Deep learning model deployment and task unloading method and system for intelligent computing

Abstract

The invention discloses an intelligent computing-oriented deep learning model deployment and task unloading method and system. The method comprises the following steps of: initializing information; model deployment and unloading decision determination are carried out; carrying out transmission compression ratio design; judging whether a preset iteration convergence limit or a maximum number of iterations is reached; if a preset convergence condition or the maximum number of iterations is reached, outputting model deployment, an unloading decision and a transmission compression ratio scheme; and if the preset convergence condition is not met or the maximum number of iterations is reached, adding 1 to the number of iterations, and returning to carry out model deployment and unloading decision determination. The application focuses on the multi-base-station multi-terminal MEC system oriented to intelligent computing, fully considers new characteristics presented by intelligent computing during deep learning model deployment and task unloading of the MEC system, and proposes a comprehensive joint design scheme of model deployment, unloading decision and transmission compression ratio of the MEC system oriented to intelligent computing.

Description

Technical Field

[0001] The present application relates to the field of mobile communications, and specifically, to a deep learning model deployment and task offloading method for intelligent computing and a system thereof. Background Art

[0002] Mobile Edge Computing (MEC) can effectively reduce service latency and improve service quality by providing communication, computing and storage services to mobile terminals nearby. With the widespread application of artificial intelligence technology, unlike traditional computing models, intelligent computing introduces neural network inference models with complex resource requirements in MEC servers. Therefore, when considering the deployment of intelligent computing involving specific characteristics of artificial intelligence models in MEC systems, such as intelligent compression, complex computing models and high computing power overhead, resource management methods at different levels continue to emerge. However, due to the limited computing and storage capabilities of MEC servers, problems such as service location and resource utilization generated when deploying intelligent computing still exist, which brings new challenges to its implementation in MEC. On the one hand, the demand for service deployment in MEC systems has existed for a long time and is a key factor affecting MEC offloading. Due to its resource-constrained characteristics, MEC servers cannot deploy all the services required by terminals on them; and with the promotion of intelligent computing technology, more and more deep learning models are widely used in computing systems, and due to their existing accuracy requirements, MEC systems face the pressure of service provision brought by the deployment of deep learning models. On the other hand, as the amount of data corresponding to the intelligent computing tasks of deep learning models continues to increase, the communication and computing requirements continue to increase, and the limited resource capacity of MEC servers makes it increasingly difficult to meet the task requirements of terminals. With the continuous development of compression transmission technology using intelligent methods, the terminal can compress and extract the task data that it needs to process and transmit it to the edge server, thereby greatly reducing the amount of data transmitted and calculated while ensuring performance, and also bringing new challenges to the offloading strategy of the MEC system. Therefore, by considering deep learning model deployment and task compression in the MEC system for intelligent computing, the reasoning task requirements and resource allocation problems of the terminal can be solved, so as to achieve the purpose of meeting the terminal needs in the MEC system. At present, some work has studied the resource management schemes for service placement and computing offloading in the MEC system, including cache placement, computing service placement, etc.; in terms of combining intelligent computing, some new intelligent compression transmission technologies have been proposed to solve the problem of limited communication capacity, including compression of multi-modal data, task-oriented compression transmission, etc., which can reduce communication pressure while ensuring transmission effect. However, the service location and resource pressure problems generated by intelligent computing have not been solved.

[0003] Research on the deployment and task offloading of deep learning models for intelligent computing in MEC systems involves the impact of task accuracy and changes in data volume, making deployment and offloading decisions complicated. At present, the deployment and task offloading of deep learning models for intelligent computing in MEC systems face the following difficulties: 1) Existing studies usually consider the placement and deployment of traditional services in MEC systems, model the communication and computing processes of the system in a traditional way, and usually study the performance changes caused by its characteristics in a traditional way when combined with intelligent computing. However, on the one hand, general deployment methods are difficult to adapt to the deployment requirements of deep learning models, and on the other hand, traditional data transmission methods are difficult to adapt to the communication and computing requirements brought about by the increasing data volume of intelligent computing. 2) Existing studies usually use task completion latency as a system performance indicator to study service deployment and task offloading, ignoring the task accuracy requirements brought about by intelligent computing. Therefore, by designing a system utility indicator that balances task accuracy and latency, the design of deep learning model deployment and task offloading strategies for different intelligent scenarios can be better adapted.

[0004] Therefore, how to provide a method to maximize the system utility based on task accuracy and latency has become an urgent problem to be solved in this field. Summary of the invention

[0005] The present application proposes a deep learning model deployment and task offloading method for intelligent computing, comprising the following steps: initializing information; after completing information initialization, determining model deployment and offloading decisions; after completing model deployment and offloading decisions, designing a transmission compression ratio; determining whether a preset iteration convergence limit is reached or a maximum number of iterations is reached; if the preset convergence condition is reached or the maximum number of iterations is reached, outputting the model deployment, offloading decision and transmission compression ratio scheme; if the preset convergence condition is not reached or the maximum number of iterations is reached, adding 1 to the number of iterations, updating the variable value and the system utility value, and returning to determine the model deployment and offloading decision.

[0006] As described above, the deep learning model deployment and task offloading method for intelligent computing, wherein information initialization includes initializing the model deployment plan of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio plan according to the location distribution of the terminal and the task situation information.

[0007] The above-mentioned deep learning model deployment and task offloading method for intelligent computing, wherein the information initialization also includes determining the offloading transmission delay It is expressed as:

[0008]

[0009] in Indicates the amount of compressed data, ruk It represents the uplink rate of terminal u accessing k, and defines the indicator variable of terminal u requesting to offload tasks to MEC server k as where α uk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload, i un Indicates the size of the data, and defines the compression ratio variable as That is, the ratio of the amount of data transmitted by the terminal to the amount of original data. N represents the number of requirements for generating compressed and unloaded intelligent computing tasks to the MEC server. The task set is N = {1,...,n,...,N}, z un represents the indicator variable generated by terminal u and task n, K represents the MEC server equipped on the base station (BS), and its set is K = {1,..,k,...,K}.

[0010] The above-mentioned deep learning model deployment and task offloading method for intelligent computing, wherein the information initialization also includes determining the signal-to-interference-to-noise ratio φ of the terminal uplink offloading communication uk , expressed as:

[0011]

[0012] P uk represents the channel gain of the transmit power, g uk represents the channel gain considering the path loss, χ represents the average co-channel interference power, B is the bandwidth of a single subcarrier, and N0 is the power spectral density.

[0013] The above-mentioned deep learning model deployment and task offloading method for intelligent computing, wherein the signal-to-interference-to-noise ratio φ of the terminal uplink offloading communication is determined uk After that, assuming that the number of subcarriers allocated to the terminal is I, then the uplink rate r uk It is expressed as:

[0014] r uk =I·Blog(1+φ uk ),

[0015] B is the bandwidth of a single subcarrier, φ uk Indicates the signal-to-interference-and-noise ratio during uplink offload communication of the terminal.

[0016] The deep learning model deployment and task offloading system for intelligent computing as described above specifically includes: an initialization unit, a model deployment and offloading decision determination unit, a transmission compression ratio design unit, a judgment unit and an output unit; the initialization unit is used to initialize information; the model deployment and offloading decision determination unit is used to determine the model deployment and offloading decisions; the transmission compression ratio design unit is used to design the transmission compression ratio; the judgment unit is used to judge whether the preset iteration convergence limit is reached or the maximum number of iterations is reached; if the preset convergence condition is reached or the maximum number of iterations is reached, the output unit outputs the model deployment, offloading decision and transmission compression ratio scheme; if the preset convergence condition is not reached or the maximum number of iterations is reached, the number of iterations is increased by 1, and the variable value and the system utility value are updated and then the execution of the model deployment and offloading decision determination unit is returned.

[0017] As described above, in the deep learning model deployment and task offloading system for intelligent computing, the initialization unit performs information initialization, including initializing the model deployment plan of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio plan according to the location distribution of the terminal and the task situation information.

[0018] The deep learning model deployment and task offloading system for intelligent computing as described above, wherein the initialization of information also includes determining the offloading transmission delay It is expressed as:

[0019]

[0020] in Indicates the amount of compressed data, r uk It represents the uplink rate of terminal u accessing k, and defines the indicator variable of terminal u requesting to offload tasks to MEC server k as where α uk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload, i un Indicates the size of the data, and defines the compression ratio variable as That is, the ratio of the amount of data transmitted by the terminal to the amount of original data. N represents the number of requirements for generating compressed and unloaded intelligent computing tasks to the MEC server. The task set is N = {1,...,n,...,N}, z un represents the indicator variable generated by terminal u and task n, K represents the MEC server equipped on the base station (BS), and its set is K = {1,..,k,...,K}.

[0021] In the above-mentioned deep learning model deployment and task offloading system for intelligent computing, the initialization unit performs information initialization and further includes determining the signal-to-interference-to-noise ratio φ of the terminal uplink offloading communication. uk, expressed as:

[0022]

[0023] P uk represents the channel gain of the transmit power, g uk represents the channel gain considering the path loss, χ represents the average co-channel interference power, B is the bandwidth of a single subcarrier, and N0 is the power spectral density.

[0024] In the above-mentioned deep learning model deployment and task offloading system for intelligent computing, in the initialization unit, the signal-to-interference-to-noise ratio φ of the terminal uplink offloading communication is determined uk After that, assuming that the number of subcarriers allocated to the terminal is I, then the uplink rate r uk It is expressed as:

[0025] r uk =I·Blog(1+φ uk ),

[0026] B is the bandwidth of a single subcarrier, φ uk Indicates the signal-to-interference-to-noise ratio during terminal uplink unloading communication

[0027] This application has the following beneficial effects:

[0028] (1) This application focuses on a multi-base station and multi-terminal MEC system for intelligent computing, fully considering the new characteristics of intelligent computing in the deployment of deep learning models and task offloading in the MEC system, and proposes a comprehensive joint design scheme for model deployment, offloading decision and transmission compression ratio of the MEC system for intelligent computing.

[0029] (2) This application can improve the task offloading efficiency of the multi-base station and multi-terminal MEC system for intelligent computing, adjust the accuracy fitting parameters according to different deep learning models, and combine the transmission compression ratio design to achieve an improvement in system utility based on task accuracy and latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0031] Figure 1 It is a flowchart of a deep learning model deployment and task offloading method for intelligent computing provided in accordance with an embodiment of the present application;

[0032] Figure 2It is a schematic diagram of the internal structure of a deep learning model deployment and task offloading system for intelligent computing provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0033] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0034] This application proposes a deep learning model deployment and task offloading method for intelligent computing in MEC systems. By jointly designing model deployment, offloading decisions, and transmission compression ratio in intelligent computing scenarios in MEC systems, the system utility based on task accuracy and latency is maximized.

[0035] Considering the intelligent monitoring scenario of the industrial Internet, there is a sensor terminal set U = {1,...,u,...,U} to collect raw data. Due to the limited computing power and communication capacity, there is a need to compress and offload intelligent computing tasks to the MEC server, whose task set is N = {1,...,n,...,N}; the MEC server is equipped on the base station (BaseStation, BS), whose set is K = {1,..,k,...,K}, and it can process the corresponding deep learning model of the task set through the cloud deployment, which can provide the computing function of the nth task. The BS allocates wireless channels to the terminal for offloading communication through Orthogonal Frequency Division Multiple Access (OFDMA).

[0036] First, consider the terminal u on the terminal side to generate the indicator variable z for task n un ∈{0,1}, where z un =1 means that the task generated by terminal u is n, otherwise it is not. The task has the following properties: 1) The data size is in are the lower and upper bounds of the data size respectively; 2) the task calculation density is δ n , defined as the ratio of the amount of computation required for task n to the amount of data, which is determined by the model of task requirements; 3) Task accuracy requirements Defined as the minimum accuracy that the task must meet; 4) Latency requirement Defined as the maximum delay in task completion.

[0037] Define the indicator variable for terminal u to request to offload tasks to MEC server k as where αuk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload. Considering that the terminal needs to compress the original data before offloading transmission, the compression ratio variable is defined as That is, the ratio of the amount of data transmitted by the terminal to the amount of original data. At the same time, considering the deployment of the deep learning model on the MEC server by the cloud, the indicator variable of the deep learning model corresponding to the deployment task n on the MEC server k is where x nk =1 means that the deep learning model of the corresponding task n is deployed on the MEC server k, otherwise it is not deployed. The variables are divided into block coordinates Z1 = {x, α} and Z2 = {ε} through the idea of ​​BCD for iterative descent.

[0038] Embodiment 1

[0039] like Figure 1 As shown, this embodiment provides a deep learning model deployment and task offloading method for intelligent computing, which specifically includes the following steps:

[0040] Step S110: Initialize information.

[0041] First, the system parameters are designed, including the model deployment plan of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio.

[0042] Specifically, the model deployment scheme of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio scheme are initialized according to the location distribution of the terminal and the task information. For the model deployment scheme, a random deployment method that meets the resource capacity constraints can be adopted initially; for the task offloading scheme of the terminal, the initial access can be initialized in a way with the minimum communication distance; for the transmission compression ratio scheme, it can be initialized in a way of uniformly distributed random assignment according to the constraints.

[0043] In the unloading process after generating a task, let the indicator variable of terminal u requesting to unload the task to MEC server k be where α uk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload, which needs to satisfy That is, a terminal can only offload tasks to one MEC server. The terminal needs to compress the original data before offloading transmission. Let the compression ratio variable be Therefore, offloading the transmission delay for:

[0044]

[0045] in Indicates the amount of compressed data, r ukrepresents the uplink rate of terminal u accessing k. Considering the OFDMA access mode, there is co-channel interference caused by other base station communications in the area. For convenience, the average co-channel interference power χ controlled by the interference management scheme is used to approximate the calculation of the interference power term in the above formula. Therefore, the signal-to-interference-to-noise ratio φ of the terminal uplink unloading communication can be obtained uk for:

[0046]

[0047] Among them, P uk , g uk represents the transmission power and the channel gain considering the path loss, B is the bandwidth of a single subcarrier, and N0 is the power spectrum density. Assuming that the number of subcarriers allocated to the terminal is I, then the uplink rate r uk for:

[0048] r uk =I·Blog(1+φ uk ), (3)

[0049] After the task is offloaded to the MEC server, let the computing capacity allocated by MEC server k to terminal u be f uk When considering task offloading for intelligent computing, it is necessary to consider the parallel computing phenomenon when allocating computing resources to the terminal, that is, the impact of computing power degradation caused by parallel computing, and the actual computing capacity becomes F uk for:

[0050]

[0051] Among them, d k represents the degradation coefficient of the server hardware calculation in MEC server k. Then the calculation delay after terminal u is offloaded to MEC server k for:

[0052]

[0053] Among them, F n (s un )=δ n s un represents the amount of computation required for task n. Considering the maximum parallel number limit of hardware, let M k , must meet

[0054]

[0055] Then, consider the deployment of deep learning models on the MEC server side, where the cloud controls the deployment of the required model on the MEC system. Suppose the indicator variable of the deep learning model corresponding to the deployment task n on the MEC server k is where x nk=1 means that the deep learning model corresponding to task n is deployed on MEC server k, otherwise it is not deployed. Deep learning models have a certain amount of parameters and will occupy a certain amount of resources on the MEC server. They cannot be deployed without restraint. Let e n Indicates the amount of resources occupied by model n, e k represents the amount of resources reserved by MEC server k, then

[0056]

[0057] Indicates that the amount of resources occupied by all deployed models cannot exceed the total resources reserved by the MEC server. At the same time, the relationship between task offloading and model deployment needs to be considered jointly, that is, it needs to meet

[0058]

[0059] The tasks representing the terminals need to be offloaded to the MEC servers with corresponding model deployments.

[0060] Since it involves the deployment of deep learning models and task offloading, the accuracy requirements of the tasks need to be considered. The computational accuracy of the inference task depends on the amount of data in the task and the type of neural network model required for the task. Suppose the accuracy y of task n offloaded by terminal u is n (s un ) is expressed as:

[0061]

[0062] in, and It represents the fitting parameters of different services n obtained by regression fitting of actual data, which is related to the neural network model adopted.

[0063] In order to take into account the accuracy and delay performance of the system, a system utility model based on task accuracy and delay is defined. Taking into account the marginal effect of utility, the total utility U of the system is defined as:

[0064]

[0065] in, and represents the accuracy and delay of terminal u, and ω is the magnitude adjustment factor. Note that there are constraints on the accuracy and delay of task n, namely and It is necessary to maximize the total utility of the system under the constraints of the above conditions: max U .

[0066] Step S120: After completing the information initialization, the model deployment and uninstallation decision are determined.

[0067] Through a given transmission compression ratio scheme, the idea of ​​variable substitution and convex optimization is adopted to transform the original problem into a directly solvable maximum convex problem and an integer programming problem. The branch and bound method is used to solve the integer programming problem, and then the idea of ​​iterative optimization is used to iteratively solve the original problem to obtain the model deployment decision and unloading decision, and obtain Z1.

[0068] Based on the above, step S120 includes the following sub-steps:

[0069] Step S1201: Input initialization information as a given value.

[0070] Step S1202: transform the original problem form by introducing auxiliary variable substitution, and decompose it into an extremum convex problem and an integer programming problem.

[0071] Step S1203: Solve the auxiliary variable maximum convex problem to obtain the solution of the auxiliary variable.

[0072] Step S1204: Solve the integer programming problem by using the branch and bound method to obtain an optimized model deployment and unloading decision plan.

[0073] Step S1205: Determine whether convergence has occurred or the maximum number of iterations has been reached.

[0074] If the iteration termination requirements are not met, update the variable values ​​and system utility values, and the number of iterations + 1.

[0075] If the iteration termination requirement is met, step S1206 is executed.

[0076] Step S1206: Output the obtained model deployment and uninstallation decision results.

[0077] Step S130: After completing the model deployment and unloading decision, the transmission compression ratio is designed.

[0078] Through the given model deployment plan and task offloading decision, the original problem is transformed into a convex problem by using the ideas of variable substitution and Successive Convex Approximation (SCA). Then, the SCA iteration method is used to optimize the original problem and obtain the transmission compression ratio scheme to obtain Z2.

[0079] Based on the above, step S130 includes the following sub-steps:

[0080] Step S1301: taking the initialization information and the obtained results of the model deployment and uninstallation decision as input.

[0081] Step S1302: Perform auxiliary variable substitution and problem conversion of the original problem to convert the original problem into a SCA solvable form.

[0082] Step S1303: Solve through the SCA iteration method to obtain an optimized transmission compression ratio design.

[0083] Step S1304: Output the obtained transmission compression ratio solution result.

[0084] Step S140: Determine whether a preset iteration convergence limit is reached or a maximum number of iterations is reached.

[0085] If the iteration does not reach the preset convergence limit and is less than the maximum number of iterations, the number of iterations is increased by 1, and the variable value and the system utility value are updated, and then the process returns to step S120.

[0086] If the iteration reaches the preset convergence limit or reaches the maximum number of iterations, step S150 is executed.

[0087] Step S150: Output model deployment, offloading decision and transmission compression ratio solution.

[0088] After an iteration is completed, the system utility U and the model deployment plan of the MEC system, the task offloading decision of the terminal and the transmission compression ratio plan are updated. If the system utility value converges to a certain range (the range can be pre-set according to the actual situation), the iteration is terminated and the optimized result is output; or the module records the number of system iterations and terminates the iteration after reaching the maximum allowed number; otherwise, the next iteration process is entered.

[0089] The above method is deployed in the cloud controller of the ground multi-base station MEC system. The sensor terminals in the coverage area will upload their respective task requirements to the MEC server of the BS, including task-related data volume, computing density, accuracy and latency requirements, and model requirements. Each MEC server uploads this information and its own resource information to the cloud controller.

[0090] After the cloud receives the information, the cloud controller will jointly design the model deployment, offloading decision and transmission compression ratio of the MEC system based on the corresponding information. The output results will be broadcast to the MEC server of each base station for the corresponding deep learning model deployment and task offloading transmission.

[0091] Embodiment 2

[0092] like Figure 2 As shown, this embodiment provides a deep learning model deployment and task offloading system for intelligent computing, wherein the system is deployed in a cloud controller of a ground multi-base station MEC system, thereby executing the deep learning model deployment and task offloading method for intelligent computing in Example 1.

[0093] The sensor terminals in the coverage area will upload their respective task requirements to the MEC server of the BS, including task-related data volume, computing density, accuracy and latency requirements, and model requirements. Each MEC server uploads this information and its own resource information to the cloud controller. After the cloud receives the information, the cloud controller controls each unit in the system to perform the joint design of the model deployment, offloading decision and transmission compression ratio of the MEC system according to the corresponding information. The output results are broadcast to the MEC servers of each base station for the corresponding deep learning model deployment and task offloading transmission.

[0094] The deep learning model deployment and task offloading system for intelligent computing specifically includes: an initialization unit 210, a model deployment and offloading decision determination unit 220, a transmission compression ratio design unit 230, a judgment unit 240 and an output unit 260.

[0095] The initialization unit 210 is used to initialize information.

[0096] First, the system parameters are designed, including the model deployment plan of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio.

[0097] Specifically, the model deployment scheme of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio scheme are initialized according to the location distribution of the terminal and the task information. For the model deployment scheme, a random deployment method that meets the resource capacity constraints can be adopted initially; for the task offloading scheme of the terminal, the initial access can be initialized by the method with the minimum communication distance; for the transmission compression ratio scheme, it can be initialized by a uniformly distributed random assignment method according to the constraints.

[0098] In the unloading process after generating a task, let the indicator variable of terminal u requesting to unload the task to MEC server k be where α uk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload, which needs to satisfy That is, a terminal can only offload tasks to one MEC server. The terminal needs to compress the original data before offloading transmission. Let the compression ratio variable be Therefore, offloading the transmission delay for:

[0099]

[0100] in Indicates the amount of compressed data, r ukrepresents the uplink rate of terminal u accessing k. Considering the OFDMA access mode, there is co-channel interference caused by other base station communications in the area. For convenience, the average co-channel interference power χ controlled by the interference management scheme is used to approximate the calculation of the interference power term in the above formula. Therefore, the signal-to-interference-to-noise ratio φ of the terminal uplink unloading communication can be obtained uk for:

[0101]

[0102] Among them, P uk represents the channel gain of the transmit power, g uk represents the channel gain considering the path loss, B is the bandwidth of a single subcarrier, and N0 is the power spectrum density. Assuming that the number of subcarriers allocated to the terminal is I, then the uplink rate r uk for:

[0103] r uk =I·Blog(1+φ uk ), (3)

[0104] After the task is offloaded to the MEC server, let the computing capacity allocated by MEC server k to terminal u be f uk When considering task offloading for intelligent computing, it is necessary to consider the parallel computing phenomenon when allocating computing resources to the terminal, that is, the impact of computing power degradation caused by parallel computing, and the actual computing capacity becomes F uk for:

[0105]

[0106] Among them, d k represents the degradation coefficient of the server hardware calculation in MEC server k. Then the calculation delay after terminal u is offloaded to MEC server k for:

[0107]

[0108] Among them, F n (s un )=δ n s un represents the amount of computation required for task n. Considering the maximum parallel number limit of hardware, let M k , must meet the following requirements:

[0109] Then, consider the deployment of deep learning models on the MEC server side, where the cloud controls the deployment of the required model on the MEC system. Suppose the indicator variable of the deep learning model corresponding to the deployment task n on the MEC server k is where x nk=1 means that the deep learning model corresponding to task n is deployed on MEC server k, otherwise it is not deployed. Deep learning models have a certain amount of parameters and will occupy a certain amount of resources on the MEC server. They cannot be deployed without restraint. Let e n Indicates the amount of resources occupied by model n, e k represents the amount of resources reserved by MEC server k, then:

[0110]

[0111] This means that the amount of resources occupied by all deployed models cannot exceed the total resources reserved by the MEC server. At the same time, the relationship between task offloading and model deployment needs to be considered jointly, that is, the following conditions need to be met:

[0112]

[0113] The tasks representing the terminals need to be offloaded to the MEC servers with corresponding model deployments.

[0114] Since it involves the deployment of deep learning models and task offloading, the accuracy requirements of the tasks need to be considered. The computational accuracy of the inference task depends on the amount of data in the task and the type of neural network model required for the task. Suppose the accuracy y of task n offloaded by terminal u is n (s un ) is expressed as:

[0115]

[0116] in, and It represents the fitting parameters of different services n obtained by regression fitting of actual data, which is related to the neural network model adopted.

[0117] In order to take into account the accuracy and delay performance of the system, a system utility model based on task accuracy and delay is defined. Taking into account the marginal effect of utility, the total utility U of the system is defined as:

[0118]

[0119] in, and represents the accuracy and delay of terminal u, and ω is the magnitude adjustment factor. Note that there are constraints on the accuracy and delay of task n, that is, and It is necessary to maximize the total utility of the system under the constraints based on the above conditions: maxU.

[0120] The model deployment and uninstallation decision determination unit 220 is used to determine the model deployment and uninstallation decision.

[0121] Through a given transmission compression ratio scheme, the idea of ​​variable substitution and convex optimization is adopted to transform the original problem into a directly solvable maximum convex problem and an integer programming problem. The branch and bound method is used to solve the integer programming problem, and then the idea of ​​iterative optimization is used to iteratively solve the original problem to obtain the model deployment decision and unloading decision, and obtain Z1.

[0122] Based on the above, the model deployment and uninstallation decision determination unit 220 performs the following sub-steps:

[0123] Step T1: Input initialization information as a given value.

[0124] Step T2: Transform the original problem form by introducing auxiliary variable substitution, and decompose it into an extremum convex problem and an integer programming problem.

[0125] Step T3: Solve the auxiliary variable maximum convex problem and obtain the solution of the auxiliary variable.

[0126] Step T4: Solve the integer programming problem through the branch and bound method to obtain the optimized model deployment and unloading decision plan.

[0127] Step T5: Determine whether convergence has occurred or the maximum number of iterations has been reached.

[0128] If the iteration termination requirements are not met, update the variable values ​​and system utility values, and the number of iterations + 1.

[0129] If the iteration termination requirement is met, step T6 is executed.

[0130] Step T6: Output the model deployment and uninstallation decision results.

[0131] The transmission compression ratio design unit 230 is used to design the transmission compression ratio.

[0132] Through the given model deployment plan and task offloading decision, the original problem is transformed into a convex problem by using the ideas of variable substitution and Successive Convex Approximation (SCA). Then, the SCA iteration method is used to optimize the original problem and obtain the transmission compression ratio scheme to obtain Z2.

[0133] Based on the above, the transmission compression ratio design unit 230 performs the following sub-steps:

[0134] Step P1: Takes as input the initialization information and the obtained results of model deployment and uninstallation decisions.

[0135] Step P2: Perform auxiliary variable substitution and problem transformation of the original problem to convert the original problem into a SCA solvable form.

[0136] Step P3: Solve through the SCA iteration method to obtain the optimized transmission compression ratio design.

[0137] Step P4: Output the obtained transmission compression ratio solution result.

[0138] The judging unit 240 is used to judge whether a preset iteration convergence limit is reached or a maximum number of iterations is reached.

[0139] If the iteration does not reach the preset convergence limit and is less than the maximum number of iterations, the number of iterations is increased by 1, and the variable values ​​and system utility values ​​are updated and then the execution model deployment and uninstallation decision determination unit 220 is returned.

[0140] If the iteration reaches the preset convergence limit or reaches the maximum number of iterations, the output unit 250 is executed.

[0141] The output unit 250 is used to output the model deployment, offloading decision and transmission compression ratio scheme.

[0142] After an iteration is completed, the system utility U and the model deployment plan of the MEC system, the task offloading decision of the terminal and the transmission compression ratio plan are updated. If the system utility value converges to a certain range (the range can be pre-set according to the actual situation), the iteration is terminated and the optimized result is output; or the module records the number of system iterations and terminates the iteration after reaching the maximum allowed number; otherwise, the next iteration process is entered.

[0143] This application has the following beneficial effects:

[0144] (1) This application focuses on a multi-base station and multi-terminal MEC system for intelligent computing, fully considering the new characteristics of intelligent computing in the deployment of deep learning models and task offloading in the MEC system, and proposes a comprehensive joint design scheme for model deployment, offloading decision and transmission compression ratio of the MEC system for intelligent computing.

[0145] (2) This application can improve the task offloading efficiency of the multi-base station and multi-terminal MEC system for intelligent computing, adjust the accuracy fitting parameters according to different deep learning models, and combine the transmission compression ratio design to achieve an improvement in system utility based on task accuracy and latency.

[0146] Although the present application is described with reference to examples, this is for illustrative purposes only and is not intended to limit the present application, and changes, additions and / or deletions to the embodiments may be made without departing from the scope of the present application.

[0147] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A deep learning model deployment and task offloading method for intelligent computing, characterized in that: The following steps are involved: Initialize information; After completing information initialization, the model deployment and uninstallation decisions are determined; After completing the model deployment and offloading decision, design the transmission compression ratio; Determine whether a preset iteration convergence limit is reached or a maximum number of iterations is reached; If the preset convergence condition is reached or the maximum number of iterations is reached, the model deployment, offloading decision and transmission compression ratio scheme are output; If the preset convergence condition is not reached or the maximum number of iterations is reached, the number of iterations is increased by 1, and the variable values ​​and system utility values ​​are updated before returning to determine the model deployment and uninstallation decisions.

2. The method for deep learning model deployment and task offloading for intelligent computing according to claim 1, characterized in that: Information initialization includes initializing the model deployment plan of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio plan according to the location distribution of the terminal and the task situation information.

3. The deep learning model deployment and task offloading method for intelligent computing according to claim 2, characterized in that: Initializing the information also includes determining the unloading transmission delay It is expressed as: in Indicates the amount of compressed data, r uk It represents the uplink rate of terminal u accessing k, and defines the indicator variable of terminal u requesting to offload tasks to MEC server k as where α uk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload, i un Indicates the size of the data, and defines the compression ratio variable as That is, the ratio of the amount of data transmitted by the terminal to the amount of original data. N represents the number of requirements for generating compressed and unloaded intelligent computing tasks to the MEC server. The task set is N = {1,...,n,...,N}, z un represents the indicator variable generated by terminal u and task n, K represents the MEC server equipped on the base station (BS), and its set is K = {1,..,k,...,K}.

4. The method for deep learning model deployment and task offloading for intelligent computing according to claim 3, characterized in that: Initializing the information also includes determining the signal-to-interference-to-noise ratio φ of the terminal uplink unloading communication uk , expressed as: P uk represents the channel gain of the transmit power, g uk represents the channel gain considering the path loss, χ represents the average co-channel interference power, B is the bandwidth of a single subcarrier, and N0 is the power spectral density.

5. The method for deep learning model deployment and task offloading for intelligent computing according to claim 4, characterized in that: Determine the signal-to-interference-to-noise ratio φ during terminal uplink unloading communication uk After that, assuming that the number of subcarriers allocated to the terminal is I, then the uplink rate r uk It is expressed as: r uk =I Blog(1+φ uk ), B is the bandwidth of a single subcarrier, φ uk Indicates the signal-to-interference-and-noise ratio during uplink offload communication of the terminal.

6. A deep learning model deployment and task offloading system for intelligent computing, characterized in that: Specifically include: Initialization unit, model deployment and unloading decision determination unit, transmission compression ratio design unit, judgment unit and output unit; An initialization unit, used for initializing information; A model deployment and uninstallation decision determination unit, used to determine model deployment and uninstallation decisions; A transmission compression ratio design unit, used for designing a transmission compression ratio; A judgment unit, used to judge whether a preset iteration convergence limit is reached or a maximum number of iterations is reached; If the preset convergence condition is reached or the maximum number of iterations is reached, the output unit outputs the model deployment, offloading decision and transmission compression ratio scheme; If the preset convergence condition is not reached or the maximum number of iterations is reached, the number of iterations is increased by 1, and the variable values ​​and system utility values ​​are updated before returning to the execution model deployment and uninstallation decision determination unit.

7. The deep learning model deployment and task offloading system for intelligent computing according to claim 6, characterized in that: The initialization unit performs information initialization, including initializing the model deployment plan of the MEC system, the task offloading decision of the terminal, and the transmission compression ratio plan according to the location distribution of the terminal and the task situation information.

8. The deep learning model deployment and task offloading system for intelligent computing according to claim 7, characterized in that: Initialization of information also includes determining the unloading transmission delay It is expressed as: in Indicates the amount of compressed data, r uk It represents the uplink rate of terminal u accessing k, and defines the indicator variable of terminal u requesting to offload tasks to MEC server k as where α uk =1 means that the terminal u offloads the task to the MEC server k, otherwise it does not offload, i un Indicates the size of the data, and defines the compression ratio variable as That is, the ratio of the amount of data transmitted by the terminal to the amount of original data. N represents the number of requirements for generating compressed and unloaded intelligent computing tasks to the MEC server. The task set is N = {1,...,n,...,N}, z un represents the indicator variable generated by terminal u and task n, K represents the MEC server equipped on the base station (BS), and its set is K = {1,..,k,...,K}.

9. The deep learning model deployment and task offloading system for intelligent computing according to claim 8, characterized in that: The initialization unit also performs information initialization, including determining the signal-to-interference-to-noise ratio φ of the terminal uplink unloading communication uk , expressed as: P uk represents the channel gain of the transmit power, g uk represents the channel gain considering the path loss, χ represents the average co-channel interference power, B is the bandwidth of a single subcarrier, and N0 is the power spectral density.

10. The deep learning model deployment and task offloading system for intelligent computing according to claim 9, characterized in that: In the initialization unit, determine the signal-to-interference-to-noise ratio φ of the terminal uplink unloading communication uk After that, assuming that the number of subcarriers allocated to the terminal is I, then the uplink rate r uk It is expressed as: r uk =I Blog(1+φ uk ), B is the bandwidth of a single subcarrier, φ uk Indicates the signal-to-interference-and-noise ratio during uplink offload communication of the terminal.

Images