Global planning-based heterogeneous computing power unified scheduling method and system

Abstract

The application relates to the technical field of computing power scheduling, in particular to a heterogeneous computing power unified scheduling method and system based on global planning. A scheduling model is created, computing power demand information is collected, pretreatment information and the computing power demand information are input into the scheduling model, the computing power of each target device is calculated through the scheduling model, the computing power levels of the target devices are divided based on the computing power, then the coupling of the computing power demand is completed in combination with the computing power levels, the standard device of each computing power demand is obtained, finally, real-time information is input into the trained scheduling model, the real-time demand is distributed to the corresponding target device through the trained scheduling model, and the demand distribution of the real-time demand is completed. The computing power of the target device is calculated, the computing power levels of the target devices are divided based on the computing power, the computing power level division and the demand-device coupling are realized, the fuzzy computing power demand is converted into definite device matching rules, and the scheduling complexity is reduced.

Description

Technical Field

[0001] This invention relates to the field of computing power scheduling technology, and more specifically, to a method and system for unified scheduling of heterogeneous computing power based on global planning. Background Technology

[0002] Heterogeneous computing refers to the construction of collaborative computing systems by integrating computing units with different instruction sets, architectures, or functions (such as CPUs, GPUs, FPGAs, ASICs, etc.). The core logic is to dynamically allocate computing resources according to task characteristics. This technology breaks through the performance bottleneck of a single architecture through hardware specialization, achieving a dual improvement in energy efficiency and computing density, and becoming the underlying technical support for the "Eastern Data, Western Computing" project.

[0003] Chinese patent CN119718692B discloses a heterogeneous computing power scheduling method, device, and storage medium. The method includes: responding to a heterogeneous computing power scheduling task sent by a user; classifying the heterogeneous computing power resources of the scheduling system according to the applicable scenarios and types of computing power resources to obtain resource adaptation scenarios and resource adaptation vendors; dividing the heterogeneous computing power resources and the heterogeneous computing power scheduling task into resource pools and task pools respectively; constructing a computing power scheduling objective function; solving the computing power scheduling objective function to obtain resource allocation results; monitoring the task pool, obtaining the current task from the task polling queue, and allocating the computing power resources in the resource pool to the current task according to the resource allocation results, the resource adaptation scenarios, and the resource adaptation vendors. However, in the prior art, the execution efficiency of different devices for different needs is not considered, making it difficult to perform point-to-point allocation of needs to devices when allocating needs, resulting in low scheduling efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a unified scheduling method and system for heterogeneous computing power based on global planning, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, one objective of this invention is to provide a unified scheduling method for heterogeneous computing power based on global planning, comprising:

[0006] 1. A unified scheduling method for heterogeneous computing power based on global planning, characterized in that it includes:

[0007] Collect device information from multiple target devices and preprocess the device information to obtain multiple preprocessed information;

[0008] Create a scheduling model;

[0009] Collect computing power demand information, input preprocessed information and computing power demand information into the scheduling model, calculate the computing power of each target device through the scheduling model, classify the computing power level of the target devices based on the computing power, and then combine the computing power level to complete the coupling of computing power demand, and obtain the standard device for each computing power demand, and obtain the trained scheduling model.

[0010] The system collects real-time information on real-time demands, inputs this information into a trained scheduling model, and then uses the trained scheduling model to allocate real-time demands to the corresponding target devices, thus completing the demand allocation for real-time demands.

[0011] Preferably, the step of collecting device information from multiple target devices and preprocessing the device information to obtain multiple preprocessed information includes the following steps:

[0012] Create a device database;

[0013] For each target device, device information is collected and all collected device information is placed into the device database; the device information includes device type, device parameters, and device status.

[0014] Randomly select all device information for a target device from the device database;

[0015] Determine if there is any missing data in the device information of the target device;

[0016] If there is missing data in the device information of the target device, it will be filled in based on the mean.

[0017] Returns all device information for a randomly selected target device from the device database, until all target devices in the device database have been selected, resulting in multiple preprocessed information entries.

[0018] Preferably, the step of collecting computing power demand information, inputting preprocessed information and computing power demand information into the scheduling model, calculating the computing power of each target device through the scheduling model, classifying the computing power levels of the target devices based on the computing power, and then combining the computing power levels to complete the coupling of computing power demand, thereby obtaining the standard device for each computing power demand, and obtaining the trained scheduling model, includes the following steps:

[0019] All preprocessed information is divided into training and test sets according to a random ratio;

[0020] The training set is input into the scheduling model, which calculates the computing power of each target device and assigns a corresponding computing power level to each target device.

[0021] Collect computing power demand information, input the computing power demand information into the scheduling model, and combine it with the computing power level to allocate the corresponding target device to obtain the trained scheduling model;

[0022] Input the test set into the trained scheduling model to verify whether the training of the scheduling model is complete.

[0023] Preferably, the step of inputting the training set into the scheduling model, calculating the computing power of each target device through the scheduling model, and assigning a corresponding computing power level to each target device includes the following steps:

[0024] Randomly select the device information of a target device from the training set;

[0025] Calculate the computing power of the target device based on the device information;

[0026] Return the device information of a target device randomly selected from the training set, until all target devices in the training set have been selected, and obtain the computing power of each target device;

[0027] Set multiple computing power levels and the corresponding computing power range for each computing power level;

[0028] The computing power level of the target device is classified based on its computing power.

[0029] Preferably, the step of collecting computing power demand information, inputting the computing power demand information into the scheduling model, and allocating corresponding target devices based on computing power levels to obtain the trained scheduling model includes the following steps:

[0030] Collect multiple computing power demand information;

[0031] Each computing power requirement is coupled with the target device to obtain the computing power level required for each computing power requirement;

[0032] Select a computing power requirement;

[0033] Based on the required computing power level, a target device is selected to perform computing power execution, and the coupling relationship between the target device and the computing power requirement is established; the target device for executing the computing power requirement is denoted as the execution device.

[0034] Return to select a computing power requirement until all computing power requirements have been selected, and obtain the execution device corresponding to each computing power requirement; record the execution device corresponding to the computing power requirement as the standard device.

[0035] Preferably, the process of collecting real-time demand information, inputting the real-time information into a trained scheduling model, and allocating the real-time demand to the corresponding target device through the trained scheduling model to complete the demand allocation for real-time needs includes the following steps:

[0036] Collect real-time information on real-time needs;

[0037] Real-time information is input into the trained scheduling model, and the trained scheduling model performs cluster analysis on the real-time demand to obtain the historical computing power demand corresponding to the real-time demand; the historical computing power demand corresponding to the real-time demand is recorded as the target demand.

[0038] Obtain the standard equipment required for the target, and allocate real-time requirements based on the real-time status of the standard equipment.

[0039] Preferably, the step of inputting real-time information into the trained scheduling model, and performing cluster analysis on the real-time demand through the trained scheduling model to obtain the historical computing power demand corresponding to the real-time demand, includes:

[0040] K demand information items are randomly selected as initial cluster centers, and other demand information items and real-time information are randomly assigned to any cluster center to form K clusters;

[0041] For each cluster, calculate the distance from any point within the cluster to the cluster center, and record the point corresponding to the average distance as the new cluster center;

[0042] Set an iteration count threshold;

[0043] Determine whether the number of iterations for each cluster is greater than or equal to the iteration count threshold;

[0044] If the number of iterations of a cluster is greater than or equal to the iteration threshold, the iteration stops, and the cluster center of the cluster containing the real-time information obtained in the last iteration is obtained; the demand information corresponding to the cluster center of the cluster containing the real-time information obtained in the last iteration is recorded as the target demand.

[0045] Preferably, the standard device for acquiring target demand, and the allocation of real-time demand based on the real-time status of the standard device, includes the following steps:

[0046] Collect the real-time operating status of standard equipment; the operating status includes the load, temperature, and remaining available memory of the standard equipment.

[0047] Determine whether the real-time operating status of the standard equipment is in normal working condition;

[0048] If the standard equipment is in normal working condition in real time, then the real-time demand will be allocated to that standard equipment.

[0049] If the real-time operating status of the standard equipment is not in normal working condition, the system returns to the training scheduling model to perform cluster analysis on the real-time demand, obtains the historical computing power demand corresponding to the real-time demand, until the real-time operating status of the standard equipment is in normal working condition.

[0050] On the other hand, this application also provides a heterogeneous computing power unified scheduling system based on global planning, applied to the heterogeneous computing power unified scheduling method based on global planning described in any one of the preceding claims. It includes a data acquisition component and a scheduling component. The data acquisition component collects device information and computing power demand information of the target device. The scheduling component is communicatively connected to the data acquisition component. All data information collected by the data acquisition component is input to the scheduling component. The scheduling component calculates the computing power of the target device and classifies the computing power level of the target device based on the computing power. Thus, the computing power demand is allocated in combination with the computing power level of the computing power demand.

[0051] Preferably, the acquisition component includes a device information acquisition module and a demand information acquisition module. The device information acquisition module acquires device information of the target device, and the demand information acquisition module acquires computing power demand information.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] By collecting device information from multiple target devices and preprocessing the information to obtain multiple preprocessed information sets, a scheduling model is created. Computational power demand information is collected, and the preprocessed information and computational power demand information are input into the scheduling model. The scheduling model calculates the computational power of each target device and classifies the target devices into computational power levels based on their computational power. Then, the computational power levels are combined to couple the computational power demands, resulting in standard devices for each computational power demand. This leads to a trained scheduling model. Finally, real-time demand information is collected and input into the trained scheduling model, which then allocates real-time demands to corresponding target devices, completing the demand allocation for real-time demands. This application achieves computational power level classification and demand-device coupling by calculating the computational power of target devices and classifying their computational power levels. This transforms fuzzy computational power demands into clear device matching rules, reducing scheduling complexity and improving scheduling efficiency. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a unified scheduling method for heterogeneous computing power based on global planning.

[0055] Figure 2 This is a schematic diagram of the framework of a unified scheduling system for heterogeneous computing power based on global planning.

[0056] Reference numerals: 100, Data Acquisition Component; 101, Equipment Information Acquisition Module; 102, Demand Information Acquisition Module;

[0057] 200. Scheduling component. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example 1

[0060] like Figure 1 As shown, one of the objectives of this invention is to provide a unified scheduling method for heterogeneous computing power based on global planning, comprising:

[0061] S100: Collect device information from multiple target devices and preprocess the device information to obtain multiple preprocessed information.

[0062] S200, create a scheduling model;

[0063] S300 collects computing power demand information, inputs preprocessed information and computing power demand information into the scheduling model, calculates the computing power of each target device through the scheduling model, classifies the computing power level of the target devices based on the computing power, and then combines the computing power level to complete the coupling of computing power demand, and obtains the standard device for each computing power demand, thus obtaining the trained scheduling model.

[0064] The S400 collects real-time information about real-time demands, inputs this information into a trained scheduling model, and then uses the trained scheduling model to allocate real-time demands to the corresponding target devices, thus completing the demand allocation for real-time demands.

[0065] It should be noted that by collecting device information from multiple target devices and preprocessing the device information to obtain multiple preprocessed information, a scheduling model is created. Computational power demand information is collected, and the preprocessed information and computational power demand information are input into the scheduling model. The scheduling model calculates the computational power of each target device and classifies the target devices into computational power levels based on the computational power. Then, the computational power levels are combined to couple the computational power demand, resulting in a standard device for each computational power demand. This leads to a trained scheduling model. Finally, real-time demand information is collected and input into the trained scheduling model. The trained scheduling model then allocates real-time demands to the corresponding target devices, completing the demand allocation for real-time demands. This application calculates the computational power of target devices and classifies the computational power levels of target devices based on the computational power, thereby achieving computational power level classification and demand-device coupling. This transforms fuzzy computational power demands into clear device matching rules, reducing scheduling complexity and improving scheduling efficiency.

[0066] In one embodiment of this application, S100 includes:

[0067] S110, Create device database;

[0068] S120, For each target device, collect the device information of the target device and put all the collected device information into the device database; the device information includes device type, device parameters and device status;

[0069] S130, randomly select all device information of a target device from the device database;

[0070] S140, Determine whether there is missing data in the device information of the target device;

[0071] S150, If there is missing data in the device information of the target device, it is filled in based on the mean;

[0072] S160, return all device information for a randomly selected target device from the device database, until all target devices in the device database have been selected, and obtain multiple preprocessed information.

[0073] It is important to note that the completeness and accuracy of device information are fundamental to training the scheduling model. This step stores the original information in a device database and uses mean imputation for missing data. For example, if a GPU's memory parameters are missing, the mean of the memory bandwidth of the same model of GPU is used to fill the gap, ensuring the availability of preprocessed information. For instance, if the memory bandwidth data for a server is missing, but the memory bandwidth data for the other 49 NVIDIA A100 GPUs is all 1935GB / s, the calculated mean is 1935GB / s. This missing value is then filled in, and mean imputation corrects the missing values, preventing computational bias caused by missing data.

[0074] In one embodiment of this application, S300 includes:

[0075] S310, divide all preprocessed information into training and test sets according to a random ratio;

[0076] S320 inputs the training set into the scheduling model, calculates the computing power of each target device, and assigns a corresponding computing power level to each target device.

[0077] S330 collects computing power demand information, inputs the computing power demand information into the scheduling model, and combines the computing power level to allocate the corresponding target device to obtain the trained scheduling model.

[0078] S340. Input the test set into the trained scheduling model to verify whether the training of the scheduling model is complete.

[0079] It should be noted that when dividing the training set and the test set, the ratio of the training set to the test set should be greater than that of the test set to ensure that there are enough training samples in the training set. For example, the ratio of the training set to the test set can be set to 70%:30% for model training and validation.

[0080] After the training set is divided, the scheduling model learns the computing power calculation rules of the devices and the demand-device coupling logic. During the testing phase, the generalization ability of the model is verified. The scheduling model outputs the computing power of each device. The computing power is actually a value used to describe the computing power of the target device. Generally speaking, the higher the computing power, the stronger the computing power of the target device, and the more complex the tasks it can handle.

[0081] After obtaining the trained scheduling model, the training samples in the test set are input into the trained scheduling model, and the response time / accuracy is used as the criterion for judging whether the scheduling model has been trained. For example, after inputting the test set into the model, if the average error between the predicted computing power and the actual computing power is 3%, which is less than the preset threshold of 5%, the model is judged to have been trained.

[0082] In one embodiment of this application, S320 includes:

[0083] S321, randomly select device information of a target device from the training set;

[0084] S322, calculate the computing power of the target device based on the device information;

[0085] S323 returns the device information of a target device randomly selected from the training set until all target devices in the training set have been selected, and obtains the computing power of each target device.

[0086] S324 sets multiple computing power levels and the computing power range corresponding to each computing power level;

[0087] S325 classifies the computing power level of the target device based on its computing power.

[0088] It should be noted that computing power calculation needs to be combined with the type and parameters of the equipment. For CPUs, serial computing power = number of cores × clock speed × IPC, and parallel computing power = number of stream processors × frequency × number of floating-point operations per cycle.

[0089] The computing power level classification needs to be set according to the business needs. For example, high computing power ≥ 100 TFLOPS, medium computing power 10-100 TFLOPS, and low computing power < 10 TFLOPS. This ensures that the level matches the actual task requirements. The level classification simplifies the requirement matching logic, so that high computing power requirements can be directly matched with high-level equipment, thereby improving scheduling efficiency.

[0090] To calculate the parallel computing power of a GPU server: number of stream processors = 16896, frequency = 1.8GHz, floating-point operations per cycle = 2 (FP32), then the computing power = 16896 × 1.8e9 × 2 = 60.8256 TFLOPS, which belongs to the high computing power level. To calculate the serial computing power of a CPU server: number of cores = 40, clock frequency = 2.3GHz, IPC = 2 (executes 2 instructions per clock cycle), then the computing power = 40 × 2.3e9 × 2 = 184 GFLOPS, which belongs to the low computing power level.

[0091] In one embodiment of this application, S330 includes:

[0092] S331 collects multiple computing power demand information;

[0093] S332, each computing power requirement is coupled with the target device to obtain the computing power level required for each computing power requirement;

[0094] S333, select a computing power requirement;

[0095] S334, Select a target device for computing power execution based on the computing power level required by the computing power demand, and establish the coupling relationship between the target device for executing the computing power demand and the computing power demand; denot the target device for executing the computing power demand as the execution device;

[0096] S335, return to select a computing power requirement, until all computing power requirements have been selected, and obtain the execution device corresponding to each computing power requirement; record the execution device corresponding to the computing power requirement as the standard device.

[0097] It should be noted that computing power requirements include task type, data volume, latency requirements, etc., and must be coupled with the computing power level of the equipment. That is, AI inference requires low latency, corresponding to a high computing power level GPU. By matching each requirement with the "minimum computing power level equipment that can meet the conditions", a standard equipment, that is, the ideal execution equipment for the requirement, is determined, providing a reference template for real-time scheduling, thereby avoiding resource waste of "high demand with low configuration" or "low demand with high configuration".

[0098] The requirements for a certain AI inference are: data volume = 10GB, latency requirement ≤20ms, accuracy = INT8. Model analysis shows that it requires high parallel computing power and low latency, matching it with a high-performance GPU server. This GPU is set as the standard device for this requirement, and the coupling relationship of "AI inference - high computing power - GPU - T4" is recorded.

[0099] In one embodiment of this application, S400 includes:

[0100] S410 collects real-time information about real-time needs;

[0101] S420: Real-time information is input into the trained scheduling model. The trained scheduling model performs cluster analysis on the real-time demand to obtain the historical computing power demand corresponding to the real-time demand. The historical computing power demand corresponding to the real-time demand is recorded as the target demand.

[0102] S430 acquires the standard equipment for the target requirements and allocates real-time requirements based on the real-time status of the standard equipment.

[0103] It should be noted that real-time demands are grouped with historical demands through cluster analysis, i.e., demands corresponding to trained standard devices are classified into one category. By leveraging historical experience, the target demand, i.e. the most similar historical demand, is quickly located, avoiding redundant calculations and improving real-time scheduling speed. By reusing historical matching results through clustering, the computational latency of real-time scheduling is reduced, and the accuracy of real-time allocation is improved based on the matching logic of historical successful cases.

[0104] The real-time requirement is "4K video rendering, resolution 3840×2160, frame rate 60fps". The trained scheduling model clusters it with 300 requirements in the historical requirement library (K=5). The Euclidean distance between the real-time requirement and each cluster center is calculated. It is found that the distance between the real-time requirement and the cluster center of "4K rendering-high computing power-GPU-RTX 4090" is the smallest (distance=0.2, threshold=0.5). Therefore, the target requirement is the historical requirement corresponding to this cluster center.

[0105] In one embodiment of this application, S420 includes:

[0106] S421, randomly select K demand information as initial cluster centers, and randomly assign other demand information and real-time information to any cluster center to form K clusters;

[0107] S422, For each cluster, calculate the distance from any point within the cluster to the cluster center, and record the point corresponding to the average distance as the new cluster center;

[0108] S423, Set the threshold for the number of iterations;

[0109] S424, determine whether the number of iterations for each cluster is greater than or equal to the iteration number threshold;

[0110] S425, if the number of iterations of a cluster is greater than or equal to the iteration threshold, then stop the iteration and obtain the cluster center of the cluster where the real-time information obtained in the last iteration is located; record the demand information corresponding to the cluster center of the cluster where the real-time information obtained in the last iteration is located as the target demand.

[0111] It should be noted that the K-means clustering algorithm is used. The K-means algorithm can still maintain a time complexity of O(nkt) (n is the amount of data, k is the number of clusters, and t is the number of iterations) in large-scale demand data, which is suitable for real-time scenarios. It updates the cluster centers iteratively, that is, the average position of the points in the cluster, until convergence (the number of iterations ≥ the threshold or the change of the cluster center ≤ ε). Random initialization of cluster centers avoids local optima, and the threshold of the number of iterations prevents infinite loops, ensuring clustering efficiency and stability. Multiple iterations optimize the cluster centers and improve the rationality of the clustering results.

[0112] In one embodiment of this application, S430 includes:

[0113] S431, Collect the real-time operating status of the standard device; the real-time operating status includes the load, temperature and remaining available memory of the standard device;

[0114] S432, determine whether the real-time working status of the standard equipment is in normal working condition;

[0115] S433, If the real-time working status of the standard equipment is in normal working state, then the real-time demand will be allocated to the standard equipment.

[0116] S434, if the real-time working status of the standard equipment is not in normal working state, then return to perform cluster analysis on the real-time demand through the trained scheduling model to obtain the historical computing power demand corresponding to the real-time demand, until the real-time working status of the standard equipment is in normal working state.

[0117] It should be noted that before allocating real-time requirements, the real-time status of standard equipment must be verified to ensure that it is within the normal operating range. If the requirement is not met, other candidate equipment should be re-clustered to avoid task failure due to equipment failure or overload, thereby eliminating abnormal equipment and reducing the risk of task interruption.

[0118] For example, if the standard device required by the target is a GPU-RTX 4090, its real-time status is: load = 90% (>80% threshold) and temperature = 88℃ (>85℃ threshold). The system determines that it is in an abnormal working state and triggers re-clustering.

[0119] Re-cluster the real-time demands, find the standard device corresponding to the suboptimal cluster center (such as GPU-RTX 3090, load=60%, temperature=75℃), and assign the real-time demands to that device.

[0120] like Figure 2As shown, in one embodiment of this application, a heterogeneous computing power unified scheduling system based on global planning is also provided, applied to the heterogeneous computing power unified scheduling method based on global planning described in any one of the preceding claims. The system includes a data acquisition component 100 and a scheduling component 200. The data acquisition component 100 collects device information and computing power demand information of the target device. The scheduling component 200 is communicatively connected to the data acquisition component 100. All data information collected by the data acquisition component 100 is input to the scheduling component 200. The scheduling component 200 calculates the computing power of the target device and classifies the computing power level of the target device based on the computing power, thereby allocating the computing power demand in combination with the computing power level of the computing power demand.

[0121] It should be noted that the acquisition component 100 is used to collect equipment information and demand information. The acquisition component 100 obtains data through a standardized interface. The scheduling component 200 integrates preprocessing, model training, and real-time allocation modules to realize an end-to-end scheduling process. The acquisition and scheduling components are decoupled, which facilitates separate upgrades.

[0122] In one embodiment of this application, the acquisition component 100 includes a device information acquisition module 101 and a demand information acquisition module 102. The device information acquisition module 101 acquires device information of the target device, and the demand information acquisition module 102 acquires computing power demand information.

[0123] It should be noted that the equipment information acquisition module 101 and the demand information acquisition module 102 have a clear division of labor: the former focuses on hardware layer data, such as type, parameters and status, while the latter focuses on business layer data, such as demand characteristics and constraints. The modules communicate with the scheduling component 200 through an internal bus to ensure that the data format is consistent.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A unified scheduling method for heterogeneous computing power based on global planning, characterized in that, include: Collect device information from multiple target devices and preprocess the device information to obtain multiple preprocessed information; Create a scheduling model; Collect computing power demand information, input preprocessed information and computing power demand information into the scheduling model, calculate the computing power of each target device through the scheduling model, classify the computing power level of the target devices based on the computing power, and then combine the computing power level to complete the coupling of computing power demand, and obtain the standard device for each computing power demand, and obtain the trained scheduling model. Collect real-time information on real-time demands, input the real-time information into the trained scheduling model, and then use the trained scheduling model to allocate real-time demands to the corresponding target devices, thus completing the demand allocation for real-time demands. The process of collecting computing power demand information, inputting preprocessed information and computing power demand information into the scheduling model, calculating the computing power of each target device through the scheduling model, classifying the target devices into computing power levels based on computing power, and then coupling the computing power demand with the computing power levels to obtain the standard device for each computing power demand, thus obtaining the trained scheduling model, includes the following steps: All preprocessed information is divided into training and test sets according to a random ratio; The training set is input into the scheduling model, which calculates the computing power level of each target device and assigns a corresponding computing power level to each target device. Collect computing power demand information, input the computing power demand information into the scheduling model, and combine it with the computing power level to allocate the corresponding target device to obtain the trained scheduling model; Input the test set into the trained scheduling model to verify whether the training of the scheduling model is complete. The process of collecting computing power demand information, inputting this information into the scheduling model, and allocating corresponding target devices based on computing power levels to obtain a trained scheduling model includes the following steps: Collect demand information from multiple computing power requirements; The demand information for each computing power requirement is coupled with the target device to obtain the computing power level required for each computing power requirement; Select a computing power requirement; Based on the required computing power level, a target device is selected for computing power execution, and the coupling relationship between the execution device and the computing power requirement is established. Return to select a computing power requirement, and continue until all computing power requirements have been selected to obtain the execution device corresponding to each computing power requirement; record the execution device corresponding to the computing power requirement as the standard device; The process of collecting real-time demand information, inputting this information into a trained scheduling model, and then allocating the real-time demand to the corresponding target device through the trained scheduling model to complete the demand allocation for real-time needs includes the following steps: Collect real-time information on real-time needs; Real-time information is input into the trained scheduling model, and the trained scheduling model performs cluster analysis on the real-time demand to obtain the historical computing power demand corresponding to the real-time demand; the historical computing power demand corresponding to the real-time demand is recorded as the target demand. Obtain the standard equipment required for the target, and allocate real-time requirements based on the real-time status of the standard equipment.

2. The heterogeneous computing power unified scheduling method based on global planning according to claim 1, characterized in that: The process of collecting device information from multiple target devices and preprocessing the device information to obtain multiple preprocessed information includes the following steps: Create a device database; For each target device, device information is collected and all collected device information is placed into the device database; the device information includes device type, device parameters, and device status. Randomly select all device information for a target device from the device database; Determine if there is any missing data in the device information of the target device; If there is missing data in the device information of the target device, it will be filled in based on the mean. Returns all device information for a randomly selected target device from the device database, until all device information for a randomly selected target device from the device database has been selected, resulting in multiple preprocessed information entries.

3. The heterogeneous computing power unified scheduling method based on global planning according to claim 1, characterized in that: The step of inputting the training set into the scheduling model, calculating the computing power of each target device through the scheduling model, and assigning a corresponding computing power level to each target device includes the following steps: Randomly select the device information of a target device from the training set; Calculate the computing power of the target device based on the device information; Return the device information of a target device randomly selected from the training set, until all target devices in the training set have been selected, and obtain the computing power of each target device; Set multiple computing power levels and the corresponding computing power range for each computing power level; The computing power level of the target device is classified based on its computing power.

4. The heterogeneous computing power unified scheduling method based on global planning according to claim 1, characterized in that: The process of inputting real-time information into the trained scheduling model, and then performing cluster analysis on the real-time demands through the trained scheduling model to obtain the historical computing power demands corresponding to the real-time demands, includes: K demand information items are randomly selected as initial cluster centers, and other demand information items and real-time information are randomly assigned to any cluster center to form K clusters; For each cluster, calculate the distance from any point within the cluster to the cluster center, and record the point corresponding to the average distance as the new cluster center; Set an iteration count threshold; Determine whether the number of iterations for each cluster is greater than or equal to the iteration count threshold; If the number of iterations of a cluster is greater than or equal to the iteration threshold, the iteration stops, and the cluster center of the cluster containing the real-time information obtained in the last iteration is obtained; the demand information corresponding to the cluster center of the cluster containing the real-time information obtained in the last iteration is recorded as the target demand.

5. The heterogeneous computing power unified scheduling method based on global planning according to claim 4, characterized in that: The standard equipment for acquiring target demand, combined with the real-time status of the standard equipment, allocates real-time demand, including the following steps: Collect the real-time operating status of standard equipment; the operating status includes the load, temperature, and remaining available memory of the standard equipment. Determine whether the real-time operating status of the standard equipment is in normal working condition; If the standard equipment is in normal working condition in real time, then the real-time demand will be allocated to that standard equipment. If the real-time operating status of the standard equipment is not in normal working condition, the system returns to the training scheduling model to perform cluster analysis on the real-time demand, obtains the historical computing power demand corresponding to the real-time demand, until the real-time operating status of the standard equipment is in normal working condition.

6. A heterogeneous computing power unified scheduling system based on global planning, applied to the heterogeneous computing power unified scheduling method based on global planning as described in any one of claims 1 to 5, characterized in that, include: Data acquisition components; The acquisition components collect device information and computing power demand information of the target device, respectively. Scheduling component; The scheduling component is communicatively connected to the acquisition component. All data information acquired by the acquisition component is input to the scheduling component. The scheduling component calculates the computing power of the target device and classifies the computing power level of the target device based on the computing power, thereby completing the allocation of computing power requirements based on the computing power level of the computing power requirements.

7. The heterogeneous computing power unified scheduling system based on global planning according to claim 6, characterized in that, The acquisition component includes a device information acquisition module and a demand information acquisition module. The device information acquisition module acquires device information of the target device, and the demand information acquisition module acquires demand information of computing power requirements.

Images