power cluster from the perspective of network access performance. First, the first location information corresponding to the task request is extracted from the task request information. This first location information is typically used to identify the starting location of the task or the storage location of its main data source, and can be in the form of a physical address, network access node, cloud region identifier, edge node city code, etc. When a task request is initiated, this location information is usually automatically attached by the requesting platform, or it can be set by the user or obtained from the task context, and is used to subsequently determine the physical and network correlation between the cluster and the task. Subsequently, based on the constructed computing power cluster profile, the second location information of each second target computing power cluster is obtained. This location information is one of the static attributes in the cluster profile, identifying the physical deployment location or network service exit location of the computing power cluster. By comparing the first location information with each second location information, the actual distance between the task request and each second target computing power cluster can be obtained. The actual distance reflects the physical proximity of the task request and the target cluster. To achieve unified comparison among multiple clusters, multiple actual distance values ​​are normalized using an extreme value method. The specific formula is:

[0095]

[0096] Calculate and obtain the normalized distance value d n Specifically, determine the maximum value d among all actual distances. max With minimum value d min Then, for each distance value d, calculate its normalized distance value so that it falls within the range of [0, 1]. The shorter the distance, the closer the normalized value is to 0, which indicates a higher degree of proximity and is beneficial for task scheduling.

[0097] Next, network latency information for each second target computing cluster is monitored and acquired. This latency information can be obtained through link detection, historical communication records, or service response times reported by cluster nodes, reflecting the average network round-trip latency from the current access point to the target cluster. Similarly, all network latency values ​​are normalized using the extreme value method, using the formula:

[0098]

[0099] Calculate and obtain the normalized network delay value t n Determine the maximum value t among all network delays. max With minimum value t min Then, for each network latency value t, calculate its normalized network latency value so that it falls within the range of [0, 1], so that the smaller the latency, the closer the normalized value is to 0.

[0100] Finally, the normalized distance value and normalized network latency value of each second target computing power cluster are weighted and summed using the formula:

[0101]

[0102] The network score (score2) is obtained, and during the weighting process, preset weight thresholds (such as w1 and w2) are used to adjust the influence of distance and latency on the final score. For tasks where network latency is the primary influencing factor, a higher weight (w2) can be set; for tasks that need to consider data proximity, the weight (w1) can be increased. The higher the final score, the more suitable the target computing cluster is in terms of network performance for the current task. It should be noted that the sum of the two weights is 100. This embodiment achieves a comprehensive analysis of the network adaptability of the target computing cluster by weighted fusion of network distance and latency. This not only enhances the ability to perceive network status but also enables scheduling decisions to dynamically adapt to the actual network requirements of the task, effectively ensuring the stability and efficiency of task execution.

[0103] In one specific embodiment, Figure 5 This is a schematic flowchart of a cost scoring method provided in an embodiment of this application. A detailed explanation of step S23 above, which involves cost scoring of the target computing power cluster, is provided. Based on the above embodiment, as... Figure 5 As shown, it includes:

[0104] S51, monitor and obtain the current available computing power of one or more second target computing power clusters, and calculate the resource occupancy ratio corresponding to the task request being scheduled to each second target computing power cluster based on the task requirement computing power and the current available computing power;

[0105] S52, based on the constructed computing power cluster profile, obtain the unit computing power price information of one or more second target computing power clusters, and calculate the execution cost corresponding to scheduling the task request to each second target computing power cluster based on the task requirement computing power and unit computing power price.

[0106] S53, obtain the preset cost of the task request, and calculate the cost matching degree corresponding to the task request being scheduled to each second target computing power cluster based on the preset cost and the execution cost;

[0107] S54. Based on the resource occupancy ratio and cost matching degree of the same second target computing power cluster, the cost score of one or more second target computing power clusters is calculated and obtained by weighted summation through preset weight thresholds.

[0108] In this embodiment, the current available computing power of each second target computing power cluster is monitored and acquired. Current available computing power refers to the amount of effective computing resources remaining in the cluster at the current moment that can be allocated to new tasks, typically including computing unit utilization, GPU or main memory usage, and computing power scheduling capabilities. Next, based on the task's required computing power and the current available computing power of each cluster, the resource occupancy ratio corresponding to the task request being scheduled to each second target computing power cluster is calculated. The resource occupancy ratio refers to the proportion of the current task's required computing power in the total currently serviceable computing power of the cluster, used to measure the tightness of the resources required for scheduling the task to the cluster. A lower resource occupancy ratio indicates that the cluster has more remaining capacity to handle the task, helping to alleviate resource pressure and improve cluster stability and overall availability. Subsequently, the unit computing power price information of each second target computing power cluster is extracted from the computing power cluster profile. The unit computing power price can be set according to different resource types and cloud service providers, such as pricing by FLOPS or by GPU card hourly unit price. Combined with the task's required computing power value, the execution cost of the task on each computing power cluster is calculated. Execution cost is used to measure the resource overhead incurred by the task when it is executed on the target cluster, and it forms the basis for subsequent cost fit analysis.

[0109] To analyze the fit between cluster execution costs and task economic expectations, it's necessary to obtain the preset cost parameters carried in the task request. The preset cost is the maximum resource cost allowed for the task as set by the user, or it can be a reference cost (i.e., the default average cost baseline) automatically generated by the system based on task type, historical scheduling experience, or user level. By comparing the execution cost with the preset cost, the cost matching degree can be calculated, reflecting the reasonableness of the current execution cost within the target cost range. Finally, the resource utilization ratio and cost matching degree of each cluster are weighted and summed according to preset weight thresholds, using the formula:

[0110]

[0111] The cost score (score3) is calculated, where rcp is the required computing power for the task, scp is the currently available computing power of the cluster, and in the second target computing power cluster, rcp / scp is less than 1. In this case, the larger the scp, the higher the score. sc is the execution cost, and ec is the expected cost in the task request. The weighting ratio (e.g., w3 for resource utilization ratio, w4 for cost matching degree) can be configured according to the task scenario, and the sum of the two weights is 100. For example, the weight of w3 can be increased for high-load scenarios, and the weight of w4 can be increased for cost-sensitive tasks. This embodiment comprehensively reflects the resource scheduling burden and economic execution capability of the computing power cluster through two dimensions: resource utilization ratio and cost matching degree, achieving a dynamic balance between task execution cost control and resource scheduling efficiency.

[0112] Figure 6 This is a flowchart illustrating the training of a computing power prediction model provided in an embodiment of this application. Based on the above embodiment, training the computing power prediction model includes:

[0113] S61, based on historical task requests, obtain the scheduling execution log of the computing power cluster, and based on the scheduling execution log, obtain the historical scheduling data of the computing power cluster; wherein, the historical scheduling data of the computing power cluster includes at least one of computing power type, parameter quantity, data scale, execution cost or computing power usage;

[0114] S62 performs data cleaning and structuring on historical scheduling data to obtain a feature dataset, and uses the feature dataset to train the computing power prediction model to obtain the trained computing power prediction model.

[0115] In this embodiment, based on historical task request records, the associated computing cluster scheduling execution logs are obtained. This log data typically originates from the operation records of resource scheduling systems, cluster management platforms, or upper-layer task orchestration systems, covering data items from each task scheduling process. For example, the scheduling execution log content includes task identifier, task type, allocated computing cluster, scheduling success or failure, allocated resource details, task execution duration, cost information, etc. This data provides the input features and target output required for model training. Next, based on the scheduling execution logs, historical scheduling data of the computing clusters is extracted and organized to construct a mapping sample between tasks and resources. Historical scheduling data includes multiple dimensions, including but not limited to computing type, parameter quantity, data scale, execution cost, and computing power usage. After initial aggregation, the data may contain outliers, missing items, or inconsistent formats, therefore, data cleaning and structuring processing are required. The cleaning process includes filling or removing missing fields, standardizing units and formats, and detecting and correcting outliers; structuring processing converts the raw log data into feature-label pairs that the model can recognize, forming a standard training sample set. Feature fields typically include task input parameters, such as the number of model parameters, data scale, and computing power type, while the target field is usually the actual computing power value used to complete the task. Based on the constructed feature dataset, machine learning algorithms are used to train the computing power prediction model. Suitable model types include, but are not limited to, multiple linear regression, decision trees, random forests, support vector regression, gradient boosting trees, and neural networks. During model training, the mapping relationship between task features and resource usage is learned based on historical samples, and model performance is optimized through cross-validation and parameter tuning to improve its prediction accuracy and generalization ability in actual scheduling. After training, the trained computing power prediction model is obtained. This embodiment, by constructing structured scheduling training samples and training the model, enables scheduling to have adaptive computing power prediction capabilities for different task types and resource characteristics. This not only improves the accuracy of supply and demand matching between tasks and resources but also reduces resource redundancy and scheduling failures.

[0116] Next, in one embodiment, obtaining the task-required computing power corresponding to the task request includes:

[0117] S211, Extract the task request information contained in the task request; wherein, the task request information includes at least one of computing power type, parameter quantity, data scale or expected cost;

[0118] S212, input at least one of the following: computing power type, parameter quantity, data scale, or expected cost, into the trained computing power prediction model to obtain the computing power required for the task request.

[0119] In this embodiment, task request information is obtained by parsing the data structure of the task request, and key fields related to computing power requirements are extracted. Task request information includes, but is not limited to, computing power type, parameter quantity, data scale, or expected cost. One or more extracted feature fields are used as input variables for the model, and these variables are formatted and input into a trained computing power prediction model. This model has established a mapping relationship between task features and actual resource consumption through historical task samples, and can predict the minimum computing power resources required to complete the requested operation based on the input features of the current task, i.e., the task's required computing power. The prediction output can be a specific numerical value (such as the number of FLOPS, the number of GPU cores required, etc.) or a standard resource specification (such as resource quota level, instance type, etc.), depending on the model settings. By introducing a model prediction mechanism, problems such as inaccurate user resource requests and mismatch between resource allocation and actual task requirements are effectively solved.

[0120] In one embodiment, the scoring results of one or more target computing power clusters obtained in step S24 above are described here. A specific implementation method is described below, based on the above embodiment, including:

[0121] S241, based on one or more target computing power clusters, calculate and obtain the score result of each target computing power cluster by weighted summation according to at least one of scheduling priority score, network score or cost score through a preset weight threshold.

[0122] In this embodiment, for each second target computing power cluster, at least one of its calculated scheduling priority score, network score, or cost score can be selected as the scoring input dimension. Weight values ​​are assigned to the selected scoring dimensions according to preset weight thresholds. For example, the weights of the scheduling priority score, network score, and cost score are set to w5, w6, and w7, respectively, where w5 + w6 + w7 = 1. The weight settings can be flexibly configured according to task type, operating strategy, or user preferences. For example, if the task is a real-time data processing task, the network score weight can be appropriately increased; if it is a training task and cost is sensitive, the cost score weight can be appropriately increased.

[0123] Next, the formula is used.

[0124]

[0125] The scores from each dimension are weighted and summed to calculate the score for the target computing power cluster. If only one or two scoring dimensions are selected, the weights of the unselected dimensions are removed, or the weights are re-normalized proportionally. The final score is used to rank all target computing power clusters; clusters with higher scores are considered more suitable candidates for the current task scheduling. Through standardized scoring and weighted fusion, a unified scheduling scoring system is established, enhancing the comprehensive judgment capability under multi-dimensional conditions. This system effectively adapts to multi-scenario, configurable, and policy-oriented resource scheduling needs, improving the overall scheduling rationality and scalability.

[0126] Furthermore, in one embodiment, based on the scoring results of one or more target computing power clusters, a final computing power cluster is selected and a task request is sent to the final computing power cluster, including:

[0127] S242: Based on the scoring results of one or more target computing power clusters, select the target computing power cluster with the highest score as the final computing power cluster, and send the task request to the final computing power cluster.

[0128] In this embodiment, the scoring results of all second target computing power clusters are sorted, typically from highest to lowest score. The target computing power cluster with the highest score is selected as the final computing power cluster for the task. If the scoring results are similar or there are other business-level control logics in the cluster scheduling strategy, other auxiliary judgment mechanisms can be introduced, such as task priority strategies, i.e., selecting the cluster with the higher scheduling priority score as the final computing power cluster. After the final computing power cluster is determined, the task request is sent to that final computing power cluster. The task request sending process includes operations such as packaging task parameters, attaching resource context, and generating scheduling identifiers, and completing the connection with the cluster scheduling node through the inter-cluster scheduling interface protocol. By selecting the optimal cluster based on the scoring results, the decision transformation from multi-objective evaluation to scheduling execution is realized, ensuring that the task scheduling result is optimal and reasonable under the current scheduling system. This not only improves the success rate and response efficiency of task execution, but also enhances the interpretability and controllability of the scheduling strategy.

[0129] Figure 7 This is a schematic flowchart of a scheduling execution monitoring method provided in an embodiment of this application. Based on the above embodiment, the scheduling method further includes:

[0130] S71, monitor the status information of the final computing power cluster, and based on the status information, when the performance of the final computing power cluster fluctuates, perform scheduling priority scoring, network scoring and cost scoring on one or more second target computing power clusters respectively; wherein, performance fluctuation includes computing power resource utilization exceeding a preset threshold and / or task execution response time exceeding a preset time threshold.

[0131] S72, based on scheduling priority score, network score and cost score, and using a weighted summation method with preset weight thresholds, calculates and obtains the score result of each second target computing power cluster;

[0132] S73: Based on the scoring results of one or more second target computing power clusters, select the second target computing power cluster with the highest score as the new final computing power cluster, and send the task request to the new final computing power cluster.

[0133] In this embodiment, after a task is scheduled to the final computing power cluster, the operational status of the final computing power cluster is continuously monitored. Upon detecting performance fluctuations, a task rescheduling mechanism is triggered to ensure that the task can be executed in the optimal resource environment. Specifically, the status information of the final computing power cluster is first monitored. This status information includes, but is not limited to, dynamic operational indicators such as currently available computing power, task queue length, node resource utilization, network latency, and system load. The monitoring process can be implemented through a cluster status reporting mechanism or by periodically pulling data from the scheduling platform. Based on the collected status information, it is determined whether the current final computing power cluster is experiencing performance fluctuations. The criteria for determining performance fluctuations can be preset by the system, including but not limited to: computing power resource utilization exceeding a preset threshold, abnormal increase in the average response time or execution latency of tasks within the cluster, increased node failure rate, and abnormal network connectivity. Once any of the above indicators is detected to meet the performance fluctuation conditions, the task rescheduling process is immediately triggered.

[0134] In the rescheduling process, scheduling priority scoring, network scoring, and cost scoring are re-performed on one or more second target computing power clusters. These second target computing power clusters are candidate clusters that have previously passed type and capability screening but have not yet been selected as the final execution cluster. The scoring rules are consistent with the initial scheduling, still based on the scheduling weight of each cluster, the network distance and latency between the task and the cluster, and the matching relationship between resource usage costs and expected task costs. Subsequently, the three types of scoring results (scheduling priority score, network score, and cost score) are weighted and summed according to preset weight thresholds to calculate the comprehensive score result for each second target computing power cluster. Based on the comprehensive score result, the second target computing power cluster with the highest score is selected as the new final computing power cluster. Finally, the task request will be resent to this new final computing power cluster for execution, while the task scheduling path is updated and this rescheduling event is recorded.

[0135] This embodiment achieves dynamic adaptation and operational assurance capabilities during the use of computing resources through performance monitoring and re-scoring mechanisms, reducing the interruption or failure of tasks due to resource bottlenecks or cluster performance degradation, and improving the stability and success rate of task execution.

[0136] Figure 8This is a schematic diagram of the scheduling device provided in an embodiment of this application. Figure 8 As shown, the scheduling device 8 includes:

[0137] The computing power acquisition module 81 is used to receive task requests and, based on the task request information carried in the task request, obtain the computing power required for the task request through a trained computing power prediction model.

[0138] The computing power cluster filtering module 82 is used to filter multiple computing power clusters across the entire domain based on task request information and task-required computing power to obtain one or more target computing power clusters.

[0139] The computing power cluster scoring module 83 is used to score one or more target computing power clusters according to preset cluster scoring rules and obtain the scoring results of one or more target computing power clusters; wherein, the cluster scoring rules include at least one of scheduling priority scoring, network scoring or cost scoring;

[0140] The task request sending module 84 is used to filter and obtain the final computing power cluster based on the scoring results of one or more target computing power clusters, and send the task request to the final computing power cluster.

[0141] The scheduling device provided in this embodiment can execute the scheduling method of the above embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0142] Figure 9 A schematic diagram of the structure of the electronic device provided in this application. Figure 9 As shown, the electronic device 9 includes at least one processor 91 and a memory 92. The electronic device 9 also includes a communication component 93. The processor 91, memory 92, and communication component 93 are connected via a bus 94.

[0143] In the specific implementation process, at least one processor 91 executes computer execution instructions stored in memory 92, causing at least one processor 91 to execute the scheduling method executed on the electronic device side as described above.

[0144] The specific implementation process of processor 91 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0145] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0146] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0147] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0148] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described method invocation embodiments at runtime.

[0149] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0150] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described methods of invocation.

[0151] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above-described method invocation embodiments.

[0152] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0153] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0154] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0155] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or in the form of software program modules.

[0156] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0157] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0158] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0159] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0160] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A scheduling method for a distributed heterogeneous computing power cluster, characterized in that, include: Receive a task request, and based on the task request information carried in the task request, obtain the task requirement computing power corresponding to the task request through a trained computing power prediction model. Based on the task request information and the computing power required by the task, multiple computing power clusters across the entire domain are filtered to obtain one or more target computing power clusters; According to the preset cluster scoring rules, one or more target computing power clusters are scored, and the scoring results of one or more target computing power clusters are obtained; wherein, the cluster scoring rules include at least one of scheduling priority scoring, network scoring, or cost scoring; Based on the scoring results of one or more of the target computing power clusters, the final computing power cluster is selected and the task request is sent to the final computing power cluster.

2. The scheduling method according to claim 1, characterized in that, The step of filtering multiple computing power clusters across the entire domain based on the task request information and the computing power required by the task to obtain one or more target computing power clusters includes: Based on the task request information, obtain the request computing power type of the task request, and based on the constructed computing power cluster profile, obtain the cluster computing power type of multiple computing power clusters in the entire domain. Based on the cluster computing power type of multiple computing power clusters in the entire domain, a computing power cluster that matches the requested computing power type is selected from the multiple computing power clusters in the entire domain as the first target computing power cluster, so as to obtain multiple first target computing power clusters; Monitor the status information of multiple first target computing power clusters and obtain the current available computing power of multiple first target computing power clusters; Based on the current available computing power of multiple first target computing power clusters, select a first target computing power cluster whose current available computing power is greater than or equal to the computing power required by the task from the multiple first target computing power clusters, and use it as a second target computing power cluster, so as to obtain one or more second target computing power clusters.

3. The scheduling method according to claim 2, characterized in that, The step of scoring one or more target computing power clusters according to preset cluster scoring rules includes: Based on the constructed computing power cluster profile, the scheduling weight of the second target computing power cluster is obtained, and the scheduling weight is recorded as the scheduling priority score of the second target computing power cluster, so as to obtain the scheduling priority score of one or more second target computing power clusters.

4. The scheduling method according to claim 2, characterized in that, The step of scoring one or more target computing power clusters according to preset cluster scoring rules includes: Based on the task request information, extract the first location information corresponding to the task request, and based on the constructed computing power cluster profile, obtain the second location information of one or more second target computing power clusters; Based on the first location information and one or more second location information, the actual distance between the task request and each second target computing power cluster is calculated and obtained, and the extreme value method is used to normalize one or more of the actual distances to obtain the normalized distance value between the task request and each second target computing power cluster. Monitor and acquire network latency information of one or more second target computing power clusters, and use the extreme value method to normalize one or more network latency information to obtain the normalized network latency value of each second target computing power cluster; Based on the normalized distance value and normalized network latency value belonging to the same second target computing power cluster, the network score of one or more second target computing power clusters is calculated and obtained by weighted summation through a preset weight threshold.

5. The scheduling method according to claim 2, characterized in that, The step of scoring one or more target computing power clusters according to preset cluster scoring rules includes: Monitor and obtain the current available computing power of one or more second target computing power clusters, and calculate and obtain the resource occupancy ratio corresponding to the task request being scheduled to each second target computing power cluster based on the task requirement computing power and the current available computing power; Based on the constructed computing power cluster profile, obtain the unit computing power price information of one or more second target computing power clusters, and calculate the execution cost corresponding to scheduling the task request to each second target computing power cluster based on the task requirement computing power and the unit computing power price. Obtain the preset cost of the task request, and calculate the cost matching degree corresponding to scheduling the task request to each of the second target computing power clusters based on the preset cost and the execution cost; Based on the resource utilization ratio and cost matching degree of the same second target computing power cluster, the cost score of one or more second target computing power clusters is calculated by weighted summation through preset weight thresholds.

6. The scheduling method according to claim 1, characterized in that, Training the computing power prediction model includes: Based on historical task requests, obtain the scheduling execution logs of the computing power cluster, and based on the scheduling execution logs, obtain the historical scheduling data of the computing power cluster; wherein, the historical scheduling data of the computing power cluster includes at least one of computing power type, parameter quantity, data size, execution cost, or computing power usage. The historical scheduling data is cleaned and structured to obtain a feature dataset, and the feature dataset is used to train the computing power prediction model to obtain the trained computing power prediction model. The step of obtaining the task requirement computing power corresponding to the task request based on the task request information carried in the task request, through a trained computing power prediction model, includes: Extract the task request information contained in the task request; wherein, the task request information includes at least one of computing power type, parameter quantity, data scale, or expected cost; Input at least one of the computing power type, parameter quantity, data scale, or expected cost into the trained computing power prediction model to obtain the computing power required for the task request.

7. The scheduling method according to any one of claims 1-5, characterized in that, The step of obtaining the scoring results of one or more of the target computing power clusters includes: Based on one or more of the target computing power clusters, according to at least one of the scheduling priority score, the network score, or the cost score, a weighted summation is performed using a preset weight threshold to calculate and obtain the score result of each target computing power cluster. Based on the scoring results of one or more of the target computing power clusters, a final computing power cluster is selected and the task request is sent to the final computing power cluster, including: Based on the scoring results of one or more target computing power clusters, the target computing power cluster with the highest score is selected as the final computing power cluster, and the task request is sent to the final computing power cluster.

8. The scheduling method according to claim 2, characterized in that, The scheduling method further includes: Monitor the status information of the final computing power cluster, and based on the status information, when the final computing power cluster experiences performance fluctuations, perform scheduling priority scoring, network scoring, and cost scoring on one or more second target computing power clusters respectively; The performance fluctuations include computing resource utilization exceeding a preset threshold or task execution response time exceeding a preset time threshold. Based on the scheduling priority score, the network score, and the cost score, a weighted summation is performed using a preset weight threshold to calculate the score result for each second target computing power cluster. Based on the scoring results of one or more second target computing power clusters, the second target computing power cluster with the highest score is selected as the new final computing power cluster, and the task request is sent to the new final computing power cluster.

9. A scheduling device, characterized in that, include: The computational power demand acquisition module is used to receive task requests and, based on the task request information carried in the task request, obtain the computational power demand corresponding to the task request through a trained computational power prediction model. The computing power cluster filtering module is used to filter multiple computing power clusters across the entire domain based on the task request information and the computing power required by the task, and obtain one or more target computing power clusters. The computing power cluster scoring module is used to score one or more target computing power clusters according to preset cluster scoring rules, and to obtain the scoring results of one or more target computing power clusters; wherein, the cluster scoring rules include at least one of scheduling priority scoring, network scoring or cost scoring; The task request sending module is used to filter and obtain the final computing power cluster based on the scoring results of one or more target computing power clusters, and send the task request to the final computing power cluster.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the scheduling method as described in any one of claims 1 to 8.

Images