An adaptive acceleration system for AI data processing based on Bayesian algorithm

Abstract

The present invention discloses an AI data processing adaptive acceleration system based on a Bayesian algorithm, which belongs to the field of data processing and computing technology, and specifically includes: locating a target operator, creating a data processing pipeline monitoring thread, collecting target data and hardware resource information and saving them as current configuration information, starting an automatic optimization function, collecting operator information and marking bottleneck operators, setting the number of iterations to perform a hyperparameter search on the operator information, using the result of each hyperparameter search to update the current configuration information as new configuration information, using the new configuration information to guide the adjustment of the operators on the data processing pipeline, after executing the number of iterations, screening out the optimal hyperparameters and storing them in a local file, and directly calling and adjusting the operators when the pipeline is started next time. The present invention can effectively improve the data processing speed during AI model training.

Description

Technical Field

[0001] The present invention relates to the field of data processing and computing technology, and in particular to an AI data processing adaptive acceleration system based on a Bayesian algorithm. Background Art

[0002] AI data processing is the basis for subsequent model training and optimization, so establishing an efficient data processing system has become a practical need. Since the operator performance is fixed in a given version of a specific AI framework, users tend to improve overall training efficiency by adjusting scripts.

[0003] The adjustment process of complex models will inevitably involve a large number of parameter adjustment operations. How to quickly and effectively find the best parameter combination has become the key to improving AI data processing efficiency.

[0004] However, the existing technology relies on sufficient computing resources and user initial settings, and has a poor effect on improving data processing efficiency in resource-constrained environments. In addition, the traditional technology lacks flexibility, and the effect of accelerating data processing in different hardware configuration environments varies greatly. Therefore, the present invention provides an AI data processing adaptive acceleration system based on a Bayesian algorithm to solve the problems existing in the existing technology. Summary of the invention

[0005] The purpose of the present invention is to provide an AI data processing adaptive acceleration system based on a Bayesian algorithm to solve the following technical problems:

[0006] Traditional AI data processing acceleration technologies rely on sufficient computing resources and user initial settings. They are not very effective in improving data processing efficiency in resource-constrained environments. In addition, traditional technologies lack flexibility and cannot effectively achieve adaptive data processing acceleration for different hardware configuration environments.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] An AI data processing adaptive acceleration system based on a Bayesian algorithm, comprising:

[0009] The information parsing module is used to locate the device queue operation operator in the data processing pipeline and mark it as the target operator. It initializes and creates a monitoring thread in the data processing pipeline, monitors the queue composed of the target operator and all previous operators, collects target data by tracking node operations and sampling node operations, and stores the target data in the current configuration information, collects hardware resource information, and also stores the hardware resource information in the current configuration information.

[0010] The automatic tuning module is used to obtain the current configuration information and perform automatic analysis. When it is determined that the efficiency of the current data processing stage is lower than the preset standard, the optimization function is automatically started, the number of iterations is set, and the information of all operators on the data processing pipeline is collected in each iteration. The information includes the operator parallelism and the operator output queue size. At the same time, all operators that limit the data processing efficiency are detected and marked as bottleneck operators. A hyperparameter search is performed on the operator information, and a hyperparameter set is obtained through the hyperparameter search. The current configuration information is updated to the new configuration information according to the hyperparameter set. After the number of iterations is executed, a set of optimal hyperparameter sets is screened out from all the hyperparameter sets obtained;

[0011] The pipeline optimization module is used to receive the new configuration information generated by each iteration in the automatic tuning module and adjust the data processing pipeline. After executing the number of iterations, the current data processing pipeline is adjusted according to the optimal hyperparameter set.

[0012] As a further solution of the present invention: in the information parsing module, the target data includes tracking node data and sampling node data. The tracking node operation collects the connector size corresponding to the last batch processing operator on the data pipeline in real time, which is marked as tracking node data. The sampling node operation is an asynchronous operation, calling the Sample function to traverse all operators on the data processing pipeline, and regularly collects the connector size corresponding to each operator and converts it into a vector representation, which is marked as sampling node data.

[0013] As a further solution of the present invention: in the information analysis module, the hardware resource information includes CPU utilization and system memory occupancy, the CPU utilization includes device CPU utilization, operator CPU utilization and processing process CPU utilization, the system memory occupancy includes operator memory occupancy and processing process memory occupancy, and the sampling time is synchronously recorded when collecting the hardware resource information.

[0014] As a further solution of the present invention: In the automatic tuning module, the specific process of automatic analysis is: using the target data, calculating the input queue utilization of the data sent to the GPU / NPU buffer; if the utilization is lower than K%, where K is a preset threshold, it is determined that the efficiency of the current data processing stage is lower than the preset standard, and the automatic optimization function is started.

[0015] As a further solution of the present invention: In the automatic tuning module, the specific process of detecting and marking the bottleneck operator is: traversing all operators on the data processing pipeline, detecting the input and output queue status of each operator in turn, and if the input queue of any operator is full and the output queue is empty, the operator is marked as a bottleneck operator.

[0016] As a further solution of the present invention: in the automatic tuning module, the process of super parameter search includes:

[0017] Set initial hyperparameters, including the initial parallelism and output queue size of each operator, set the search space, which is the range of the number of working threads of each operator, and the range size is set to [1, number of CPU cores], and set the number of iterations B, where B is a natural number;

[0018] The optimization index of the data processing pipeline is set to the execution time of the data processing pipeline and the system memory occupancy rate, and a comprehensive score definition formula is set, and the comprehensive score definition formula is:

[0019] f(X)=k(X)×r+m(X)×(1-r);

[0020] Where X represents the hyperparameter set, X={x1, x2, ..., x n}, the elements in X represent the parallelism of each operator and the size of the operator's output queue, n is the total number of all operators on the data processing pipeline, f(X) represents the comprehensive score of the data processing pipeline under X, k(X) represents the execution time of the data processing pipeline under X, m(X) represents the system memory occupancy under X, r is the weighted parameter for balancing the two optimization indicators, 0<r<1, k(X) and m(X) are both normalized results;

[0021] After starting the automatic optimization function, the comprehensive score definition and search space are used in combination with the Gaussian process model, random search algorithm and Bayesian algorithm to perform hyperparameter search. Each hyperparameter search obtains a new set of hyperparameters, and the new set of hyperparameters is serialized in JSON format and stored in the current configuration information. At the same time, the current configuration information is updated to the new configuration information.

[0022] As a further solution of the present invention: in the automatic tuning module, the process of super parameter search further includes:

[0023] Construct an observation data set, which contains the current hyperparameter search results and historical hyperparameter search results, as well as the corresponding data processing pipeline comprehensive scores. The formula for the observation data set is:

[0024] D b =D b-1 ∪{(X * ,f(X * ))};

[0025] Where b is the current iteration number, b=1, 2, ..., B, B represents the set iteration number X * It represents the set of hyperparameters obtained in the current iteration, f(X *) indicates the corresponding X * The comprehensive score of the data processing pipeline under b represents the observed data set obtained in the current iteration, D b-1 represents the observed data set obtained at the end of the b-1th iteration, D b-1 The expression is:

[0026] D b-1 ={(X1,f(X1)),(X2,f(X2)),...,(X b-1 ,f(X b-1 ))};

[0027] where X1, X2, ..., X b-1 represents the set of hyperparameters obtained by historical hyperparameter search, f(X1), f(X2), ..., f(X b-1 ) represents the comprehensive score of the data processing pipeline under the corresponding hyperparameter set.

[0028] As a further solution of the present invention: the specific process of adjusting the data processing pipeline in the pipeline optimization module is: receiving new configuration information in the automatic tuning module through the API interface, deserializing and reading the new configuration information, and then adjusting the operators on the data processing pipeline, wherein the adjustment method includes adjusting the number of working threads of each operator, adjusting the output queue size of each operator, and merging or separating MAP operator operators;

[0029] The merge operation means that when there are N consecutive MAP operators for data processing, these N operators are merged into one MAP operator. The split operation means that when the execution time of any MAP operator for processing data exceeds the preset time, the operator is separated into M sub-MAP operators according to the processing logic, and each sub-MAP operator processes data in parallel.

[0030] As a further solution of the present invention: the working mode between the information analysis module, the automatic tuning module and the pipeline optimization module is concurrent execution. After the pipeline optimization module receives the new configuration information obtained by the current iteration of the automatic tuning module, it adjusts the data processing pipeline. At the same time, the information analysis module executes the adjusted data processing pipeline and collects the current configuration information of the data processing pipeline.

[0031] As a further solution of the present invention: in the automatic tuning module, after obtaining a set of optimal hyperparameter sets, the hyperparameter sets are serialized in JSON format and stored in a local file. When the data processing pipeline is started next time, the pipeline optimization module reads the optimal hyperparameter set through deserialization and adjusts the operators on the data processing pipeline.

[0032] Beneficial effects of the present invention:

[0033] The automatic optimization algorithm designed by the present invention adopts a random search algorithm to reduce the size of the search space, and then combines the Bayesian optimization algorithm to iteratively optimize the hyperparameter search results, so that the hyperparameter results are more accurate and stable, effectively improving the data processing efficiency of the operator, and avoiding the problem of memory overflow under resource-constrained conditions. The present invention uses three groups of modules to coordinate and cooperate, automatically adjust the operator parallelism of the data processing pipeline according to environmental resources, and make full use of system resources to accelerate data processing speed and improve AI training performance, thereby achieving efficient utilization of hardware resources. At the same time, the present invention defines a comprehensive scoring indicator for the data processing pipeline, balances execution time and memory occupancy, and ensures that the best data processing acceleration performance can be obtained under different hardware environments, with excellent flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below in conjunction with the accompanying drawings.

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 It is a schematic diagram of the data processing pipeline update process of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] See also Figure 1-Figure 2 As shown, the present invention is an AI data processing adaptive acceleration system based on Bayesian algorithm, comprising:

[0039] The information parsing module is used to locate the device queue operation operator (device_queue_op) in the data processing pipeline and mark it as the target operator. At the same time, a monitoring thread is initialized and created in the data processing pipeline to monitor the queue composed of the target operator and all previous operators. The current data processing system is pipeline parallel, so the present invention selects the queue size rather than the quantitative time consumption as the monitoring indicator to obtain the changes in the data processing process in real time, and uses the tracking node operation (TracingNode) and the sampling node operation (SamplingNode) to collect the target data, and the target data includes the tracking node data and the sampling node data, and the target data is stored in the current configuration information in JSON format.

[0040] The tracking node operation collects the connector size corresponding to the last batch operator on the data pipeline in real time, and marks it as tracking node data, so that the status of the host can be tracked every time data is collected. The sampling node operation consists of functions such as Init, Sample, and SaveToFile, and is an asynchronous operation. The sampling node operation will call the Sample function to traverse all operators on the data processing pipeline, regularly collect the connector size corresponding to each operator and convert it into a vector representation, which is marked as sampling node data. Combining this data with the queue computing capability, the average queue utilization rate in the entire sampling process can be calculated.

[0041] The information parsing module also collects hardware resource information, which includes CPU utilization and system memory occupancy. The CPU utilization includes device CPU utilization, operator CPU utilization and processing CPU utilization. The system memory occupancy includes operator memory occupancy and processing memory occupancy. The sampling time is recorded synchronously when collecting hardware resource information, and the hardware resource information is also stored in the current configuration information in JSON format.

[0042] By collecting the device CPU utilization, we can obtain the changes and fluctuations of the current system CPU resources. The process CPU utilization can help the system understand the CPU consumption of the current process and determine whether other tasks occupy CPU resources. The operator CPU utilization helps the system find which operators consume more CPU resources during execution time and adjust the number of threads according to the actual situation.

[0043] The automatic tuning module is used to call the Profiler API interface to obtain the current configuration information and perform automatic analysis. The specific process of automatic analysis is as follows: using the target data, calculate the input queue utilization of data sent to the GPU / NPU buffer. If the utilization is lower than K%, where K is the preset threshold, it is judged that the efficiency of the current data processing stage is lower than the preset standard, and the automatic optimization function is started.

[0044] In the automatic tuning module, relevant initial settings need to be performed before designing the adaptive acceleration algorithm, including: setting the initial hyperparameters, which include the initial parallelism and output queue size of each operator, setting the search space, which is the range of the number of working threads of each operator, and the range size is set to [1, number of CPU cores]. It should be noted that since the queue length depends on the speed at which the previous operator processes data, the maximum queue length must also be less than the number of working threads of the operator. In the user view, the data set input and preprocessing pipeline are executed concurrently by multiple threads according to the preset hyperparameters in actual operation. At this time, the queue size can be approximately regarded as the thread queue size, and the number of iterations B is set at the same time, where B is a natural number.

[0045] According to the design goal of the present invention, the optimization index of the data processing pipeline is set to the execution time of the data processing pipeline and the system memory occupancy rate. In the algorithm design process, these two indicators are first balanced, and a comprehensive score is defined for the data processing pipeline. Here, the pipeline execution time and the system memory occupancy rate are first normalized, then multiplied by the corresponding weight ratio and added to obtain the pipeline comprehensive score, and the definition formula obtained is:

[0046] f(X)=k(X)×r+m(X)×(1-r);

[0047] Where X represents the hyperparameter set, X={x1, x2, ..., x n}, the elements in X represent the parallelism of each operator and the output queue size of the operator, n is the total number of all operators on the data processing pipeline, f(X) represents the comprehensive score of the data processing pipeline under X, the lower the comprehensive score, the higher the data processing efficiency of the data processing pipeline, k(X) represents the execution time of the data processing pipeline under X, m(X) represents the system memory occupancy under X, r is the weighted parameter for balancing the two optimization indicators, 0<r<1.

[0048] When designing the algorithm, it is clear that the results of the hyperparameter search include the parallelism of each data processing operator and the size of the internal output queue. The size of the search space depends on the scale of the data enhancement operation and the concurrency supported by the hardware. Considering that there may be more data enhancement operations in typical scenarios such as CV and other fields, resulting in a larger search space, a random search algorithm is used to reduce the size of the search space when selecting the algorithm, and the Bayesian algorithm is used for iterative optimization on this basis. The optimization result of the algorithm is to minimize the execution time of the data processing pipeline. This method can be used to search for a better parallelism hyperparameter.

[0049] When designing an adaptive acceleration algorithm, it is also necessary to construct an observation data set within the algorithm. The observation data set contains the current hyperparameter search results and historical hyperparameter search results, as well as the corresponding data processing pipeline comprehensive score. The formula for the observation data set is:

[0050] D b =D b-1 ∪{(X * ,f(X * ))};

[0051] Where b is the current iteration number, b=1, 2, ..., B, B represents the set iteration number, X * It represents the set of hyperparameters obtained in the current iteration, f(X * ) indicates the corresponding X * The comprehensive score of the data processing pipeline under b represents the observed data set obtained in the current iteration, D b-1 represents the observed data set obtained at the end of the b-1th iteration, D b-1 The expression is:

[0052] D b-1 ={(X1,f(X1)),(X2,f(X2)),...,(X b-1 ,f(X b-1 ))};

[0053] where X1, X2, ..., X b-1 represents the set of hyperparameters obtained by historical hyperparameter search, f(X1), f(X2), ..., f(X b-1 ) represents the comprehensive score of the data processing pipeline under the corresponding hyperparameter set. The observation dataset can not only be used as a reference for each iterative optimization, but also facilitates the return of the optimal hyperparameter set in the automatic tuning module.

[0054] When the automatic tuning module starts the automatic optimization function, it traverses the operators on the data processing pipeline in each iteration, and checks the input and output queues of each operator in turn. If the input queue of any operator is full and the output queue is empty, the operator is marked as a bottleneck operator. At the same time, all operator information on the pipeline is collected, and the operator information is searched for hyperparameters using an adaptive acceleration algorithm. A hyperparameter set is obtained through hyperparameter search in each iteration, and the hyperparameter set is serialized in JSON format and stored in the current configuration information. At the same time, the current configuration information is updated to the new configuration information. After the number of iterations is completed, a set of hyperparameter sets that achieve the highest data processing efficiency of the data processing pipeline is selected from all the hyperparameter sets obtained and marked as the optimal hyperparameter set.

[0055] The serialization mentioned above is a way to convert an object into a byte stream. The corresponding method is deserialization, which converts a binary stream into an object. Since serialization converts the state information of an object into a byte stream, the information can be written to a file. Even in different environments, just reading these byte streams can construct an object that is exactly the same as the initial object state.

[0056] The pipeline optimization module is used to receive the new configuration information generated by the automatic tuning module each iteration through the API interface, deserialize and read the new configuration information, and then adjust the operators on the data processing pipeline. The adjustment method includes adjusting the number of working threads of each operator, thereby changing the parallelism of the operator to avoid CPU resource preemption under resource-constrained conditions, thereby reducing the overall processing efficiency, adjusting the output queue size of each operator to minimize memory consumption while maintaining peak performance levels, thereby freeing up more memory for other modules, and merging or separating MAP operator operators.

[0057] The merge operation refers to replacing multiple consecutive MAP operator operators with a high-performance operator when they are used for data processing. The split operation refers to splitting the operator into multiple sub-MAP operator operators according to the processing logic when the execution time of any MAP operator operator for processing data exceeds the preset time. Each sub-MAP operator operator processes data in parallel. The purpose of merging and splitting operators is to balance the operator thread overhead.

[0058] The working mode among the information analysis module, the automatic tuning module and the pipeline optimization module in the present invention is concurrent execution, and when the automatic optimization function is started, the working process among the three modules is also an iterative cycle process: after the pipeline optimization module receives the new configuration information obtained by the current iteration of the automatic optimization module, it adjusts the data processing pipeline, and at the same time, the information analysis module executes the adjusted data processing pipeline and collects the current configuration information of the data processing pipeline, and the automatic tuning module receives the current configuration information again to complete one iteration.

[0059] After all iterations are executed, the automatic tuning module will select a set of optimal hyperparameter sets, and serialize the hyperparameter set in JSON format and store it in a local file. Depending on the number of cards trained, a corresponding number of tuning files will be obtained. For example, in a 4-card training environment, 4 tuning files will be obtained, corresponding to the tuning configuration of the data pipeline on the 4 cards. In a single-card environment, there will only be one tuning file. The pipeline optimization module adjusts the current data processing pipeline according to the optimal hyperparameter set, and when the data processing pipeline is started next time, the pipeline optimization module directly reads the optimal hyperparameter set through deserialization, realizes automatic real-time adjustment of the hyperparameters related to each concurrent execution operator, and improves user-friendliness.

[0060] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0061] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0062] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An AI data processing adaptive acceleration system based on Bayesian algorithm, characterized in that: include: The information parsing module is used to locate the device queue operation operator in the data processing pipeline and mark it as the target operator. It initializes and creates a monitoring thread in the data processing pipeline, monitors the queue composed of the target operator and all previous operators, collects target data by tracking node operations and sampling node operations, and stores the target data in the current configuration information, collects hardware resource information, and also stores the hardware resource information in the current configuration information. The automatic tuning module is used to obtain the current configuration information and perform automatic analysis. When it is determined that the efficiency of the current data processing stage is lower than the preset standard, the optimization function is automatically started, the number of iterations is set, and the information of all operators on the data processing pipeline is collected in each iteration. The information includes the operator parallelism and the operator output queue size. At the same time, all operators that limit the data processing efficiency are detected and marked as bottleneck operators. A hyperparameter search is performed on the operator information, and a hyperparameter set is obtained through the hyperparameter search. The current configuration information is updated to the new configuration information according to the hyperparameter set. After the number of iterations is executed, a set of optimal hyperparameter sets is screened out from all the hyperparameter sets obtained; The pipeline optimization module is used to receive the new configuration information generated by each iteration in the automatic tuning module and adjust the data processing pipeline. After executing the number of iterations, the current data processing pipeline is adjusted according to the optimal hyperparameter set.

2. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 1, characterized in that: In the information parsing module, the target data includes tracking node data and sampling node data. The tracking node operation collects the connector size corresponding to the last batch processing operator on the data pipeline in real time, which is marked as tracking node data. The sampling node operation is an asynchronous operation, calling the Sample function to traverse all operators on the data processing pipeline, and regularly collects the connector size corresponding to each operator and converts it into a vector representation, which is marked as sampling node data.

3. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 1, characterized in that: In the information analysis module, the hardware resource information includes CPU utilization and system memory occupancy, the CPU utilization includes device CPU utilization, operator CPU utilization and processing CPU utilization, the system memory occupancy includes operator memory occupancy and processing memory occupancy, and the sampling time is synchronously recorded when collecting the hardware resource information.

4. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 2, characterized in that: In the automatic tuning module, the specific process of automatic analysis is: using the target data, calculate the input queue utilization of the data sent to the GPU / NPU buffer. If the utilization is lower than K%, where K is the preset threshold, it is judged that the efficiency of the current data processing stage is lower than the preset standard, and the automatic optimization function is started.

5. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 4 is characterized in that: In the automatic tuning module, the specific process of detecting and marking the bottleneck operator is: traversing all operators on the data processing pipeline, detecting the input and output queue conditions of each operator in turn, and if the input queue of any operator is full and the output queue is empty, the operator is marked as a bottleneck operator.

6. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 5, characterized in that: In the automatic tuning module, the hyper-parameter search process includes: Set initial hyperparameters, including the initial parallelism and output queue size of each operator, set the search space, which is the range of the number of working threads of each operator, and the range size is set to [1, number of CPU cores], and set the number of iterations B, where B is a natural number; The optimization index of the data processing pipeline is set to the execution time of the data processing pipeline and the system memory occupancy rate, and a comprehensive score definition formula is set, and the comprehensive score definition formula is: f(X)=k(X)×r+m(X)×(1-r); Where X represents the hyperparameter set, X={x1, x2, ..., x n }, the elements in X represent the parallelism of each operator and the size of the operator's output queue, n is the total number of all operators on the data processing pipeline, f(X) represents the comprehensive score of the data processing pipeline under X, k(X) represents the execution time of the data processing pipeline under X, m(X) represents the system memory occupancy under X, r is the weighted parameter for balancing the two optimization indicators, 0<r<1, k(X) and m(X) are both normalized results; After starting the automatic optimization function, the comprehensive score definition and search space are used in combination with the Gaussian process model, random search algorithm and Bayesian algorithm to perform hyperparameter search. Each hyperparameter search obtains a new set of hyperparameters, and the new set of hyperparameters is serialized in JSON format and stored in the current configuration information. At the same time, the current configuration information is updated to the new configuration information.

7. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 6, characterized in that: In the automatic tuning module, the hyper-parameter search process also includes: Construct an observation data set, which contains the current hyperparameter search results and historical hyperparameter search results, as well as the corresponding data processing pipeline comprehensive scores. The formula for the observation data set is: D b =D b-1 ∪{(X * ,f(X * ))}; Where b is the current iteration number, b=1, 2, ..., B, B represents the set iteration number, X * It represents the set of hyperparameters obtained in the current iteration, f(X * ) indicates the corresponding X * The comprehensive score of the data processing pipeline under b represents the observed data set obtained in the current iteration, D b-1 represents the observed data set obtained at the end of the b-1th iteration, D b-1 The expression is: D b-1 ={(X1,f(X1)),(X2,f(X2)),...,(X b-1 ,f(X b-1 ))}; where X1, X2, ..., X b-1 represents the set of hyperparameters obtained by historical hyperparameter search, f(X1), f(X2), ..., f(X b-1 ) represents the comprehensive score of the data processing pipeline under the corresponding hyperparameter set.

8. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 7, characterized in that: The specific process of adjusting the data processing pipeline in the pipeline optimization module is: receiving new configuration information in the automatic tuning module through the API interface, deserializing and reading the new configuration information, and adjusting the operators on the data processing pipeline. The adjustment method includes adjusting the number of working threads of each operator, adjusting the output queue size of each operator, and merging or separating MAP operator operators; The merge operation means that when there are N consecutive MAP operators for data processing, these N operators are merged into one MAP operator. The split operation means that when the execution time of any MAP operator for processing data exceeds the preset time, the operator is separated into M sub-MAP operators according to the processing logic, and each sub-MAP operator processes data in parallel.

9. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 8, characterized in that: The information analysis module, the automatic tuning module and the pipeline optimization module work in a concurrent execution mode. After the pipeline optimization module receives the new configuration information obtained by the current iteration of the automatic tuning module, it adjusts the data processing pipeline. At the same time, the information analysis module executes the adjusted data processing pipeline and collects the current configuration information of the data processing pipeline.

10. The AI ​​data processing adaptive acceleration system based on Bayesian algorithm according to claim 1, characterized in that: In the automatic tuning module, after obtaining a set of optimal hyperparameter sets, the hyperparameter sets are serialized in JSON format and stored in a local file. When the data processing pipeline is started next time, the pipeline optimization module reads the optimal hyperparameter set through deserialization and adjusts the operators on the data processing pipeline.

Images