Hybrid heterogeneous deployment method for improving operation efficiency of embedded system

Abstract

The invention discloses a hybrid heterogeneous deployment method for improving the running efficiency of an embedded system, and relates to the field of computer embedded operating systems, a neural network is used for performing convolution calculation on a calculation task in a full connection layer to obtain a feature extract as a distribution basis, the task is distributed to Windows and Linux operating systems for the first time, and the operation efficiency of the Windows and Linux operating systems is improved. According to the scheme, real-time level models are established for a first computing task allocated to Windows and a second computing task allocated to Linux respectively, the tasks are allocated to kernel spaces of corresponding operating systems to complete classification operation after real-time evaluation, hybrid heterogeneous deployment is achieved, the advantages of different systems are brought into play, auxiliary allocation of the neural network is more accurate, and the operation efficiency is improved. The real-time level model ensures that high-real-time tasks are processed preferentially, and the requirements for efficiency and real-time performance of an embedded system in a complex scene are met.

Description

Technical Field

[0001] The present invention relates to the field of computer embedded operating systems, and in particular to a hybrid heterogeneous deployment method for improving the operating efficiency of embedded systems. Background Art

[0002] Affected by the development trends of intelligentization, networking, information security isolation and defense of controlled equipment, embedded software systems are becoming more and more complex. The challenges faced by a single OS carrying all functions are increasing. In embedded scenarios, some comprehensive application scenarios are gradually emerging, which require both the high real-time, high reliability and high security of real-time operating systems, and the management capabilities and rich ecology of general-purpose rich-function operating systems. At the same time, benefiting from the rapid development of integrated circuit technology, processor capabilities are becoming more and more powerful. The improvement of single-core capabilities, single-core to multi-core, heterogeneous multi-core and even many-core have laid a solid foundation for deploying multiple OS in a highly integrated processor and module. Hybrid heterogeneous deployment structure provides a variety of hybrid heterogeneous operation solutions for different hardware conditions and security requirements, which is bound to be one of the important directions for the future evolution of embedded systems.

[0003] As computer applications require more processor computing power and computing needs become more diverse, processors have developed rapidly. Although the computing power provided by processors is getting stronger and stronger, it is increasingly difficult to increase the frequency of processors because the integrated circuits in processors have reached the nanometer level. The room for improving the computing power of a single core is already very limited. Therefore, improving the overall computing power of processors through multi-core parallel computing has become a new research hotspot. To give full play to the overall performance of multi-core processors, the design of parallel algorithms for programs is very critical, which also increases the complexity of program design and compilation. Due to the limitations of traditional embedded architectures, the number of cores of a processor cannot be too many, so the method of increasing the number of cores to improve the overall computing power has also encountered bottlenecks. Especially now that artificial intelligence and big data are developing rapidly, the demand for computer computing power has increased dramatically, and traditional embedded systems are difficult to adapt to the growth in demand. Summary of the invention

[0004] Aiming at the problem of improving the computing power of popular multi-core high-performance chips on the market, the present invention makes improvements on the traditional embedded operating system. The present invention proposes a hybrid heterogeneous deployment method for improving the operating efficiency of embedded systems. The scheme proposes to adopt multi-core heterogeneous operation technology to support the mixed operation environment of Linux operating system and Windows operating system. The core technology of the designed hybrid kernel mode lies in the problem of computing task allocation between multi-core processors. The computing task allocation algorithm of the present invention is defined as: a computing task scheduling algorithm based on convolutional deep learning.

[0005] The hybrid heterogeneous deployment solution of the present invention first uses a convolutional neural network to extract features of computing tasks, initially assigns the computing tasks to an adapted operating system for computing, and then analyzes the real-time level of the computing tasks, evaluates the computing tasks, and then assigns the computing tasks to the corresponding kernels, thereby finally completing the classification operation of the computing tasks.

[0006] Convolutional neural network is used to extract features from computational tasks. Convolutional neural network (CNN) is the most commonly used network structure in the field of computer vision. The most important task of convolutional neural network in image processing is feature extraction. The core of convolutional neural network is convolution layer, and the core part of convolution layer is convolution operation. The classic structure diagram of typical convolutional neural network mainly consists of convolution layer, pooling layer, fully connected layer (FC), non-linear units, loss function and other parts.

[0007] The present invention uses the basic features of the computing task as the input layer, and adapts the executed tasks and the operating system through deep learning. In the neural network algorithm, the fully connected layer convolution is located at the last layer of the convolution calculation. However, the convolution of the fully connected layer is different from the convolution calculation of other layers. There are two main differences. First, the amount of data is large. Since the number of input and output channels in the last layer of convolution is much larger than that of other layers of convolution, the number of parameters is significantly increased; second, the feature data needs to be multiplied by the weights. These two characteristics determine that the computing power requirements of the fully connected layer convolution calculation are not high. In order to improve the energy efficiency of the fully connected layer convolution calculation, the present invention designs an acceleration algorithm for the fully connected layer calculation, which can achieve better system performance. The mathematical model is as follows:

[0008]

[0009] The neural network performs convolution calculations on the computing tasks in the fully connected layer to obtain the convolution calculation results. n is the number of input channels in the fully connected layer, and m is the number of feature vectors in a single channel. In the deep learning process, it is necessary to read each weight value Weight and the input feature vector Feature value from the memory device in turn to perform dot product operations, and add the dot product operation result to the bias value Bias, and then shift to obtain the final convolution result Conv result , the convolution result Conv result As feature extracts and as a basis for allocating computing tasks, computing tasks are initially allocated to Windows and Linux operating systems.

[0010] The neural network performs convolution calculations on the computing tasks in the fully connected layer. The convolution calculation results are the feature extracts of the computing tasks. The convolution calculation results are used as the basis for allocating computing tasks. The computing tasks are initially allocated to the Windows operating system and the Linux operating system. After the first allocation of computing tasks is completed, the real-time level models are established for the computing tasks allocated to the Windows operating system and the Linux operating system, respectively, and the computing tasks are allocated according to the real-time level:

[0011] Computing tasks have response level requirements. R represents the response requirement of the task, which is the maximum time interval from the initial preparation state to the computing state. It is defined as follows:

[0012] {R|R∈(0,1,2,3)}

[0013] In the above formula, if R = 0, it means that the task has the lowest real-time response level; if R = 3, it means that the task has the highest real-time response level, and so on. Calculate the priority of task response: Test is determined by the four most important parameters: the energy consumption of the task H, the importance of the task I, the number of execution statements of the task K, and the response requirement of the task R. The mathematical modeling is as follows:

[0014]

[0015] Among them, β is the response priority adjustment coefficient, which can be synchronized with the weight coefficient of the same convolutional neural network.

[0016] The task assigned to the Windows operating system is the first computing task. A first real-time level model is established for the first computing task. After evaluating the first computing task according to the real-time level, the first computing task is assigned to the kernel space of the Windows operating system to complete the reclassification operation; the task assigned to the Linux operating system is the second computing task. A second real-time level model is established for the second computing task. After evaluating the second computing task according to the real-time level, the second computing task is assigned to the kernel space of the Linux operating system to complete the reclassification operation.

[0017] In summary, the beneficial technical effects of this application are:

[0018] 1) The hybrid heterogeneous deployment solution of the present invention is applicable to multi-core processor computing platforms that do not have hardware-assisted virtualization support, supports the Windows operating system and the Linux operating system to work together in the kernel space, provides a unified process running environment and resources in the user space, and seamlessly runs Windows operating system real-time applications and Linux operating system applications.

[0019] 2) Compared with other algorithms, the computing task scheduling algorithm based on convolutional deep learning of the present invention can significantly improve the system's response speed to high real-time tasks.

[0020] 3) Currently, it has mature implementations on multiple processors with different instruction set architectures, and can be applied to scenarios such as the fusion of real-time control and dynamic safety measurement, heterogeneous parallelization of multiple control systems, isolation and protection of PLC real-time domain and security domain, and heterogeneous parallelization of real-time control and intelligent computing.

[0021] 4) The hybrid kernel provides a flexible, scalable and adaptive environment for the operating system through a unified interrupt management framework, supports the coexistence of hybrid kernels and sharing of hardware resources, and provides POSIX and real-time operating system compatible APIs in user space, facilitating the porting of applications to the hybrid operating environment while ensuring strong real-time performance.

[0022] 5) Applicable to single-core or multi-core processor computing platforms, application systems with strong real-time requirements and compatible with Linux operating system and Windows operating system running environment, such as the fusion of motion control and image processing, the fusion of real-time tasks and human-computer interaction interface, and the situation where the computing tasks are complicated. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flow chart of a hybrid heterogeneous deployment method for improving the operating efficiency of an embedded system according to the present invention;

[0024] Figure 2 Compare the computational efficiency of the computing task scheduling algorithm based on convolutional deep learning with other traditional scheduling algorithms. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be described clearly and completely below in conjunction with the drawings 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.

[0026] Figure 1 The figure is a flow chart of a hybrid heterogeneous deployment solution according to an embodiment of the present invention.

[0027] The hybrid heterogeneous deployment solution of this embodiment first uses a convolutional neural network to extract features of computing tasks, initially assigns the computing tasks to an adapted operating system for computing, and then analyzes the real-time level of the computing tasks, evaluates the computing tasks, and then assigns the computing tasks to the corresponding kernels, thereby finally completing the classification operation of the computing tasks.

[0028] This embodiment uses the basic features of the computing task as the input layer, adapts the executed task and the operating system through deep learning, designs an acceleration algorithm for the fully connected layer calculation, and optimizes the system performance. The mathematical model is as follows:

[0029]

[0030] The neural network performs convolution calculations on the computing tasks in the fully connected layer to obtain the convolution calculation results. n is the number of input channels in the fully connected layer, and m is the number of feature vectors in a single channel. In the deep learning process, it is necessary to read each weight value Weight and the input feature vector Feature value from the memory device in turn to perform dot product operations, and add the dot product operation result to the bias value Bias, and then shift to obtain the final convolution result Conv result , the convolution result Conv result As feature extracts and as a basis for allocating computing tasks, computing tasks are initially allocated to Windows and Linux operating systems.

[0031] The neural network performs convolution calculations on the computing tasks in the fully connected layer, and the calculation results obtained are the feature extracts of the computing tasks. The calculation results are used as the basis for allocating computing tasks. The computing tasks are initially allocated to the Windows operating system and the Linux operating system. After the first allocation of computing tasks is completed, the real-time level models are established for the computing tasks allocated to the Windows operating system and the Linux operating system, respectively, and the computing tasks are allocated according to the real-time level:

[0032] Computing tasks have response level requirements. R represents the response requirement of the task, which is the maximum time interval from the initial preparation state to the computing state. It is defined as follows:

[0033] {R|R∈(0,1,2,3)}

[0034] In the above formula, if R = 0, it means that the task has the lowest real-time response level; if R = 3, it means that the task has the highest real-time response level, and so on. Calculate the priority of task response: Test is determined by the four most important parameters: the energy consumption of the task H, the importance of the task I, the number of execution statements of the task K, and the response requirement of the task R. The mathematical modeling is as follows:

[0035]

[0036] Among them, β is the response priority adjustment coefficient, which can be synchronized with the weight coefficient of the same convolutional neural network.

[0037] The task assigned to the Windows operating system is the first computing task. A first real-time level model is established for the first computing task. After evaluating the first computing task according to the real-time level, the first computing task is assigned to the kernel space of the Windows operating system to complete the reclassification operation; the task assigned to the Linux operating system is the second computing task. A second real-time level model is established for the second computing task. After evaluating the second computing task according to the real-time level, the second computing task is assigned to the kernel space of the Linux operating system to complete the reclassification operation.

[0038] In the case of a large number of computing tasks, the computing task scheduling algorithm based on convolutional deep learning of the present invention can significantly improve the system's response speed to high real-time tasks compared with other algorithms; the computing efficiency comparison of the computing task scheduling algorithm based on convolutional deep learning and other traditional scheduling algorithms is shown in the figure below. Figure 2 shown.

Claims

1. A hybrid heterogeneous deployment method for improving the operating efficiency of an embedded system, characterized in that: The neural network performs convolution calculation on the computing task in the fully connected layer, and the obtained calculation result is a feature extract of the computing task. The calculation result is used as the basis for allocating the computing task, and the computing task is initially allocated to the Windows operating system and the Linux operating system; The task assigned to the Windows operating system is a first computing task, a first real-time grade model is established for the first computing task, and after evaluating the first computing task according to the real-time grade, the first computing task is assigned to the kernel space of the Windows operating system to complete the reclassification operation; The task assigned to the Linux operating system is a second computing task. A second real-time level model is established for the second computing task. After the second computing task is evaluated according to the real-time level, the second computing task is assigned to the kernel space of the Linux operating system to complete the re-classification operation.

2. The hybrid heterogeneous deployment solution according to claim 1, characterized in that: The convolution calculation result is calculated by the following formula; Among them, Conv result Convolution is the result of convolution. n is the number of input channels in the fully connected layer. m is the number of feature vectors in a single channel. Feature is the input feature vector. Weight is the weight value. Bias is the bias value. Offset is the offset. Convolution result Conv result Satisfy the value that is much larger than the offset.

3. The hybrid heterogeneous deployment solution according to claim 2, characterized in that: The first real-time level model is expressed by the following formula: Among them, Test represents the priority of the computing task in the first real-time level model, H represents energy consumption, I represents the importance of the task, K represents the number of execution statements of the task, and R represents the response requirement of the task, and the response requirement R can take an integer from 0 to 3.

4. The hybrid heterogeneous deployment solution according to claim 2, characterized in that: The second real-time level model is expressed by the following formula: Among them, Test represents the priority of the computing task in the second real-time level model, H represents energy consumption, I represents the importance of the task, K represents the number of execution statements of the task, and R represents the response requirement of the task, and the response requirement R can take an integer from 0 to 3.

Images