Edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction

Abstract

The present invention discloses a method for collaborative optimization and offloading of edge DPU computing power with adaptive demand prediction. The method comprises: establishing a computing power demand function model and improving it according to the environment; constructing a Laplace distribution model based on historical computing power resource demand data to predict the computing power resource demand of the device; determining the cost matrix for allocating resources to each device by each edge node; establishing a computing power resource optimization function model and solving it to achieve the optimal allocation of computing power resources. The present invention fully considers the random changes in the computing power demand of the tasks proposed by the device, and proposes a method for collaborative optimization and offloading of edge DPU computing power with adaptive demand prediction. The computing power demand function model is improved according to environmental factors, and the computing power demand of the tasks proposed by the device is predicted through the Laplace distribution model. Considering the limited computing power resources of the edge node, the computing power demand of the tasks proposed by the device can be met with the least computing power resources.

Description

Technical Field

[0001] The present invention relates to the field of computing power resources, and in particular to an edge DPU computing power collaborative optimization and unloading method based on demand adaptive prediction. Background Art

[0002] As two important technologies in the integration of information and communication technologies in the 5G era, edge computing and computing power networks are important supports for the development and implementation of emerging businesses. As a new network technology solution to the problem of unified resource supply when multiple levels of computing power resources coexist, computing power networks can be used to provide a unified resource supply. Figure 2 As shown in the figure, the computing power network distributes the computing power, storage, algorithm and other resource information of the service nodes through the network control plane, and combines network information and resource requirements to provide distribution, association, transaction and allocation of computing, storage, network and other resources, thereby achieving the optimal configuration and use of the entire network resources.

[0003] The computing power requirements of a task posed by a device depend not only on the task type, data size, algorithm complexity, and the device's own DPU capabilities, but also on the environment in which the device operates. The computing power requirements of a task vary randomly. Therefore, given the limited computing resources of edge nodes, accurately predicting the computing power requirements of a device and using the minimum computing resources to meet the computing power requirements of the task posed by the device remains an unresolved issue. Therefore, it is necessary to predict computing power requirements based on historical computing power demand data collected by sensors to achieve optimal resource allocation. Summary of the Invention

[0004] To address the above problems, the present invention provides an edge DPU computing power collaborative optimization unloading method with demand adaptive prediction, establishes a computing power demand function model, and predicts the computing power resource demand through the Laplace distribution model to solve the problem of using the least computing power resources to meet the computing power requirements of the tasks proposed by the device.

[0005] The present invention provides an edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction, which includes the following steps:

[0006] Step 1: Establish a computing power demand function model based on the computing power resources required by the device and the task type, data volume, algorithm complexity, and DPU capability of the device's computing task.

[0007] Step 2: Introduce environmental factors of the device environment to improve the computing power demand function model;

[0008] Step 3: Based on the historical data of computing resource requirements of the device, a probability density function of the Laplace distribution model is established to predict the computing resource requirements of the device;

[0009] Step 4: Establish a cost matrix for allocating resources from each edge node to each device;

[0010] Step 5: Establish a computing power resource optimization function model;

[0011] Step 6: Solve the computing power resource optimization model to achieve the optimal allocation of computing power resources.

[0012] Furthermore, the computing power demand function model in step 1 is specifically expressed as follows:

[0013]

[0014] in, is the computing resource requirement of device i at time t; S st Indicates the computing resource requirement value for a fixed task type, data volume, algorithm complexity, and device DPU capability; is the task type of the computing task proposed by device i at time t; is the data size of the computing task proposed by device i at time t; is the algorithm complexity of the computing task proposed by device i at time t; is the DPU capability of device i.

[0015] Furthermore, the improved computing power demand function model in step 2 is specifically expressed as follows:

[0016]

[0017] in, is the computing resource requirement of the improved device i at time t, is the environmental factor of device i at time t, is the temperature of the environment where device i is located at time t, is the humidity of the environment where device i is located at time t, is the power stability of the environment where device i is located at time t, is the network bandwidth of the environment where device i is located at time t, and λ1, λ2, λ3, and λ4 are parameters.

[0018] Furthermore, the probability density function of the Laplace distribution model in step 3 is specifically expressed as follows:

[0019]

[0020] in express The probability density of for The sample value of computing power resource demand of device i in the time period, T is the collection period of historical data of computing power resource demand, n is the number of equal parts of the collection period, and They are The mean and variance of

[0021] According to the probability density function, the maximum and minimum predicted values ​​of the computing power resource requirements of the device are obtained, which are specifically expressed as follows:

[0022]

[0023]

[0024] in, and are the maximum and minimum predicted values ​​of the computing resource requirements of the device, respectively. P[·] represents the probability, and the parameter θ = 0.0001;

[0025] Establish a task complexity function and use the artificial neural network prediction method to train the historical data of computing resource requirements to obtain the task complexity at time t. And normalized to the task complexity factor The final predicted value of the computing power resource demand of the device is then calculated using the maximum and minimum predicted values ​​of the computing power resource demand of the device. The specific implementation is as follows:

[0026]

[0027]

[0028]

[0029] in, is the final predicted value of computing resource demand of device i at time t, is the task type of the computing task proposed by device i at time t-1, is the data size of the computing task proposed by device i at time t-1; is the algorithm complexity of the computing task proposed by device i at time t-1, and β_Max is the maximum value of the task complexity.

[0030] Furthermore, in step 4, the cost matrix for allocating resources from each edge node to each device is established through the analytic hierarchy process. The specific steps are as follows:

[0031] A hierarchical structure model is established, with the comprehensive cost of allocating computing resources from edge nodes to devices as the highest target layer, the distance between edge nodes and devices, obstacle density, channel conditions, interference between channels, and resource matching as factors to be considered in the intermediate criterion layer, and m×n resource allocation methods for m devices and n edge nodes in the computing network as the lowest solution layer; a comparison judgment matrix is ​​constructed based on the five factors in the intermediate criterion layer; the consistency of the comparison judgment matrix is ​​tested using a single-rank ranking method, and a consistency matrix is ​​constructed, using the matrix A=(a ij ) 5×5 , (i=1,2,…,m,j=1,2,…,n), where:

[0032]

[0033] The factor weight values ​​of each factor at the criterion layer are obtained by the arithmetic mean method and recorded as h1, h2, h3, h4, and h5, respectively, forming the factor weight matrix H = [h1, h2, h3, h4, h5]. The factor comparison judgment matrix of the m×n resource allocation methods at the solution layer relative to the factors at the criterion layer is constructed respectively, and the consistency test is performed. The solution weight values ​​of the m×n resource allocation methods at the solution layer relative to the factors at the criterion layer are obtained by the arithmetic mean method and recorded as Construct the scheme weight matrix E, which is specifically expressed as follows:

[0034]

[0035] According to the factor weight matrix H and the solution weight matrix E, the cost matrix W is obtained, which is specifically expressed as follows:

[0036] W=H×E=[ω 1,1 ,…,ω i,j ,…,ω m,n ],(i=1,2,…,m,j=1,2,…,n)

[0037] Among them, ω i,j Represents the cost weight of edge node j allocating computing resources to device i.

[0038] Furthermore, the computing power resource optimization function model in step 5 is specifically expressed as follows:

[0039]

[0040]

[0041]

[0042] Among them, Y t* is the optimal solution vector, The optimal computing power resource allocated by edge node j to device i at time t; It represents the size of the computing power resource allocated by edge node j to device i at time t. It is the maximum value of the computing power resource of edge node j, ψ i (ε) represents the degree of relaxation of the computing power demand of device i.

[0043] Furthermore, in step 6, the computing power resource optimization function model is solved based on the simulated annealing algorithm and the relaxation method. The specific steps are as follows:

[0044] S1: Initialization Among them, the parameter ε = 0.0001. The iteration step τ = 0, ψ i (ε) is the relaxation factor representing the degree of relaxation of the computing power resource demand of device i.

[0045] S2: Use the simulated annealing algorithm to solve the computing power resource optimization function model;

[0046] S3: If there is a solution in step S2, output the optimal solution vector Y t* ; If there is no solution, jump to step S4;

[0047] S4: Let τ = τ + 1 and recalculate

[0048] S5: Judge the iteration step τ. If τ > L, end the iteration and the computing power resource offloading fails, where L is the maximum iteration step; if τ < L, jump to step S2.

[0049] The present invention also provides a demand - adaptive prediction - based edge DPU computing power collaborative optimization offloading system. The system includes:

[0050] A computing power demand function model construction unit, which is used to establish a computing power demand function model between the computing power resources required by a device and the task type, data volume size, algorithm complexity of the computing tasks proposed by the device, and the DPU capabilities of the device itself, and improve it by introducing the environmental factors of the environment where the device is located;

[0051] A computing power resource demand prediction unit, which is used to establish a probability density function of the Laplace distribution model based on the historical data of the computing power resource demand of the device and predict the computing power resource demand of the device; <00001​​​​​​​The computing power resource allocation unit is used to solve the computing power resource optimization model to achieve the optimal allocation of computing power resources.

[0055] The present invention also provides a device comprising:

[0056] Memory;

[0057] processor;

[0058] as well as

[0059] computer programs;

[0060] The computer program is stored in the memory and configured to be executed by the processor to implement the method described above.

[0061] The present invention also provides a storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0062] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0063] (1) The present invention fully considers various factors affecting the computing power resource requirements of the tasks proposed by the device, among which the task type, data size, algorithm complexity and the DPU capability of the device itself play an important role, and establishes a computing power demand function model between the computing power demand and the task type, data size, algorithm complexity and the DPU capability of the device itself;

[0064] (2) The present invention establishes an environmental loss function based on environmental factors such as the temperature, humidity, power supply stability, and network bandwidth of the environment in which the device is located, obtains environmental factors, and then further improves the computing power demand function model based on the environmental factors;

[0065] (3) The present invention considers that the task type, data size, and algorithm complexity of the computing task proposed by the device are all random variables. Based on the historical data of computing power requirements collected by sensors in the recent period, a Laplace distribution model is established to predict the computing power resource requirements of the device;

[0066] (4) The present invention considers factors such as the distance between edge nodes and devices, obstacle density, channel conditions, interference between channels, and resource matching, and uses the hierarchical analysis method to derive the cost matrix of each edge node allocating resources to each device. Based on the computing power resources allocated by the edge nodes to the devices and the cost matrix of each edge node allocating resources to each device, a computing power resource optimization function model is established;

[0067] (5) The present invention solves the computing power resource optimization model based on the random optimization algorithm-simulated annealing algorithm to achieve the optimal allocation of computing power resources;

[0068] (6) The present invention can use the least amount of computing power resources to meet the computing power requirements of the tasks proposed by the device under the condition that the computing power resources of the edge node are limited. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of the present invention.

[0070] Figure 2 This is a model diagram of the computing power network system.

[0071] Figure 3 It is the flow chart of resource allocation algorithm. DETAILED DESCRIPTION

[0072] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0073] The present invention proposes a method for collaborative optimization and offloading of edge DPU computing power based on demand adaptive prediction. Figure 1 The specific steps are as follows:

[0074] In step S101, the task type, data size, algorithm complexity, and DPU (Data Processing Unit) capability of the computing task proposed by the device are the main factors affecting the computing power resources required by the device in the computing network. Each device can only propose one computing task at a time. A computing power demand function model is established between the computing power resources required by the device and the task type, data size, algorithm complexity, and DPU capability of the computing task. The specific representation method is as follows:

[0075]

[0076] in, is the computing resource requirement of device i at time t; S st The computing resource requirement value for a fixed task type, data volume, algorithm complexity, and device DPU capability; The task type factor of the computing task requested by device i at time t. For example, basic image and text processing tasks require relatively low computing resources, while relatively large-scale scientific computing, machine learning and other tasks require higher computing resources. The data size of the computing task requested by device i at time t. Under the premise of the same task type, tasks with larger data volumes require higher computing resources; is the algorithm complexity of the computing task requested by device i at time t. When the task type and data size are the same, the computing resource demand of the device is proportional to the algorithm complexity; is the DPU capability of device i. Under the same task conditions, devices with stronger DPU capabilities require less computing resources.

[0077] In step S102, considering that the environment in which the device is located will affect the computing power demand, an environmental loss function is established based on environmental factors such as the temperature, humidity, power supply stability, and network bandwidth of the device environment to obtain an environmental factor. The computing power demand function model is further improved based on the environmental factor. The specific implementation method is as follows:

[0078]

[0079] in, For environmental factors, is the temperature of the environment where device i is located at time t. The higher the temperature, the lower the performance of the device or even crashes. Therefore, the greater the environmental loss, the more computing power resources are needed to dissipate the heat to ensure the normal operation of the task. is the humidity of the environment where device i is located at time t. The higher the humidity, the more likely the electronic components of the device will be corroded or short-circuited, resulting in increased losses and requiring more computing power for error detection and correction. The power supply stability of the environment where device i is located at time t. A less stable power supply will lead to voltage fluctuations and increased environmental losses. The device needs to consume more computing power to handle voltage anomalies to protect the device. is the network bandwidth of the environment where device i is located at time t. The limitation of network bandwidth will cause the data transmission speed to slow down or be interrupted. In order to meet the real-time requirements, the device needs higher computing power to process and cache data, which increases the environmental loss. λ1, λ2, λ3, and λ4 are parameters.

[0080] According to environmental factors Further improve the computing power demand function model, the specific formula is as follows:

[0081]

[0082] in, For the improved computing power demand function model, the greater the loss caused by the environment in which device i is located, the greater the environmental factor The larger it is, the more computing power resources are needed to ensure that the task proposed by device i can run normally, so that the improved computing power demand function model can be applied to different physical environments.

[0083] Step S103: Considering that the task type factor, data size, and algorithm complexity of the computing task proposed by the device are all complex random variables, if these random variables are predicted separately and then integrated into one expression based on the prediction results, it will be difficult to form a closed-loop expression. It is still a random variable, so first put multiple random variables in one expression and output them as one random variable, and finally make a unified prediction for one random variable.

[0084] Based on the historical data such as task type factors, data volume, and algorithm complexity collected by sensors in the recent period, the computing power demand sample value is calculated based on the computing power demand sample function model, and the frequency histogram of the computing power demand sample value is obtained. It is found that the computing power demand sample value follows the Laplace distribution model. The mean and variance of the computing power demand sample value are calculated, and the computing power demand of device i at time t is predicted. The specific implementation method is as follows:

[0085] Divide the most recent T time period into n equal parts. Use sensors to collect sample values ​​of the task type factor, data volume, and algorithm complexity of device i in the most recent T time period, and establish a computing power demand sample function model as follows:

[0086]

[0087] in, are the sampling values ​​of task type factor, task data size and task complexity of device i at time t, It is the computing power requirement sample value calculated by device i at time t based on the sampling values ​​of task type factor, task data size, and task complexity.

[0088] for The sample value of computing power demand of device i in time period, It is a sequence of computing power demand sample values ​​composed of computing power demand sample values ​​in n time periods, specifically expressed as follows:

[0089]

[0090] According to the frequency histogram of the computing power resource demand sample values, it is concluded that the computing power resource demand conforms to the Laplace distribution, so the computing power demand sample value sequence of the computing device i in the recent T time period is The mean and variance of are as follows:

[0091]

[0092]

[0093] Based on the computing power demand sample value sequence of device i in the recent T time period The mean and variance of the Laplace distribution model are established as follows:

[0094]

[0095] According to the probability density function, the maximum and minimum predicted values ​​of computing power demand are obtained as follows:

[0096]

[0097]

[0098] The computing power resource requirements of a device have a range. First, determine the range of computing power resource requirements and then determine the specific predicted computing power resource value within this range. The above two formulas are used to solve for the maximum and minimum values ​​within this range. There are countless solutions that can meet the minimum computing power resource requirement. The solution with the largest value among these countless solutions is used as the minimum predicted computing power resource requirement. There are countless solutions that can meet the maximum computing power resource requirement. The solution with the smallest value among these countless solutions is used as the maximum predicted computing power resource requirement.

[0099] Establish a task complexity function and train historical data through artificial neural network prediction method to obtain the task complexity at the current time t And normalized to the task complexity factor The final predicted value of computing power demand is then calculated using the maximum and minimum predicted values ​​of computing power demand. The specific implementation is as follows:

[0100]

[0101]

[0102]

[0103] in, represents the minimum predicted value of computing power required by device i, is the maximum predicted value of computing power demand of device i; formula and Indicates the computing power requirement of device i Range express The probability of express probability; is the predicted value of the final computing power requirement of device i, is the task complexity function, is the task complexity at the current moment t predicted by the artificial neural network method through training historical data, β_Max is the maximum value of the task complexity, For the general After normalization, the task complexity factor, the final computing power demand forecast value is proportional to the task complexity factor. If the predicted task complexity factor If it is 0, the final computing power demand forecast value is like If it is 1, the final computing power demand forecast value is

[0104] In step S104, a cost matrix for allocating resources from each edge node to each device is obtained by using the analytic hierarchy process, taking into account factors such as the distance between the edge node and the device, obstacle density, channel conditions, interference between channels, and resource matching. The specific steps are as follows:

[0105] First, a hierarchical structure model is established, with the comprehensive cost of allocating resources from edge nodes to devices as the highest target layer, the distance between edge nodes and devices, obstacle density, channel conditions, interference between channels, and resource matching as factors considered in the intermediate criterion layer, and the m×n resource allocation methods for m devices and n edge nodes in the computing power network as the lowest solution layer; secondly, a comparison judgment matrix is ​​constructed based on the five factors in the intermediate criterion layer, and the consistency test of the comparison judgment matrix is ​​performed using the single ranking method to construct a consistency matrix, which is expressed as follows: matrix A=(a ij ) 5×5 Indicates that:

[0106]

[0107] The weight values ​​of each factor in the criterion layer are obtained by the arithmetic mean method, which are h1, h2, h3, h4, and h5. The weight values ​​of each factor in the criterion layer constitute the factor weight matrix H = [h1, h2, h3, h4, h5]. Then, the comparison judgment matrix of the m×n resource allocation methods in the solution layer relative to each factor in the criterion layer is constructed respectively, and the consistency test is performed. The weight values ​​of the m×n resource allocation methods in the solution layer relative to each factor in the criterion layer are obtained by the arithmetic mean method. The weight values ​​of the m×n resource allocation methods relative to the distance between the edge node and the device, the obstacle density, the channel condition, the interference between the channels, and the resource matching degree are respectively used. The weight values ​​are then used to construct the solution weight matrix. The solution weight matrix E is expressed as follows:

[0108]

[0109] The cost matrix W is obtained based on the factor weight matrix H and the solution weight matrix E. The matrix is ​​expressed as follows:

[0110] W=H×E[ω 1,1 ,…,ω i,j ,…,ω m,n ],(i=1,2,…,m,j=1,2,…,n)

[0111] Among them, ω i,j,(i=1,2,…,m,j=1,2,…,n) represents the cost weight of edge node j allocating resources to device i. The larger the cost weight, the greater the distance between edge node j and device i, the higher the obstacle density, the worse the channel condition, the greater the interference between channels and the lower the resource matching degree.

[0112] Step S105: Based on the computing resources allocated by the edge nodes to the devices and the cost matrix of the resources allocated by each edge node to each device, a computing resource optimization function model is established. The specific method is as follows:

[0113]

[0114]

[0115]

[0116] Among them, Y t* The optimal solution vector for the computing resource optimization model, is the optimal computing power resource allocated by edge node j to device i at time t; the optimization target is the total computing power resource allocated to the edge node, Represents the cost weight ω i,j Impact on the computing power resources allocated by edge node j to device i Cost weight ω i,j The larger it is, the smaller the computing power resources to be allocated will be. Indicates that the total computing power resources that edge node j can allocate to the device are limited. is the maximum computing power resource of edge node j, and the constraint condition Indicates that the total computing power resources received by device i from each edge node are allocated according to the device's computing power demand prediction value, ψ i (ε) is the relaxation factor representing the degree of relaxation of computing resource requirements of device i, if ψ i If (ε) is 0, it means that the computing power resources allocated to the device must strictly meet the computing power requirements of the device.

[0117] Step S106, as Figure 3 As shown in the figure, the computing power resource optimization model is solved based on the simulated annealing algorithm and relaxation method. i (ε) function value is set to change the degree of relaxation between resource allocation and computing power requirements, thereby achieving the optimal allocation of computing power resources. The specific steps are as follows:

[0118] S1: Initialization The parameter ε=0.0001, Iteration steps τ=0,ψ i(ε) is a relaxation factor representing the degree of relaxation of the computing power resource requirements of device i.

[0119] S2: Use the simulated annealing algorithm to solve the computing power resource optimization function model.

[0120] S3: If there is a solution in step S2, output the optimal solution vector Y. t* ; If there is no solution, jump to step S4.

[0121] S4: Let τ = τ + 1, reassign the relaxation factor. Each device has different requirements for latency. Compare the required latency of each device, and relax the constraint conditions of the computing power resource demand relaxation degree from high to low according to the required latency of the device. By changing the value of the relaxation factor, change the computing power demand resource constraint conditions. The larger the value of the relaxation factor, the more relaxed the computing power resource demand, and the larger the floating space of the allocated computing power resource around the demand value; recalculate. [[ID=IS=17]]

[0122] S5: Judge the iteration step τ. If τ > L, end the iteration and the computing power resource offloading fails, where L is the maximum number of iteration steps; if τ < L, jump to step S2. <IS=

[0123] The present invention also proposes an edge DPU computing power collaborative optimization offloading system with demand adaptive prediction. The system includes:

[0124] A computing power demand function model construction unit for establishing a computing power demand function model between the computing power resources required by a device and the task type, data volume size, algorithm complexity of the computing tasks proposed by the device, and the DPU capabilities of the device itself, and improving it by introducing environmental factors of the environment where the device is located.

[0125] A computing power resource demand prediction unit for establishing a probability density function of the Laplace distribution model according to the historical data of the computing power resource demand of the device and predicting the computing power resource demand of the device.

[0126] A cost matrix construction unit for establishing a cost matrix for each edge node to allocate resources to each device.

[0127] A computing power resource optimization function model construction unit for establishing a computing power resource optimization function model.

[0128]

[0128] A computing power resource allocation unit for solving the computing power resource optimization model and achieving the best allocation of computing power resources.

[0129] The technical solution of the above system is similar to the foregoing edge DPU computing power collaborative optimization offloading method with demand adaptive prediction, and will not be elaborated here.

[0130] Based on the same technical solution, the present invention also provides a device, comprising:

[0131] Memory;

[0132] processor;

[0133] as well as

[0134] computer programs;

[0135] The computer program is stored in the memory and is configured to be executed by the processor to implement the aforementioned method of collaborative optimization and unloading of edge DPU computing power based on demand adaptive prediction.

[0136] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the aforementioned edge DPU computing power collaborative optimization unloading method for demand adaptive prediction.

[0137] It should be understood that the specific examples in the present invention are only intended to help those skilled in the art better understand the embodiments of this specification, rather than to limit the scope of the present invention.

[0138] It can be understood that in the various implementations of this specification, the size of the serial number of each process 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 implementation methods of this specification.

[0139] It can be understood that the various embodiments described in this specification can be implemented individually or in combination, and the embodiments in this specification are not limited to this.

[0140] Unless otherwise indicated, all technical and scientific terms used in the embodiments of this specification have the same meaning as those commonly understood by those skilled in the art in the technical field of this specification. The terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit the scope of this specification. The term "and / or" used in this specification includes any and all combinations of one or more related listed items. The singular forms "a", "above", and "the" used in the embodiments of this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

[0141] It is understood that the processor in the embodiments of this specification can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiment can be completed by hardware integrated logic circuits in the processor or software instructions. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The various methods, steps, and logic block diagrams disclosed in the embodiments of this specification can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this specification can be directly implemented as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0142] It will be understood that the memory in the embodiments of this specification may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0143] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0144] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems and units can refer to the corresponding processes in the aforementioned method implementation, and will not be repeated here.

[0145] In the several embodiments provided in this specification, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0146] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of this embodiment.

[0147] In addition, each functional unit in each embodiment of this specification may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0148] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this specification, or the part that contributes to the prior art, or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this specification. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0149] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. The edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction is characterized by: The following steps are involved: Step 1: Establish a computing power demand function model based on the computing power resources required by the device and the task type, data volume, algorithm complexity, and DPU capability of the device's computing task. Step 2: Introduce environmental factors of the device environment to improve the computing power demand function model; Step 3: Based on the historical data of computing resource requirements of the device, a probability density function of the Laplace distribution model is established to predict the computing resource requirements of the device; Step 4: Establish a cost matrix for allocating resources from each edge node to each device; Step 5: Establish a computing power resource optimization function model; Step 6: Solve the computing power resource optimization model to achieve the optimal allocation of computing power resources; The improved computing power demand function model in step 2 is specifically expressed as follows: ， in, is the computing resource requirement of the improved device i at time t, is the environmental factor of device i at time t, is the temperature of the environment where device i is located at time t, is the humidity of the environment where device i is located at time t, is the power stability of the environment where device i is located at time t, is the network bandwidth of the environment where device i is located at time t, 、 、 、 is a parameter; The probability density function of the Laplace distribution model in step 3 is specifically expressed as follows: ， in express The probability density of for The sample value of computing resource demand of device i in the time period, This is the collection period for historical data on computing resource requirements. is the number of equal parts of the collection period, and They are The mean and variance of According to the probability density function, the maximum and minimum predicted values ​​of the computing power resource requirements of the device are obtained, which are specifically expressed as follows: ， ， in, and are the maximum and minimum predicted values ​​of the computing resource requirements of the device, Represents probability, parameter ; Establish a task complexity function and use the artificial neural network prediction method to train the historical data of computing resource requirements to obtain the task complexity at time t. , and normalized to the task complexity factor , and then calculate the final predicted value of the computing power resource demand of the device through the maximum predicted value and minimum predicted value of the computing power resource demand of the device. The specific implementation is as follows: ， ， ， in, is the final predicted value of computing resource demand of device i at time t, is the task type of the computing task proposed by device i at time t-1, is the data size of the computing task proposed by device i at time t-1; is the algorithm complexity of the computing task proposed by device i at time t-1, is the maximum value of task complexity.

2. The edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction according to claim 1 is characterized in that: The computing power demand function model in step 1 is specifically expressed as follows: ， in, is the computing resource requirement of device i at time t; Indicates the computing resource requirement value for a fixed task type, data volume, algorithm complexity, and device DPU capability; is the task type of the computing task proposed by device i at time t; is the data size of the computing task proposed by device i at time t; is the algorithm complexity of the computing task proposed by device i at time t; is the DPU capability of device i.

3. The edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction according to claim 1 is characterized in that: In step 4, the cost matrix for allocating resources from each edge node to each device is established through the analytic hierarchy process. The specific steps are as follows: A hierarchical structure model is established, with the comprehensive cost of allocating computing resources from edge nodes to devices as the highest target layer, and the distance between edge nodes and devices, obstacle density, channel conditions, interference between channels, and resource matching as factors considered in the intermediate criterion layer. The resource allocation method is used as the lowest scheme layer; the five factors of the intermediate criterion layer are used to construct a comparative judgment matrix; the consistency test of the comparative judgment matrix is ​​carried out using the single ranking of the levels, and the consistency matrix is ​​constructed. Indicates that: ， The factor weights of each factor in the criterion layer are obtained by arithmetic mean method and recorded as , the weight matrix of the constituent factors ; Construct solution layers separately Compare the resource allocation methods with the factors of the criterion layer and conduct consistency test; calculate the solution layer by arithmetic average method. The weights of the resource allocation methods relative to the factors in the criterion layer are recorded as 、 、 、 、 , construct the scheme weight matrix E, which is specifically expressed as follows: ； According to the factor weight matrix And the solution weight matrix E, the cost matrix W is obtained, which is specifically expressed as follows: ， in, Represents the cost weight of edge node j allocating computing resources to device i.

4. The edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction according to claim 1 is characterized in that: The computing power resource optimization function model in step 5 is specifically expressed as follows: ， ， ， in, is the optimal solution vector, is the optimal computing resource allocated by edge node j to device i at time t; Indicates the amount of computing resources allocated by edge node j to device i at time t, is the maximum computing power resource of edge node j, is the relaxation factor that represents the degree of relaxation of computing resource requirements of device i.

5. The edge DPU computing power collaborative optimization offloading method based on demand adaptive prediction according to claim 1 is characterized in that: In step 6, the computing power resource optimization function model is solved based on the simulated annealing algorithm and relaxation method. The specific steps are as follows: S1: Initialization , where the parameters , , the number of iterations , is the relaxation factor that represents the degree of relaxation of computing resource requirements of device i, S2: Use the simulated annealing algorithm to solve the computing power resource optimization function model; S3: If there is a solution in step S2, then output the optimal solution vector ; If there is no solution, jump to step S4; S4: Command , recalculate ; S5: number of iterations Make a judgment, if , then the iteration ends and the computing resource unloading fails, where L is the maximum number of iteration steps; if , jump to step S2.

6. A system for collaborative optimization and offloading of edge DPU computing power based on the method for adaptive demand prediction according to any one of claims 1 to 5, characterized in that: The system comprises: The computing power demand function model building unit is used to establish a computing power demand function model between the computing power resources required by the device and the task type, data volume, algorithm complexity and DPU capability of the computing task proposed by the device, and to improve it by introducing environmental factors of the environment in which the device is located; A computing resource demand prediction unit is used to establish a probability density function of a Laplace distribution model based on the historical computing resource demand data of the device to predict the computing resource demand of the device; A cost matrix construction unit is used to establish a cost matrix for allocating resources to each device at each edge node; A computing power resource optimization function model building unit is used to build a computing power resource optimization function model; The computing power resource allocation unit is used to solve the computing power resource optimization model to achieve the optimal allocation of computing power resources.

7. A device, characterized in that include: Memory; processor; as well as computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method according to any one of claims 1 to 5.

8. A storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images