A GPU intelligent computing power control method and system based on load identification and recommendation

Abstract

The present application relates to a method and system for intelligent GPU computing power control based on load identification and recommendation, including obtaining characteristic data of each GPU load in a preset GPU load set, generating a corresponding training sample set based on the characteristic data; based on a preset training scenario, feeding the data of the training sample set to a neural network for supervised training of different scenarios to obtain a model that associates the category information of the GPU load with the training sample set; performing computing power control on GPU loads of different category information and recording and storing the optimal control parameters of the PID controller, associating the optimal control parameters with the GPU load of the corresponding category information; when any characteristic data is identified, the category information of the corresponding GPU load is identified through the model, and the corresponding optimal control parameters are retrieved to adjust the computing power resource allocation of the current GPU load. The present application has the effect of improving the efficiency and accuracy of the dynamic adjustment of the control parameters of the PID controller, so that the GPU computing power resource allocation remains accurate and convergent.

Description

Technical Field

[0001] The present application relates to the technical field of GPU, and in particular to a method and system for controlling GPU intelligent computing power based on load identification and recommendation. Background Art

[0002] At present, the PID algorithm is usually used for scheduling control of computing resource allocation of virtual GPUs. The classic PID controller only guarantees its correctness and convergence when the system performance is linear and time-invariant.

[0003] System performance linearity means that there should be a linear relationship between the control parameter input and the performance output. Time invariance means that when the same control parameters are input at any time, the system performance output should also be the same and will not be affected by time changes.

[0004] However, the linearity and time invariance of system performance are difficult to achieve on GPU intelligent computing devices. First, the kernel functions and data volume of the running system calls often change on a large scale, breaking the linearity and time invariance principles. Second, the performance of large-scale parallel computing changes instantly. Limited by parameters such as video memory, computing core, and data transmission, the performance of running at different times varies greatly, making it difficult to achieve linear system performance. Therefore, the control parameters of the PID controller need to be adjusted dynamically, and how to efficiently and accurately adjust the control parameters of the PID controller to maintain the correctness and convergence of computing resource allocation is a problem worth studying. Summary of the invention

[0005] In order to improve the efficiency and accuracy of dynamic adjustment of PID controller control parameters so that GPU computing power resource allocation remains accurate and convergent, the present application provides a GPU intelligent computing power control method and system based on load identification and recommendation.

[0006] The above-mentioned invention objective of the present application is achieved through the following technical solutions:

[0007] A GPU intelligent computing power control method based on load identification recommendation includes the following steps:

[0008] Obtain a preset GPU load set and obtain characteristic data of each GPU load, the characteristic data including the number of threads of the kernel function corresponding to the GPU load and the requested video memory;

[0009] Generate a corresponding training sample set based on the feature data;

[0010] Based on the preset training scenarios, the data of the training sample set is fed into the neural network for supervised training in different scenarios to obtain a model that associates the category information of the GPU load with the training sample set.

[0011] The computing power is controlled for the GPU load of different categories of information, and the optimal control parameters of the PID controller are recorded and stored, and the optimal control parameters are associated with the GPU load of the corresponding category of information;

[0012] When any characteristic data is identified, the category information of the corresponding GPU load is identified through the model, and the corresponding optimal control parameters are called to adjust the computing power resource allocation of the current GPU load.

[0013] Optionally, the generating a corresponding training sample set based on the feature data includes:

[0014] Get the number of threads and requested video memory of kernel functions corresponding to multiple GPU loads;

[0015] Based on the bucketing method, each preset number of thread number records is used as a thread number training sample, and each preset number of requested video memory records is used as an requested video memory training sample; so as to obtain a thread number training sample set and a requested video memory training sample set;

[0016] The thread number training sample set and the video memory application training sample set are normalized respectively to serve as the training sample sets.

[0017] Optionally, based on a preset training scenario, the data of the training sample set is fed to a neural network for supervised training in different scenarios to obtain a model associating the category information of the GPU load with the training sample set, including:

[0018] The training scenarios include load training scenarios and load reasoning scenarios, and the load training scenarios and the load reasoning scenarios respectively correspond to different thread number training samples collected and different video memory application training samples;

[0019] The GPU load is labeled based on known types to obtain category information of the GPU load, and supervised neural network training is performed on the thread number training sample set, the video memory application training sample set and the category information of the GPU load.

[0020] Optionally, the computing power control is performed on the GPU load of different categories of information and the optimal control parameters of the PID controller are recorded and stored, and the optimal control parameters are associated with the GPU load of the corresponding category of information, wherein the optimal control parameters include load recommendation parameters and PID gain parameters, the load recommendation parameters include the optimal parameters under the computing power control ratio, and the PID gain parameters include the proportional gain parameter Kp, the integral gain parameter Ki, and the differential gain parameter Kd;

[0021] When performing computing power control on the GPU load, the recommended load parameters and PID gain parameters of the PID controller are recorded and stored.

[0022] Optionally, when any feature data is identified, identifying the category information of the corresponding GPU load through the model includes:

[0023] When any characteristic data is identified, the model selects several thread number training samples and applies for video memory training samples through a preset algorithm;

[0024] The number of threads and the amount of memory requested are identified one by one for a number of training samples of the number of threads and training samples of the amount of memory requested, so as to determine the category information of the current GPU load.

[0025] Optionally, when any feature data is identified, the model selects a number of thread number training samples and application memory training samples through a preset algorithm; including:

[0026] When any characteristic data is identified, the number of threads of the characteristic data and the requested video memory are obtained, where the number of threads is equal to the number of thread blocks multiplied by the number of threads in each thread block;

[0027] Select a number of training samples with the same number of threads and the same amount of requested memory;

[0028] Obtain the number of thread blocks and the number of threads in each thread block in the selected thread number training sample, and combine them with the requested video memory to form a sample to be judged;

[0029] Based on the combined features of the number of thread blocks, the number of threads in each thread block, and the amount of requested video memory, the category information of the current GPU load is determined.

[0030] Optionally, when any feature data is identified, the model selects a number of thread number training samples and application video memory training samples through a preset algorithm; and further includes:

[0031] When the identified GPU load does not belong to the preset GPU load set, the model screens out several thread number training samples and video memory application training samples whose error values ​​are within the preset range through load inference scenario inference;

[0032] For several training samples of the number of threads and the training samples of the amount of video memory requested, the number of threads and the amount of video memory requested are identified one by one, and several possible pending category information of the current GPU load is obtained;

[0033] Filter out several optimal control parameters corresponding to the obtained pending category information, adjust the computing power resource allocation of the current GPU load in stages based on the several optimal control parameters, obtain the adjustment result, and select the one with the best adjustment result from the several optimal control parameters to associate with the GPU load.

[0034] The second object of the invention is achieved by the following technical solutions:

[0035] A GPU intelligent computing power control system based on load identification and recommendation is used to execute a GPU intelligent computing power control method based on load identification and recommendation.

[0036] The third objective of the present application is achieved through the following technical solutions:

[0037] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for controlling GPU intelligent computing power based on load identification recommendation are implemented.

[0038] The fourth objective of the present application is achieved through the following technical solutions:

[0039] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned GPU intelligent computing power control method based on load identification recommendation.

[0040] In summary, the present application includes at least one of the following beneficial technical effects:

[0041] 1. By adopting the above method, a large number of GPU loads are classified through the parameter distribution of the kernel function. Since the kernel function sends a large number of data per second, the fault tolerance of the neural network for GPU load recognition is improved by summarizing a preset number of feature data as a sample; the characteristics of the load can be reflected from two dimensions, and the sample is established in a bucketing manner to avoid the loss of information due to the parameters being too general. Furthermore, the computing power is controlled for each type of GPU load after training, and the optimal control parameters are recorded through multiple adjustments. Finally, by identifying the feature data, the model identifies the category information of the corresponding GPU load, and calls the corresponding optimal control parameters to adjust the computing power resource allocation of the current GPU load, which is not limited by the linearity of system performance and the influence of different elimination times, and improves the efficiency and accuracy of the dynamic adjustment of the PID controller control parameters, so that the GPU computing power resource allocation remains accurate and convergent.

[0042] 2. If the current GPU load is not in the preset GPU load set, by distinguishing between training and inference scenarios, the load can still be identified and classified for the feature information that is not in the GPU load set. The recommended computing power control parameters still have reference value and can be further adjusted and compared to match the optimal control parameters for the newly identified GPU load;

[0043] 3. The Kp, Ki, and Kd parameters are dependent on the computing power control ratio, for example, 25%. Therefore, when making GPU load parameter recommendations, the computing power control ratio will also be used as an input parameter. During the GPU load identification training process, the optimal parameters of Kp, Ki, and Kd will also be recorded, and the optimal parameters under different computing power control ratios will be recorded, so that appropriate and accurate final control parameter recommendations can be achieved after load identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flowchart of an implementation of a method for controlling GPU intelligent computing power based on load identification and recommendation according to an embodiment of the present application;

[0045] Figure 2 This is a schematic diagram of a bucketing method in an embodiment of a GPU intelligent computing power control method based on load identification recommendation in the present application;

[0046] Figure 3 This is a PID algorithm control effect diagram in an embodiment of a GPU intelligent computing power control method based on load identification recommendation of the present application;

[0047] Figure 4 This is another PID algorithm control effect diagram in an embodiment of a GPU intelligent computing power control method based on load identification recommendation of the present application;

[0048] Figure 5 It is a principle block diagram of a computer device of the present application. DETAILED DESCRIPTION

[0049] The present application is further described in detail below in conjunction with the accompanying drawings.

[0050] In one embodiment, if Figure 1-5 As shown, the present application discloses a GPU intelligent computing power control method based on load identification recommendation, which specifically includes the following steps:

[0051] S1: Obtain a preset GPU load set, and obtain characteristic data of each GPU load, wherein the characteristic data includes the number of threads of the kernel function corresponding to the GPU load and the requested video memory amount.

[0052] In this embodiment, there are obvious differences in the distribution of kernel function parameters corresponding to different GPU loads, such as the number of threads used by each kernel function, the number of thread blocks multiplied by the number of threads in each thread block, and resource data such as the amount of video memory requested. The amount of video memory requested is the size of video memory requested by the GPU load. For training tasks, the kernel function uses more computing resources than the inference task. Therefore, the load can be analyzed and classified by obtaining the parameters in the kernel function.

[0053] S2: Generate corresponding training sample sets based on feature data.

[0054] Specifically, obtain the number of threads and requested video memory of the kernel function corresponding to multiple GPU loads;

[0055] Based on the bucketing method, each preset number of thread number records is used as a thread number training sample, and each preset number of requested video memory records is used as an requested video memory training sample; so as to obtain a thread number training sample set and a requested video memory training sample set;

[0056] Normalize the thread number training sample set and the video memory application training sample set respectively to serve as the training sample set;

[0057] In this embodiment, bucketing is a data processing technology, which is usually used to divide a data set into several non-overlapping subsets, each of which is called a "bucket". When training a machine learning model, bucketing can be used to deal with class imbalance problems or for feature engineering.

[0058] Every 100 kernel functions are preset, and the number of threads and the amount of video memory requested by each 100 kernel functions are recorded as a sample. These two dimensions can reflect the characteristics of the load. However, if the average value of the parameters of these 100 kernel functions is simply used as a sample, a lot of information will be lost. Therefore, a bucketing method is adopted to establish samples. Including: expanding the dimension of the video memory size requested by the kernel function, as shown in 2, the video memory size requested by these 100 kernel functions is classified according to the power of 2, that is, 2 1 , 2 2 , 2 3 Bytes, retaining the distribution information of the video memory application size in these 100 kernel functions. Based on the same consideration, similar processing is also performed on the number of threads used by the kernel function, which will not be repeated here. Finally, normalization is used to eliminate unnecessary effects caused by different lengths.

[0059] S3: Based on the preset training scenarios, the data of the training sample set is fed into the neural network for supervised training in different scenarios to obtain a model that associates the category information of the GPU load with the training sample set.

[0060] In this embodiment, the training scenarios include a load training scenario and a load reasoning scenario, and the load training scenario and the load reasoning scenario respectively correspond to different thread number training samples and video memory application training samples;

[0061] The GPU load is labeled based on known types to obtain category information of the GPU load, and supervised neural network training is performed on the thread number training sample set, the video memory application training sample set and the category information of the GPU load.

[0062] S4: performing computing power control on GPU loads of different categories of information and recording and storing optimal control parameters of a PID controller, and associating the optimal control parameters with GPU loads of corresponding categories of information;

[0063] In this embodiment, the optimal control parameters include load recommended parameters and PID gain parameters. The load recommended parameters include the optimal parameters under the computing power control ratio. The PID gain parameters include the proportional gain parameter Kp, the integral gain parameter Ki, and the differential gain parameter Kd.

[0064] When performing computing power control on the GPU load, the recommended load parameters and PID gain parameters of the PID controller are recorded and stored.

[0065] The Kp, Ki, and Kd parameters are dependent on the computing power control ratio. Therefore, when recommending the optimal control parameters for GPU load, the computing power control ratio will also be used as an input parameter. During the GPU load identification training, the optimal parameters of Kp, Ki, and Kd will also be recorded, and the optimal parameters under different computing power control ratios will be recorded, so that after GPU load identification, appropriate and accurate control parameters can be recommended.

[0066] S5: When any feature data is identified, the category information of the corresponding GPU load is identified through the model, and the corresponding optimal control parameters are called to adjust the computing power resource allocation of the current GPU load.

[0067] Specifically, when any feature data is identified, the model selects a number of thread number training samples and application memory training samples through a preset algorithm;

[0068] The number of threads and the amount of memory requested are identified one by one for a number of training samples of the number of threads and training samples of the amount of memory requested, so as to determine the category information of the current GPU load.

[0069] In this embodiment, since each sample is obtained by processing 100 feature data, if only one sample is used to determine whether it belongs to the current GPU load, it may not reflect the overall characteristics of the GPU load. Therefore, when identifying the GPU load, the algorithm will select multiple samples, and when each sample is identified, the identification results of all samples are finally combined to finally determine the classification of the GPU load.

[0070] When the feature data is identified, the model obtains the number of threads of the feature data and applies for video memory, where the number of threads is equal to the number of thread blocks multiplied by the number of threads in each thread block;

[0071] Select a number of training samples with the same number of threads and the same amount of requested memory;

[0072] Obtain the number of thread blocks and the number of threads in each thread block in the selected thread number training sample, and combine them with the requested video memory to form a sample to be judged;

[0073] Based on the combined features of the number of thread blocks, the number of threads in each thread block, and the amount of requested video memory, the category information of the current GPU load is determined.

[0074] Furthermore, when the identified GPU load does not belong to the preset GPU load set, the model screens out several thread number training samples and video memory application training samples whose error values ​​are within the preset range through load inference scenario inference;

[0075] The number of threads and the amount of memory requested are identified one by one for several training samples of the number of threads and training samples of the amount of memory requested, and several possible pending category information of the current GPU load is obtained; if only one training sample of the number of threads and one training sample of the amount of memory requested are screened out, the corresponding optimal control parameters are directly identified.

[0076] Filter out several optimal control parameters corresponding to the obtained pending category information, and adjust the computing power resource allocation of the current GPU load in stages based on the several optimal control parameters, wherein the staged adjustment refers to controlling several optimal control parameters one by one, and recording the adjusted results, obtaining the adjustment results and selecting the one with the best adjustment result from several optimal control parameters to associate with the GPU load, and updating the GPU load to a preset GPU load set.

[0077] Optionally, after the computing power resource allocation of the current GPU load is adjusted in stages based on several optimal control parameters, if the adjustment result still does not meet the preset standards, new optimal control parameters suitable for the current GPU load are inferred within the range of several optimal control parameters through load inference scenarios.

[0078] In one embodiment, referring to Figure 3 , which is the PID algorithm control effect diagram (dual process equal proportion) when GPU load identification and optimal control parameters are not matched, refer to Figure 4 , is the PID algorithm control effect diagram after adopting the recommended optimal control parameters, Figure 4 Comparison Figure 3 , which quickly enters the convergence state.

[0079] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0080] In one embodiment, a GPU intelligent computing power control system based on load identification and recommendation is provided, which is used to execute a GPU intelligent computing power control method based on load identification and recommendation in the above-mentioned embodiment.

[0081] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the computer program is executed by the processor, a GPU intelligent computing power control method based on load identification recommendation is implemented.

[0082] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a method for controlling GPU intelligent computing power based on load identification and recommendation is implemented.

[0083] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, a method for controlling GPU intelligent computing power based on load identification recommendation is implemented.

[0084] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0085] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0086] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A GPU intelligent computing power control method based on load identification and recommendation, characterized in that: Obtain a preset GPU load set and obtain characteristic data of each GPU load, the characteristic data including the number of threads and requested video memory of the kernel function corresponding to the GPU load; Generate a corresponding training sample set based on the feature data; Based on the preset training scenarios, the data of the training sample set is fed into the neural network for supervised training in different scenarios to obtain a model that associates the category information of the GPU load with the training sample set. The computing power is controlled for the GPU load of different categories of information, and the optimal control parameters of the PID controller are recorded and stored, and the optimal control parameters are associated with the GPU load of the corresponding category of information; When any characteristic data is identified, the category information of the corresponding GPU load is identified through the model, and the corresponding optimal control parameters are called to adjust the computing resource allocation of the current GPU load; The method of generating a corresponding training sample set based on the feature data comprises: Get the number of threads and requested video memory of kernel functions corresponding to multiple GPU loads; Based on the bucketing method, each preset number of thread number records is used as a thread number training sample, and each preset number of requested video memory records is used as an requested video memory training sample; so as to obtain a thread number training sample set and a requested video memory training sample set; Normalize the thread number training sample set and the video memory application training sample set respectively to serve as the training sample set; When any feature data is identified, the corresponding GPU load category information is identified through the model, including: When any characteristic data is identified, the model selects several thread number training samples and applies for video memory training samples through a preset algorithm; For a number of training samples of the number of threads and training samples of the amount of video memory requested, the number of threads and the amount of video memory requested are identified one by one to determine the category information of the current GPU load; When any characteristic data is identified, the model selects a number of thread number training samples and application memory training samples through a preset algorithm; including: When any characteristic data is identified, the number of threads of the characteristic data and the requested video memory are obtained, where the number of threads is equal to the number of thread blocks multiplied by the number of threads in each thread block; Select a number of training samples with the same number of threads and the same amount of requested memory; Obtain the number of thread blocks and the number of threads in each thread block in the selected thread number training sample, and combine them with the requested video memory to form a sample to be judged; Based on the combined features of the number of thread blocks, the number of threads in each thread block, and the amount of requested video memory, the category information of the current GPU load is determined.

2. The method for controlling GPU intelligent computing power based on load identification and recommendation according to claim 1, characterized in that: Based on the preset training scenarios, the data of the training sample set is fed to the neural network for supervised training of different scenarios to obtain a model that associates the category information of the GPU load with the training sample set, including: the training scenarios include load training scenarios and load reasoning scenarios, and the load training scenarios and the load reasoning scenarios respectively correspond to different collected thread number training samples and application video memory training samples; The GPU load is labeled based on known types to obtain category information of the GPU load, and supervised neural network training is performed on the thread number training sample set, the video memory application training sample set and the category information of the GPU load.

3. The method for controlling GPU intelligent computing power based on load identification and recommendation according to claim 1, characterized in that: The computing power control is performed on the GPU load of different categories of information and the optimal control parameters of the PID controller are recorded and stored, and the optimal control parameters are associated with the GPU load of the corresponding category of information, wherein the optimal control parameters include load recommendation parameters and PID gain parameters, the load recommendation parameters include the optimal parameters under the computing power control ratio, and the PID gain parameters include proportional gain parameters Kp, integral gain parameters Ki, and differential gain parameters Kd; When performing computing power control on the GPU load, the recommended load parameters and PID gain parameters of the PID controller are recorded and stored.

4. The method for controlling GPU intelligent computing power based on load identification and recommendation according to claim 1, characterized in that: When any characteristic data is identified, the model selects a number of thread number training samples and application memory training samples through a preset algorithm; and also includes: When the identified GPU load does not belong to the preset GPU load set, the model screens out several thread number training samples and video memory application training samples whose error values ​​are within the preset range through load inference scenario inference; For several training samples of the number of threads and the training samples of the amount of video memory requested, the number of threads and the amount of video memory requested are identified one by one, and several possible pending category information of the current GPU load is obtained; Filter out several optimal control parameters corresponding to the obtained pending category information, adjust the computing power resource allocation of the current GPU load in stages based on the several optimal control parameters, obtain the adjustment result, and select the one with the best adjustment result from the several optimal control parameters to associate with the GPU load.

5. A GPU intelligent computing power control system based on load identification and recommendation, characterized in that: Used to execute a GPU intelligent computing power control method based on load identification recommendation as described in any one of claims 1 to 4.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of a GPU intelligent computing power control method based on load identification recommendation as described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a GPU intelligent computing power control method based on load identification recommendation as described in any one of claims 1 to 4 are implemented.

Images