A workflow scheduling method based on multi-neural networks for edge intelligence

Abstract

The present invention discloses a multi-neural network workflow scheduling method for edge intelligence. The training tasks of multiple CNNs are accelerated by offloading the computing tasks from the edge device (MD) to the edge node server (eNB), and the CPU and GPU of the MD are dynamically adjusted to minimize the system training delay and the energy consumption of the MD. The present invention designs a multi-neural network computing workflow scheduling method for edge intelligence, which can enhance the computing power of the terminal device by offloading the multi-neural network training tasks to the edge nodes with sufficient resources, thereby reducing delays and saving energy. In addition, an adaptive strategy is proposed to obtain the inertia weight parameter w, so as to better adapt to the directed acyclic graph DAG structure of the DNN model and improve the search ability of the neural network DNN.

Description

Technical Field

[0001] The present invention belongs to the technical field of workflow task scheduling, and relates to a workflow scheduling method for implementing multi-neural networks for edge intelligence. Aiming at edge computing scenarios, a DeepINT method for solving multi-neural network training tasks in edge computing is proposed, that is, the training tasks of multiple CNNs are accelerated by offloading computing tasks from edge devices (MDs) to edge node servers (eNBs), and the CPU and GPU of MDs are dynamically frequency-adjusted to minimize system training delays and energy consumption of MDs. Background Art

[0002] Artificial intelligence (AI) technology, especially deep neural networks (DNNs), has achieved state-of-the-art performance and is widely used in various application fields. Recently, with the emergence of high-performance system-on-chip (SoC), running DNN applications on mobile devices and IoT devices has gained great popularity and attention. However, it is challenging to perform all computationally intensive DNN inference tasks on resource-limited terminal devices, which may result in unbearable latency and energy consumption. To alleviate the above problems, the emergence of the edge computing paradigm, which deploys computing, storage, and network resources to the edge of the network, can accelerate DNN inference by offloading part of the computing tasks to edge nodes.

[0003] Edge devices are currently capable of completing CNN tasks, but because edge devices have limited memory, computing resources, and power processing capabilities, they cannot meet the needs of completing multiple neural network tasks, saving computing resources, and reducing training delays in actual scenarios. This patent proposes a multi-neural network task workflow scheduling method in edge intelligence scenarios, based on the meta-heuristic optimization technology particle swarm optimization (PSO), and provides possible scheduling schemes from workflow to task execution location, CPU frequency, and GPU frequency to minimize latency and energy consumption in edge computing. Summary of the invention

[0004] The purpose of the present invention is to propose a multi-neural network workflow scheduling method for edge intelligence in an environment where multiple neural network model training tasks are executed in an edge computing scenario.

[0005] Edge devices are currently capable of completing CNN tasks. However, because edge devices have limited memory, computing resources, and power processing capabilities, they cannot meet the needs of completing multiple neural network tasks, saving computing resources, and reducing training delays in actual scenarios. This paper proposes a multi-neural network task workflow scheduling method DeepINT in edge intelligence scenarios. It is based on the meta-heuristic optimization technology particle swarm optimization (PSO) and provides possible scheduling schemes from workflow to task execution location, CPU frequency, and GPU frequency to minimize delays and energy consumption in edge computing.

[0006] In a first aspect, the present invention provides a method for implementing workflow scheduling using a multi-neural network for edge intelligence, comprising the following steps:

[0007] Step (1), establishing an optimal workflow scheduling model under resource constraints;

[0008] Step (2), initializing the workflow;

[0009] Step (3), use the particle swarm algorithm to perform optimization iterations to obtain the optimal solution of the optimal workflow scheduling model. By comparing these optimal solutions, the optimal unloading point of the multi-neural network DNNs and the optimal operating frequency of the CPU and GPU are obtained; the interaction in the particle swarm is based on the topological neighborhood, and each particle The vector of the best neighbor Communicate to obtain the vector The fitness value gbest is set, and corresponding adjustments are made in its individual behaviors;

[0010] In the process of particle swarm optimization, the iteration is repeated until the maximum number of iterations is reached or the fitness value reaches the preset threshold, and the optimal particle is output; in each iteration, the speed and position of the particle are expressed by equations (1)-(2) respectively:

[0011]

[0012]

[0013] Where w represents the inertia weight parameter; c1 represents the individual learning factor, c2 represents the group learning factor; r1 and r2 represent random numbers;

[0014] Adaptive strategies are used to adjust the spatial search capability of the front particles. The inertia weight parameter w is obtained according to the network structure of a single neural network DNN and the mass of the particle to better adapt to the DAG structure of the DNN model. The search capability of the particle is adaptively adjusted according to the difference between the global optimal particle and the current particle. The specific calculation formula is as follows:

[0015]

[0016] in Indicates the global optimal particle position gbest and the current particle The spatial coordinate difference between max and w min They represent the maximum and minimum values ​​of w given in the initial stage respectively;

[0017] Step (4): According to the optimal workflow offloading solution obtained in step (3), part of the training tasks are offloaded from the edge device MD to the edge node server eNB, and the computing tasks are performed according to the obtained optimal operating frequency of the CPU and the optimal operating frequency of the GPU, so as to optimize the computing delay of the multi-neural network DNNs training tasks and the energy consumption of the edge device MD.

[0018] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.

[0019] In a third aspect, the present invention provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the described method is implemented.

[0020] Beneficial effects of the present invention:

[0021] The present invention designs a multi-neural network computing workflow scheduling method for edge intelligence. By offloading multi-neural network training tasks to edge nodes with sufficient resources, the computing power of terminal devices can be enhanced, thereby reducing delays and saving energy. In addition, an adaptive strategy is proposed to obtain the inertia weight parameter w to better adapt to the directed acyclic graph DAG structure of the DNN model and improve the search ability of the neural network DNN.

[0022] The present invention provides a solution with just the right time performance and lower energy consumption of edge devices according to the actual needs of complex scenarios of training multiple neural network models in edge computing scenarios. The present invention proposes the DeepINT method, which is a method of accelerating multi-DNN training by offloading calculations from edge devices to edge nodes while reducing system energy consumption. The results show that DeepINT is superior to existing systems, reduces joint reasoning latency and energy consumption, and improves the feasibility and efficiency of joint reasoning in edge intelligence scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A simple workflow model is formed for the process of jointly computing CNNs tasks between MD and eNB.

[0024] Figure 2 This is the workflow scheduling diagram obtained by the DeepINT method.

[0025] Figure 3 The figure shows the MD energy consumption comparison of DeepINT method, Server-only method, and Local-only method.

[0026] Figure 4Comparison chart of total latency calculation for DeepINT method, Server-only method, and Local-only method tasks.

[0027] Figure 5 This is the relationship diagram between the DeepINT method control parameter V and energy consumption and inference delay. DETAILED DESCRIPTION

[0028] The present invention will be further analyzed below in conjunction with the accompanying drawings.

[0029] In edge computing scenarios, we can use eNB to assist multiple MDs in joint model training for DNNs computing tasks. eNB is geographically close to most MDs and provides relatively reliable short-distance transmission, which can significantly reduce the system communication cost in terms of data collection and model training.

[0030] A multi-neural network workflow scheduling method for edge intelligence (DeepINT), the steps are:

[0031] Step (1), establishing an optimal workflow scheduling model under resource constraints;

[0032] The simple workflow model formed by the joint reasoning and calculation of CNNs by MD and eNB is shown in the attached figure. Figure 1 As shown in the figure. The empty nodes at the beginning of the figure represent the input data of multiple CNNs, and the empty nodes at the end represent the collection of calculation results of multiple CNNs. The convolution kernel Conv Kernel of the model is used as the task node. The calculation process of task nodes t1, t2, t3, and t4 represents the complete reasoning process of model CNN1, and the calculation process of task nodes t5 and t6 represents the complete reasoning process of model CNN2. Combined with the structure of CNN's directed acyclic graph (DAG), each task node is constrained by all its predecessor task nodes before execution.

[0033] 1.1. Energy consumption model for completing multi-neural network DNNs training tasks on edge devices and edge node servers:

[0034] 1.1.1. The edge device MD is equipped with two CPU architectures: high-performance CPU and low-performance CPU. In order to minimize power consumption, it is assumed that all low-performance CPU cores and high-performance CPU cores have the same operating frequency; the CPU computing power P of this edge device MD is CPu It is expressed as the following formula:

[0035] P CPU =P L +P H Formula (4)

[0036]

[0037]

[0038] Among them, α L and α H are two constants related to the power of low-performance CPU and high-performance CPU, respectively. l represents the operating frequency of the lth low-performance CPU core, f h represents the operating frequency of the hth high-performance CPU core; N L Indicates the number of low-performance CPUs, N H Indicates the number of high-performance CPUs, P L Represents the computing power of low-performance CPU, P H Indicates the computing power of a high-performance CPU;

[0039] 1.1.2. Not only is CPU computing power related to frequency, but its computing power is also controlled by frequency; therefore, the relationship between a single-core CPU computing power and CPU frequency is expressed by the following formula:

[0040] c=βf c Formula (7)

[0041] Among them, c represents the computing power of a single CPU core, that is, floating-point operations, β represents the floating-point calculation value of a single CPU cycle, and f c Indicates the operating frequency of a single CPU core; therefore, the CPU computing power of the edge device MD is expressed by the following formula:

[0042] C MD =β L N L f l +β H N H f h Formula (8)

[0043] Among them, β L Represents the floating-point calculation value of a low-performance CPU in a single cycle; β H Represents the floating-point calculation value of a single cycle of a high-performance CPU;

[0044] 1.1.3. The CPU architecture and GPU architecture design are very different, so the GPU computing power P of the edge device MD is GPU The calculation method is completely different. The following formula is used to represent GPU computing power:

[0045] P GPu =ηU G f G Formula (9)

[0046] Where η represents the coefficient of energy conversion efficiency; f G Indicates the frequency of the GPU, U G Indicates the operating voltage;

[0047] The GPU computing capability of the edge device MD is expressed as:

[0048] c G =γf G Formula (10)

[0049] Where γ represents the number of floating-point calculations that the GPU can process in a single period;

[0050] 1.1.4. Time for edge device MD to execute tasks for:

[0051]

[0052] in, Indicates the workload processed by the edge device MD;

[0053] Therefore, the total energy consumption E consumed in the process of transmitting data from the edge device MD to the edge node server eNB is expressed as:

[0054]

[0055] Among them, P CPU Indicates the computing power of the CPU; P GPU Indicates the computing power of the GPU; The time it takes to transmit the data with integrity on the MD;

[0056] 1.2. When performing terminal-edge collaborative training, the kth edge device MDk trains I in the multi-neural network DNNs k layer sub-model, while the edge node server eNBk trains the remaining II in the multi-neural network DNNs k Layer sub-model; I represents the number of sub-model layers of multiple neural networks DNNs; therefore, the training delay model construction process is as follows:

[0057] 1.2.1、Build edge device MD local execution training time model:

[0058] Assume that the edge device MDk performs the training computation time locally Layer I of the sub-model with multiple neural networks DNNs k , batch size is b k , local data size d k and GPU frequency f k and memory μ k Impact on computing power The heterogeneity of edge device MD computing power indicates that different edge devices MD have different local computing power; assuming that edge device MD k The batch size for local training is b k , MD k The total batch is B k , whose expression is:

[0059]

[0060] 1.2.2. Training data transmission time modeling:

[0061] In the terminal-edge collaborative training process, after the edge device MD completes local training, the edge device MD uploads its own training result data to the edge node server eNB; the edge node server eNB continues to train the current edge device MD to train the remaining II k layer, and then the edge node server eNB sends the corresponding gradient back to the current edge device MD;

[0062] Therefore, the edge device MD k The communication overhead in the uplink and downlink data transmission process of each round of training is defined as UL(I k ) and DL(I k ); therefore, the training data transmission time It is expressed as:

[0063]

[0064] where r k Indicates edge device MD k The network bandwidth used to transmit data;

[0065] 1.2.3. Construct edge node server eNB execution training time model:

[0066] The edge node server eNB will evenly allocate a fixed amount of computing resources to each heterogeneous edge device MD to ensure fairness among edge devices MDs; therefore, the edge node server eNB performs training computing time It is expressed as:

[0067]

[0068] in Indicates that the edge node server eNB is assigned to the MD k The computational amount of the bth batch of Indicates that the edge node server eNB is training its remaining II k Layer DNN model for MD k Allocated computing power;

[0069] 1.2.4. Build a full model training delay model:

[0070] When performing device-edge co-training, MD k Training has I k The device-side sub-model of the layer, while the edge node server eNB trains the remaining II k Layer; for each edge device MD, the full model training delay T in each batch training k It is expressed as:

[0071]

[0072] in, represents the computation time of the edge device MD performing training locally, represents the training data transmission time, Indicates the computing time of edge node server eNB to perform training;

[0073] 1.3 In order to efficiently perform multi-neural network DNNs computing tasks in edge computing scenarios, improve computing efficiency and reduce energy consumption; therefore, the optimization problem P is expressed as follows:

[0074]

[0075] Where V represents the control parameter that balances low energy consumption and low latency, V∈[0.01, 100]; represents the total number of tasks, x i Indicates the execution location of the task, x i = 0 means that the i-th task is executed on the edge device MD, x i =1 means that the i-th task is offloaded to the edge node server eNB for execution;

[0076] Step (2), initializing the workflow;

[0077] The workflow composed of the initial tasks all executes tasks on the edge device MD, and randomly sets the mutation rate mR for each task, mR∈[0,1];

[0078] When initializing the workflow, determine whether the predecessor task of the current task is executed on the edge node server eNB or the mutation rate exceeds the threshold. If so, the current task will be assigned to the edge node server eNB for execution. Otherwise, the current task is executed on the edge device MD and the successor task is executed on the edge node server eNB, and the task becomes a task offloading point for DNNs;

[0079] The objective function (17) is a function of a nonlinear, non-differentiable optimization problem, so a multi-dimensional coding strategy is used to solve the problem; specifically:

[0080] Initialize the number of particles N in the workflow p , the dimension of the particle is Dim, each particle consists of 3 vectors with Dim dimension, namely the current position of the particle Previous best position and speed And the dimension of the particle Dim is the product of the number of parameters in the objective function (17) and the number of tasks in the workflow; the current position represents a candidate solution to the optimization problem;

[0081] In this experiment, three parameters are used (network bandwidth r, operating frequency f of MD's high-performance CPU and low-performance CPU of equal size, and MD's GPU operating frequency f G ) and the number of workflow tasks is 6, then the value of Dim is 18, so the position of the particle is defined by 18 coordinates. Current position represents a candidate solution to the optimization problem. If the current position is better than any position so far, the current position is stored in the vector In addition, the value of the best fitness function so far is stored in pbest k middle.

[0082] Step (3), use the particle swarm algorithm to perform optimization iterations to obtain the optimal value of the objective function (17). By comparing these optimal values, the optimal unloading point of the multi-neural network DNNs and the optimal operating frequency of the CPU and GPU are obtained; the interaction in the particle swarm is based on the topological neighborhood, and each particle The vector of the best neighbor Communicate to obtain the vector The fitness value gbest is set, and corresponding adjustments are made in its individual behaviors;

[0083] In the process of particle swarm optimization, the iteration is repeated until the maximum number of iterations is reached or the fitness value reaches the preset threshold, and the optimal particle is output; in each iteration, the speed and position of the particle are expressed by equations (18)-(19) respectively:

[0084]

[0085]

[0086] Where w represents the inertia weight parameter, w∈[0.4, 2]; c1 represents the individual learning factor, c2 represents the group learning factor, c1, c2∈[0, 4]; r1 and r2 represent random numbers, the purpose is to increase the randomness of spatial search, r1, r2∈[0, 1];

[0087] The present invention improves the traditional linear change strategy, adopts an adaptive strategy to adjust the spatial search capability of the front particle, obtains the inertia weight parameter w according to the network structure of a single neural network DNN and the mass of the particle, so as to better adapt to the directed acyclic graph DAG structure of the DNN model; this method can adaptively adjust the search capability of the particle according to the difference between the global optimal particle and the current particle; when the difference between the two is small, the algorithm will enhance the local search capability to accelerate the convergence speed and find the optimal solution; the specific calculation formula is as follows:

[0088]

[0089] in Indicates the global optimal particle position gbest and the current particle The spatial coordinate difference between max and w min They represent the maximum and minimum values ​​of w given in the initial stage respectively;

[0090] As attached Figure 2 As shown in , the x of each task i The value of the coordinate corresponds to an index. The first coordinate of each task indicates the location where the corresponding task is executed. When the coordinate value is 0, it means that the task is executed on the MD, and when the coordinate value is 1, it means that the task is executed on the eNB. The second coordinate indicates the frequency of the CPU of the current execution device; the third coordinate indicates the frequency of the GPU of the current execution device. Possible scheduling schemes from workflow to task execution location, CPU frequency, and GPU frequency are given.

[0091] Step (4): According to the optimal workflow offloading solution obtained in step (3), part of the training tasks are offloaded from the edge device MD to the edge node server eNB, and the computing tasks are performed according to the obtained optimal operating frequency of the CPU and the optimal operating frequency of the GPU, so as to optimize the computing delay of the multi-neural network DNNs training tasks and the energy consumption of the edge device MD.

[0092] The present invention adopts DeepINT method, Server-only method and Local-only method to compare MD in terms of energy consumption and system delay.

[0093] 1) Comparison of energy consumption of three methods

[0094] In the experiment, we set the network bandwidth to fluctuate between 10 and 100 Mbps, and keep several other variables unchanged. We set the CPU frequency of MD to 1.8 GHz, the computing power of MD GPU to 2.6 GFLOPs, the CPU frequency of eNB to 3.6 GHz, the computing power of eNB GPU to 4.4 GFLOPs, and the control parameter V to 1. The comparison of MD energy consumption of the three methods under different network bandwidths is shown in the figure below. Figure 3 As shown, we can see that in the bandwidth range of 10 to 20 Mbps, compared with the Local-only and Server-only methods, the DeepINT method performs well in MD energy consumption. When the bandwidth is 10 Mbps, the average energy consumption of the DeepINT method when performing task reasoning in MD is 45% lower than that of the Server-only method; in the range of 30 to 100 Mbps, the average energy consumption of the DeepINT method when performing task reasoning in MD is more than 50% lower than that of the Local-only method, and the difference with the Server-only method is small.

[0095] 2) Comparison of the total system delay of the three methods

[0096] The total delay of task reasoning of the three methods under different network bandwidths is shown in the figure below. Figure 4 As shown, we can see that when the bandwidth is 10 to 100Mbps, the total reasoning delay of the DeepINT method is better than the other two methods, and it can be seen that under the current experimental setting parameters, the DeepINT method has a certain stability. As the bandwidth increases linearly, the total delay of task reasoning decreases linearly. Therefore, this method performs well in the case of network bandwidth fluctuations. However, due to the task integrity transmission mechanism of the Server-only method and the large amount of initial data transmitted, this method is greatly affected by the network bandwidth when the bandwidth is small. When the network bandwidth is above 40Mbps, the DeepINT method reduces the total delay of task reasoning by at least 25% compared with the Local-only method and at least 10% compared with the Server-only method.

[0097] 3) Selection of control parameter V

[0098] Compare the control parameters of the DeepINT method in terms of energy consumption and inference latency. Figure 5As shown in the figure, as V changes from 0.01 to 10, the energy consumption of DeepINT is gradually increasing, and the rate of increase is also increasing, while the average latency of task reasoning is gradually decreasing. When the V value is small, the system focuses on the energy consumption of MD, resulting in high task latency but low energy consumption. On the contrary, when the V value is very large, it indicates that the system emphasizes the task reasoning latency of the system rather than the device energy consumption. This shows that the control parameter V can indeed achieve a balance between MD energy consumption and reasoning latency in the system. Therefore, an appropriate V can be selected in actual scenarios to meet customized energy consumption and latency requirements.

[0099] The experimental results show that the DeepINT method of the present invention has excellent performance, can significantly reduce energy consumption and improve system reasoning delay, and can be applied to most real-world scenarios. In addition, the method designs a control parameter V, which can balance energy consumption and delay according to actual needs in the usage scenario. From the above two performances, regardless of the network bandwidth transmission rate, the performance of the DeepINT method is better than the other two methods, especially when the network fluctuates greatly, the DeepINT method shows excellent stability.

[0100] Table 1 Energy consumption of different methods under different bandwidths (J)

[0101]

[0102]

[0103] Table 2 Total system delay of different methods under different bandwidths (s)

[0104]

[0105] Table 3 Comparison of average energy consumption (J) and average delay (s) of DeepINT method under different control parameters V

[0106]

Claims

1. A multi-neural network workflow scheduling method for edge intelligence, characterized in that The method comprises the following steps: Step (1), establishing an optimal workflow scheduling model under resource constraints; 1.

1. Build an energy consumption model for completing multi-neural network DNNs training tasks on edge devices and edge node servers; 1.

2. When performing terminal-edge collaborative training, the kth edge device MD k Training multiple neural networks (DNNs) k Layer sub-model, and the edge node server eNB k Training Multiple Neural Networks (DNNs) in Residual II k Layer sub-model; I represents the number of sub-model layers of multiple neural networks DNNs; 1.

3. In order to efficiently perform multi-neural network DNNs computing tasks in edge computing scenarios, improve computing efficiency and reduce energy consumption; therefore, the optimization problem P is expressed as follows: Where V represents the control parameter that balances low energy consumption and low latency; represents the total number of tasks, x i Indicates the execution location of the task, x i = 0 means that the i-th task is executed on the edge device MD, x i =1 means that the i-th task is offloaded to the edge node server eNB for execution; Step (2), initializing the workflow; Step (3), use the particle swarm algorithm to perform optimization iterations to obtain the optimal value of the objective function (1). By comparing these optimal values, the optimal unloading point of the multi-neural network DNNs and the optimal operating frequency of the CPU and GPU are obtained; the interaction in the particle swarm is based on the topological neighborhood, and each particle The vector of the best neighbor Communicate to obtain the vector The fitness value gbest is set, and corresponding adjustments are made in its individual behaviors; In the process of particle swarm optimization, the iteration is repeated until the maximum number of iterations is reached or the fitness value reaches the preset threshold, and the optimal particle is output; in each iteration, the speed and position of the particle are expressed by equations (2)-(3) respectively: Where w represents the inertia weight parameter; c1 represents the individual learning factor, c2 represents the group learning factor; r1 and r2 represent random numbers; Adaptive strategies are used to adjust the spatial search capability of the front particles. The inertia weight parameter w is obtained according to the network structure of a single neural network DNN and the mass of the particle to better adapt to the DAG structure of the DNN model. The search capability of the particle is adaptively adjusted according to the difference between the global optimal particle and the current particle. The specific calculation formula is as follows: in Indicates the global optimal particle position gbest and the current particle The spatial coordinate difference between max and w min They represent the maximum and minimum values ​​of w given in the initial stage respectively; Step (4): According to the optimal workflow offloading solution obtained in step (3), part of the training tasks are offloaded from the edge device MD to the edge node server eNB, and the computing tasks are run according to the obtained optimal operating frequency of the CPU and the optimal operating frequency of the GPU, so as to optimize the computing delay of the multi-neural network DNNs training tasks and the energy consumption of the edge device MD.

2. The method according to claim 1, characterized in that Step 1.1 is as follows: 1.1.

1. The edge device MD is equipped with two CPU architectures: high-performance CPU and low-performance CPU. In order to minimize power consumption, it is assumed that all low-performance CPU cores and high-performance CPU cores have the same operating frequency; the CPU computing power P of this edge device MD is CPU It is expressed as the following formula: P CPU =P L +P H Formula (5) in, α L and α H are two constants related to the power of low-performance CPU and high-performance CPU, respectively. l represents the operating frequency of the lth low-performance CPU core, f h represents the operating frequency of the hth high-performance CPU core; N L Indicates the number of low-performance CPUs, N H Indicates the number of high-performance CPUs, P L Represents the computing power of low-performance CPU, P H Indicates the computing power of a high-performance CPU; 1.1.

2. Not only is CPU computing power related to frequency, but its computing power is also controlled by frequency; therefore, the relationship between a single-core CPU computing power and CPU frequency is expressed by the following formula: c=βf c Formula (8) in, c represents the computing power of a single CPU core, i.e. floating-point operations, β represents the floating-point calculation value of a single CPU cycle, and f c Indicates the operating frequency of a single CPU core; therefore, the CPU computing power of the edge device MD is expressed by the following formula: c MD =b L N L f l +b H N H f h expression(9) Among them, β L Represents the floating-point calculation value of a low-performance CPU in a single cycle; β H Represents the floating-point calculation value of a single cycle of a high-performance CPU; 1.1.

3. The CPU architecture and GPU architecture design are very different, so the GPU computing power P of the edge device MD is GPU The calculation method is completely different. The GPU computing power is expressed using formula (10): P GPU =ηU G f G Formula (10) Where η represents the coefficient of energy conversion efficiency; f G Indicates the frequency of GPU, U G Indicates the operating voltage; The GPU computing capability of the edge device MD is expressed as: c G =γf G Formula (11) Where γ represents the number of floating-point calculations that the GPU can process in a single period; 1.1.

4. Time for edge device MD to execute tasks for: in, Indicates the workload processed by the edge device MD; Therefore, the total energy consumption E consumed in the process of transmitting data from the edge device MD to the edge node server eNB is expressed as: Among them, P CPU Indicates the computing power of the CPU; P GPU Indicates the computing power of the GPU; Indicates the time taken to transmit the integrity data on the MD.

3. The method according to claim 1, characterized in that The delay model construction process in step 1.2 is as follows: 1.2.1、Build edge device MD local execution training time model: Assume that the edge device MD k Compute time for local training execution Layer I of the sub-model with multiple neural networks DNNs k , batch size is b k , local data size d k and GPU frequency f k and memory μ k Impact on computing power The heterogeneity of edge device MD computing power indicates that different edge devices MD have different local computing power; assuming that edge device MD k The batch size for local training is b k , MD k The total batch is B k , whose expression is: 1.2.

2. Training data transmission time modeling: In the terminal-edge collaborative training process, after the edge device MD completes local training, the edge device MD uploads its own training result data to the edge node server eNB; The edge node server eNB continues to train the current edge device MD to train the remaining II k layer, and then the edge node server eNB sends the corresponding gradient back to the current edge device MD; Therefore, the edge device MD k The communication overhead in the uplink and downlink data transmission process of each round of training is defined as UL(I k ) and DL(I k ); therefore, the training data transmission time It is expressed as: where r k Indicates edge device MD k The network bandwidth used to transmit data; 1.2.

3. Construct edge node server eNB execution training time model: The edge node server eNB will evenly allocate a fixed amount of computing resources to each heterogeneous edge device MD to ensure fairness among edge devices MDs; therefore, the edge node server eNB performs training computing time It is expressed as: in Indicates that the edge node server eNB is assigned to the MD k The computational amount of the bth batch of Indicates that the edge node server eNB is training its remaining II k Layer DNN model for MD k Allocated computing power; 1.2.

4. Build a full model training delay model: When performing device-edge co-training, MD k Training has I k The device-side sub-model of the layer, while the edge node server eNB trains the remaining II k Layer; for each edge device MD, the full model training delay T in each batch training k It is expressed as: in, represents the computation time of the edge device MD performing training locally, represents the training data transmission time, Indicates the computing time of the edge node server eNB to perform training.

4. The method according to claim 1, characterized in that Step (2) is specifically: The workflow composed of the initial tasks all executes tasks on the edge device MD, and randomly sets the mutation rate mR for each task, mR∈[0,1]; When initializing the workflow, determine whether the predecessor task of the current task is executed on the edge node server eNB or the mutation rate exceeds the threshold. If so, the current task will be assigned to the edge node server eNB for execution. Otherwise, the current task is executed on the edge device MD and the successor task is executed on the edge node server eNB, and the task becomes a task offloading point for DNNs; The objective function (1) is a function of a nonlinear, non-differentiable optimization problem, so a multi-dimensional coding strategy is used to solve the problem; Specifically: Initialize the number of particles N in the workflow p , the dimension of the particle is Dim, each particle consists of 3 vectors with Dim dimension, namely the current position of the particle Previous best position and speed And the dimension of the particle Dim is the product of the number of parameters in the objective function (1) and the number of tasks in the workflow; the current position represents a candidate solution to the optimization problem.

5. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 4.

6. A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 4 is implemented.

Images