An efficient home computing power optimization method based on edge computing

Abstract

The present application belongs to the technical field of edge computing, and particularly relates to an efficient home computing capability optimization method based on edge computing. In order to solve the problem of shortage of home computing power caused by home intelligence, the present application introduces edge computing in home computing. The present application constructs a home computing power scheduling model, which is more suitable for the home scene associated with sub-tasks and reduces the task processing delay of home devices. The present application uses the block coordinate descent method to decompose the non-convex optimization problem into four sub-problems of unloading, computing power optimization, power distribution and bandwidth distribution. The present application proposes a HOME algorithm for overall optimization, which realizes the lightweight deployment in the home edge server to make task processing and resource allocation decisions quickly. Finally, a large number of experiments on real data prove the effectiveness and scalability of the method proposed in the present application for different scenarios.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of edge computing, and particularly relates to an efficient home computing capability optimization method based on edge computing. BACKGROUND

[0002] At present, home devices gradually develop towards automation and intelligence, which will provide people with more convenient and immersive home experience. The types of home electronic devices are also increasingly rich. In addition to common smart phones and personal computers, smart cameras, smart refrigerators, smart sound boxes and other smart home appliances also appear in many families. Even virtual reality (VR), augmented reality (AR) and home robots, which have high demand for computing power, have appeared.

[0003] Most of the existing devices rely on their own computing power to process tasks, and some complex tasks are uploaded to cloud servers for processing. However, with the growth of user demand, computing tasks are becoming increasingly complex. The computing power of the device itself needs to be continuously increased to adapt to changes in demand, but the increase in computing power will inevitably change its size, weight and increase the heat dissipation function, which will seriously reduce the comfort of home devices. Uploading to the cloud is a common solution at present, but relying on uploading to the cloud to process a large number of complex tasks in the future will result in high latency. The proposal of edge computing (EC) solves some limitations of cloud computing (CC). Since it is closer to users at the edge of the network, it can provide users with lower latency. Edge computing has been extensively researched in many fields, effectively solving the problem of tight computing power in various scenarios. For the home scene, edge servers (ES) can be embedded in essential network devices such as routers, or placed independently according to demand. A part of the computing task is offloaded to the edge server to improve the response speed of the device and improve user satisfaction.

[0004] Edge computing has made some progress in many fields. For example, edge computing enables intelligent medical care, edge fog nodes collect user data, and attention mechanisms are used to realize automatic monitoring of medical systems. Meta-reinforcement learning is used to solve the task offloading problem of train automatic bypass scenarios. There are also a large number of studies on vehicle networking scenarios based on edge computing. There are few studies on edge computing for the home. Only some documents are directed at this scenario. The concept of home edge computing (HEC) is first proposed in the research, and a three-layer architecture of home edge computing, edge computing and cloud computing is proposed. Existing work studies the computing offloading of smart TV interaction computing in the home. The document uses a deep neural network architecture to evaluate human fall detection in smart homes. The above studies on edge computing for smart homes are all directed at a certain aspect, and the generality of processing home tasks is poor.

[0005] Common computing tasks can be divided into independent tasks and tasks with dependencies. The inter-task correlation in a home scenario is strong, and even involves multiple device cooperation, for example, in a home intelligent security system, after an abnormality is detected by a home door (burglary) face recognition task, a series of operations such as a task of sending information to family members and a task of taking a photo and uploading need to be performed. SUMMARY

[0006] The application provides an efficient home computing capability optimization method based on edge computing, which realizes efficient processing of home computing tasks based on edge computing technology, and solves the problem of shortage of home computing power caused by limited computing power of home intelligentization and home devices.

[0007] The technical scheme of the application is as follows:

[0008] An efficient home computing capability optimization method based on edge computing comprises the following steps:

[0009] S1, constructing a home computing power scheduling model: the model comprises a communication model and a computing model, the communication model is used to calculate the task offloading delay, and the computing model is used to calculate the delay generated in the computing process of the task;

[0010] S2, defining a problem; the established model is formulated into a problem objective function, that is, solving what problem, and then the problem is defined for selected solution;

[0011] S3, algorithm design and complexity analysis, the BCD method is used to decompose the problem in step S2 into four sub-problems for solution, and then a HOME algorithm is designed to solve the whole problem, specifically including sub-problem solving, overall optimization and complexity analysis, the sub-problem solving is to simplify the problem for solution, the overall optimization is to solve the whole problem after decomposition, and the complexity analysis is to analyze whether the problem is complex and whether it can be solved by a polynomial.

[0012] Optionally, in step S1, the communication model models the channel gain as follows:

[0013]

[0014] Wherein, d0 represents unit path loss, d u,e represents the distance between the device and the edge server, a represents the path loss index, and r is the Rice factor;

[0015] The signal received by the edge server from the home device offloading task can be expressed as:

[0016]

[0017] Wherein, p u,e , xu,e respectively represent the transmit power and the transmit signal of the home device, and η represents the additive white Gaussian noise between the home device and the edge server, and the noise power is ρ 2 ;

[0018] The signal-to-noise ratio between the home device and the edge server is:

[0019]

[0020] The device offloading transmission rate is calculated according to the Shannon formula:

[0021] R u,e =γ u Blog2(1+K u,e );

[0022] wherein γ u is the bandwidth ratio allocated to the device, and B is the bandwidth size.

[0023] Optionally, in step S1, in the calculation model, the user device is calculated in the following manner:

[0024] The size of the s-th subtask of the u-th device is denoted by a u,s , and if the task is calculated on the device, the calculation time can be represented as:

[0025]

[0026] wherein c u,s represents the number of CPU cycles required to process one unit byte of task data, i.e., the task complexity, represents the CPU clock frequency size of the device, and the energy consumed by the calculation process is calculated by the following formula:

[0027]

[0028] wherein, represents the CPU chip architecture related parameter of HEu.

[0029] Optionally, in step S1, in the calculation model, the edge calculation is in the following manner:

[0030] The offloading delay of the task is:

[0031]

[0032] The EC time is similar to the device calculation:

[0033]

[0034] wherein f ESCPU clock frequency of the edge server;

[0035] The time required for the user to offload to the edge server is

[0036] The energy consumption generated in the process of edge computing is:

[0037]

[0038] wherein, the first half of the formula represents the energy consumption of the sub-task transmission of the home device, and the second half represents the energy consumption generated by edge computing, κ ES represents a parameter related to the CPU chip architecture of the edge server.

[0039] Optionally, in step S2, the problem definition is represented as:

[0040]

[0041] wherein, ω u,s represents the offloading strategy, ω u,s = 1 indicates that the sub-task is calculated by the device, ω u,s = 0 indicates that the task is calculated by the edge server, constraint C1 indicates that each sub-task can be calculated by the edge server or itself, constraints C2 and C3 limit the CPU clock frequency range of the home device and the edge server, respectively represent the upper limit of the CPU clock frequency of the home device and the edge server, and the energy consumption constraint when the home device calculates the task is represented by C4, represents the upper limit of the energy consumption, constraints C5 and C6 limit the transmission power range of the home device, constraint C6 sets the minimum SNR, K min guarantees the communication conditions of the HE, and the maximum SNR, K max makes it possible to prevent the additional energy consumption caused by the excessive transmission power under the premise of meeting the communication quality of the home device, and constraint C7 represents the size of the bandwidth ratio that can be allocated to each home device.

[0042] Optionally, in step S3, the sub-problem decomposition includes offloading decision optimization, computing power optimization, power allocation optimization, and bandwidth allocation optimization, wherein, in the offloading decision optimization, the offloading strategy optimization problem is represented as:

[0043]

[0044] {ω u,s} represents the offloading decision, represents the computing time of the u-th device and the s-th sub-task on the device, denotes the computing time of the u-th device, s-th subtask on the edge server;

[0045] {ω u,s} is expressed as:

[0046]

[0047] According to the above gradient, iterative solution is carried out;

[0048] In the calculation power optimization, the CPU clock frequency of the edge server and the home device is further solved, and the variables {p u,s ,γ u} of the fixed power allocation and bandwidth allocation, then the sub-problem can be expressed as follows:

[0049]

[0050] The Lagrange function of problem P3 can be formulated as:

[0051]

[0052] Wherein, the variables λ u ,μ and v u,s are non-negative Lagrange multipliers corresponding to constraints C1, constraints C2 and constraints C3 respectively, according to the following KKT conditions, the optimal calculation power CPU clock frequency and f ES* of the home device and the edge server, and the optimal Lagrange multipliers μ * and For KKT conditions are expressed as:

[0053]

[0054] μ(f ES -f ES,max )=0;

[0055]

[0056] According to the above equation group, we can get:

[0057]

[0058] When μ≠0, f ES =f ES,max , μ is calculated as:

[0059]

[0060] In power allocation optimization, based on the calculated unloading decision and simplified calculations, the subproblem can be represented as follows:

[0061]

[0062] Analyze the characteristics of the above function and differentiate the objective function:

[0063]

[0064] The optimal solution for power allocation is:

[0065]

[0066] In bandwidth allocation optimization, the bandwidth allocation subproblem can be represented as follows:

[0067]

[0068] The objective function for bandwidth allocation optimization is about γ u A monotonically decreasing function whose objective function is minimized in the variable γ u At the maximum value:

[0069]

[0070] Optionally, in step S3, during the overall optimization, the vector variables and numerical variables {ω(0)} and {f} are first initialized. hd (0)f ES (0)}, {p(0)} and {γ(0)}, and then optimize the four subproblems, fixing {f hd ,f ES ,p,γ}, after relaxation transformation according to algorithm1, iteratively optimize {ω(0)}, and restore the obtained results to integers 0 and 1. Then, based on the result of the unloading decision optimization, for {f hd (n), f ES (n)} uses the KKT method to find a closed-form solution, and finally follows the... and The calculation method is used to gradually obtain {p(n)} and {γ(n)}.

[0071] Optionally, in step S3, during the complexity analysis, the algorithm complexity is... in, Let E represent the total number of subtasks across all home devices, and E represent the number of edges connecting all tasks to their subtasks. For the offloading decision optimization subproblem, the computational complexity mainly depends on the number of devices, subtasks, and the number of iterations. If the algorithm converges after I iterations, then the complexity of this subproblem is...

[0072] After the above technical solutions are adopted, the application has the following beneficial effects:

[0073] 1. The application constructs an edge computing family computing power problem optimization model of subtask dependence to realize family computing power concentration and provide computing power guarantee for complex computing scenarios in the family.

[0074] 2. The application designs a block coordinate descent algorithm (BCD method) to divide it into an offloading subproblem, a computing power optimization subproblem, a power distribution subproblem and a bandwidth distribution subproblem. Relaxation change, Karush-Kuhn-Tucker (KKT) and other methods are used to solve different subproblem characteristics, realizing optimal resource scheduling and task processing of the scene.

[0075] 3. The application designs a family computing power optimization algorithm (HOME algorithm), which realizes the overall optimization of the problem. And through the analysis of the algorithm complexity of the method, the lightweight of the deployment is illustrated.

[0076] 4. The application carries out all aspects of experiments based on real data sets and deploys tests on real edge servers. Finally, through the comparison of different experiments, the effectiveness and scalability of the HOME algorithm are proved. BRIEF DESCRIPTION OF DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0078] Figure 1 is the overall process of the HCPS model;

[0079] Figure 2 is the convergence analysis of the HOME algorithm under different learning rates;

[0080] Figure 3 is the delay comparison of the HOME algorithm with other algorithms under different device numbers. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0082] The embodiment of the application discloses a high-efficiency home computing capability optimization method based on edge computing.

[0083] Embodiment

[0084] According to Figures 1-3 The high-efficiency home computing capability optimization method based on edge computing.

[0085] 1, system model and problem definition

[0086] The application studies an edge server assisted home area, wherein U home devices (HE) are randomly distributed in the home area covered by the edge server. Each home device contains multiple dependent sub-tasks, such as file downloading, online gaming and other tasks. The size of the s-th sub-task of the u-th device is represented by a u,s The scene flow chart of the home computing power scheduling model (HCPS) constructed by the application is shown in Figure 1 The application integrates the edge server and the wireless router, so that the future home only needs to purchase one product to realize the integration of computing and networking. The communication distance of the home is short, and the delay of the wired connection between the home device and the network device is obviously lower than that of the wireless connection. Since the application takes the task processing delay as the optimization target, it is assumed that all the home devices transmit signals through the home wifi. The dependent sub-tasks are first sequentially processed, and at the same time, the edge server collects information such as channel state and device computing power of the tasks to be processed. Finally, the edge server makes an optimal low-delay offloading decision based on the HOME algorithm according to the collected information.

[0087] 1.1 Communication model

[0088] Due to the characteristics of multiple obstacles in the home scene, the application considers adopting OFDMA technology for wireless signal transmission, and divides the transmission bandwidth into multiple orthogonal sub-carrier sets allocated to each device to reduce signal interference between devices. Therefore, the transmission of signals between the home device and the edge server mainly considers the large-scale fading caused by path loss and the small-scale fading caused by signal multipath propagation. The application models the channel gain as follows:

[0089]

[0090] Wherein, d0 represents the unit path loss, d u,e represents the distance between the device and the edge server, and a represents the path loss exponent. The Reyleigh factor r is considered. It is assumed that the channel is quasi-static during the offloading process of the device. The signal received by the edge server after the home device offloads the task can be represented as:

[0091]

[0092] Wherein, p u,e, x u,e denote the transmit power and the transmit signal of HE, respectively. η denotes the additive white Gaussian noise (AWGN) between the home devices and the edge server, with noise power 2 . The signal-to-noise ratio (SNR) between the home devices and the edge server is:

[0093]

[0094] Since the present application adopts OFDMA technology, the interference between devices is negligible. Thus, the device offloading transmission rate can be calculated according to the Shannon formula:

[0095] R u,e = γ u Blog2(1+K u,e ); (4)

[0096] wherein γ u is the bandwidth ratio allocated to the device, and B is the bandwidth size.

[0097] 1.2 Calculation model

[0098] 1.2.1 User device calculation

[0099] Each device has a different number of sub-tasks, denoted by sub u . The present application adopts the classic Kahn algorithm to sort the association topology graph of the sub-tasks, and stores the output queue in the list S u in order. The size of the s-th sub-task of the u-th device is denoted by a u,s . If the task is calculated on the device, the calculation time can be represented as:

[0100]

[0101] wherein c u,s denotes the number of CPU cycles required to process one unit byte of task data, i.e., the task complexity. denotes the CPU clock frequency size of the device. The energy consumed by the calculation process is calculated by the following formula:

[0102]

[0103] wherein denotes the CPU chip architecture-related parameters of HEu.

[0104] 1.2.2 Edge calculation

[0105] When the home devices offload tasks to the edge server for auxiliary computing, the present application mainly considers the time delay and energy consumption of two processes: offloading process and edge computing process. Since the data size of the computing result is much smaller than the offloading data size, this part is negligible. The offloading time delay of the task is:

[0106]

[0107] The time required for the s-th task on the u-th device to perform edge computing is calculated as follows, and the calculation method is similar to the task computing on the device:

[0108]

[0109] wherein f ES represents the CPU clock frequency size of the edge server.

[0110] In summary, the time required for the user to offload to the edge server is The energy consumption generated in the process of edge computing is:

[0111]

[0112] wherein the first half of the formula represents the transmission energy consumption of the sub-task of the home device, and the latter half represents the energy consumption generated by edge computing. K ES represents the CPU chip architecture related parameter of the edge server.

[0113] 1.3 Problem definition

[0114] According to the communication model and the computing model of HCPS, the processing time of the system mainly includes local computing time, offloading time delay and edge computing time. In view of the above time delay, the present application fully considers the CPU clock frequency resources of the devices and the edge server in the home environment, the device power and the bandwidth, schedules the resources in the HCPS model, and obtains the optimal task processing offloading strategy with low time delay. The optimization problem can be represented as:

[0115]

[0116] wherein, ω u,s represents the offloading strategy, ω u,s = 1 indicates that the sub-task is calculated by the device, ω u,s = 0 indicates that the task is calculated by the edge server. The constraint C1 indicates that each sub-task can be calculated by the edge server or itself, and the constraints C2 and C3 limit the CPU clock frequency range of the home device and the edge server, respectively represent the upper limit of CPU clock frequency of the home devices and edge servers. Since the edge servers are wired and have sufficient power supply, the present application only restricts the energy consumption of the home devices when performing tasks, represented by C4, represent the upper limit of energy consumption to prevent excessive consumption of device power. The constraints C5 and C6 limit the transmission power range of the home devices, and constraint C6 sets the minimum SNR (signal-to-noise ratio), K min guarantees the communication conditions of the HEs, while the maximum SNR, K max prevents excessive energy consumption caused by high transmission power under the premise of meeting the communication quality of the home devices. Constraint C7 represents the size of the bandwidth proportion that each home device can be allocated, and the present application allocates the required bandwidth size according to the actual number of sub-tasks of the devices.

[0117] 2. Algorithm design and complexity analysis

[0118] As can be seen from the variables and constraints of the optimization problem (10), it is a mixed integer programming problem and a non-convex optimization problem. To solve this NP-hard problem, the present application uses the BCD method to first decompose it into four sub-problems for solving. Then, the HOME algorithm is designed to solve the whole problem. The decomposition process of the optimization problem of the present application first fixes the variables {ω u,s} of the offloading sub-problem, and then obtains the solution of the optimal computing power sub-problem according to the obtained offloading decision {ω u,s} and the fixed {p u,s , γ u}. Similar to the above process, the bandwidth allocation {p u,s} is solved, and then the device bandwidth γ u is allocated according to the obtained optimal variables.

[0119] 2.1 Sub-problem decomposition

[0120] 2.1.1 Offloading decision optimization

[0121] The offloading decision is determined by the variable {ω u,s}, which indicates whether the s-th sub-task of the u-th device needs to be offloaded for computing. According to the fixed variables the present application represents the offloading strategy optimization problem as:

[0122]

[0123] where {ω u,s} represents the offloading decision, This represents the computation time of the u-th device and the s-th subtask on the device. Let represent the computation time of the u-th device and the s-th subtask on the edge server; according to constraint C1, this problem is still a non-convex problem. This invention uses a relaxation transformation method to change the variable {ω}. u,s Relax the continuous variables in the interval [0,1] and change the constraint C1 to 0≤ω u,s ≤1, It can be seen that the transformed objective function (11) is a linear programming problem, which is solved using the gradient descent method in this invention. The result is {ω}. u,s The gradient of}

[0124]

[0125] The solution is obtained iteratively based on the gradient described above.

[0126] 2.1.2 Computing power optimization

[0127] Based on the results of the offloading decision optimization, the CPU clock frequencies of the edge server and home devices are further calculated. The variables for fixed power allocation and bandwidth allocation are {p}. u,s ,γ u Then the subproblem can be represented as follows:

[0128]

[0129] Due to variables f ES In the optimization objective, (10b) and (10c) are used as the denominator, so the constraints are made more stringent and cannot be 0, which are used as constraints C1 and C2 for problem P3.

[0130] Theorem: Problem P3 in formula (13) is a convex optimization problem.

[0131] Proof: In formula (13), the objective function and constraints of problem P3 are expressed as follows:

[0132]

[0133] for and f ES objective function and The second derivative is:

[0134]

[0135] Where, ω u,s When = 1, the task is computed on HE, where a u,s cu,s and are positive, then and ω u,s = 0, the task is offloaded to the edge server computing, and similarly Therefore, the objective function and the constraint function are convex functions. And the constraints and m1(f ES ) are linear functions. So all the constraints are convex sets. Therefore, the problem P3 in equation (13) is a convex problem.

[0136] According to the optimization objective and constraints of problem P3, Karush-Kuhn-Tucker (KKT) is used to solve it. The Lagrangian function of problem P3 can be formulated as:

[0137]

[0138] where variables λ u , μ and v u,s are non-negative Lagrange multipliers corresponding to constraints C1, C2 and C3 respectively. According to the following KKT conditions, the optimal CPU clock frequencies of the family device and the edge server and f ES* , and the optimal Lagrange multipliers μ * and For KKT conditions are expressed as:

[0139]

[0140] μ(f ES -f ES,max ) = 0; (18)

[0141]

[0142] According to the above equations, we can get:

[0143]

[0144] From the complementary slackness condition (18) of KKT, we know that when μ ≠ 0, f ES = f ES,max , so the μ in (23) is calculated as:

[0145]

[0146] 2.1.3 Power allocation optimization

[0147] For this sub-problem, the transmit power of the home devices can be optimized according to the results of the offloading decision optimization and the computing power optimization. The variable of bandwidth allocation is {γ u} since the power allocation is only related to the offloading to the edge server computing, i.e. ω u,s = 0 involves the optimization of this variable. Therefore, according to the calculated offloading decision, the simplified calculation is as follows:

[0148]

[0149] In order to analyze the function characteristics, the present application derives the target function:

[0150]

[0151] It can be seen from the above formula that the optimization target of power allocation is a decreasing function, and the minimum value is obtained at the maximum value of p u,e . According to the constraint, the optimal solution of power allocation is:

[0152]

[0153] 2.1.4 Bandwidth allocation optimization

[0154] The bandwidth allocation problem is similar to the power allocation problem, and only needs to consider the offloading to the edge server, and the case of ω u,s = 0 can be considered. The bandwidth allocation sub-problem can be expressed as follows:

[0155]

[0156] Since γ u is the denominator, formula (10g) is rewritten as formula (28a). After decomposing the sub-problem, it can be seen that the target function of bandwidth allocation optimization is a monotonically decreasing function about γ u , and the minimum value of the target function is obtained at the maximum value of the variable γ u :

[0157]

[0158] In addition, in order to better achieve the optimization of the overall optimization problem, the present application proposes HOME to iteratively optimize the four sub-problems of the above block coordinate descent method.

[0159] 2.2 Overall optimization

[0160] The algorithm process is shown in Figure 1 , which first initializes the vector variables and numerical variables {ω(0)},{f hd (0)f ES(0)},{p(0)} and {γ(0)}. The four sub-problems are optimized in turn, fixing {f hd ,f ES ,p,γ} are relaxed according to algorithm 1 in table 1, and the obtained results are restored to integers 0 and 1. Then, according to the optimization results of the offloading decision, the closed solution of {f hd (n),f ES (n)} is obtained using the KKT method. Finally, {p(n)} and {γ(n)} are obtained step by step according to formula (27) and formula (29). The convergence is judged after each iteration, and the iteration is continuously iterated until the iteration times reach the acceptable set maximum or the convergence is reached. The algorithm is shown in table 1:

[0161] Table 1 algorithm

[0162]

[0163] 2.3 Complexity analysis

[0164] Before the problem is optimized, the invention uses Kahn algorithm to sort the topological graph of sub-tasks, and the algorithm complexity is wherein, is the total number of sub-tasks of all home devices, and E represents the number of edges associated with all tasks containing sub-tasks. Then, for the offloading decision optimization sub-problem, the calculation complexity is mainly related to the number of devices, the number of sub-tasks, and the number of iterations. If the algorithm converges after I iterations, the complexity of the sub-problem is In the sub-problem algorithm optimization, power allocation and bandwidth allocation, the invention obtains the closed solution of the corresponding sub-problem in formula (21), formula (23), formula (27) and formula (29) through mathematical derivation. Therefore, the solution of the three sub-problems is only related to the problem size. Because the calculation on the device and the calculation on the edge server are not coupled, the calculation complexity of the algorithm optimization is Similarly, the calculation complexity of power allocation and bandwidth allocation is only related to the device itself and is not related to the sub-task. Then the calculation complexity is Finally, the overall optimization of the algorithm problem is performed. If the algorithm converges after N iterations, the calculation complexity of the overall optimization of HOME is It can be seen that the complexity of the algorithm is a polynomial level. Since the number of devices and the number of sub-tasks in the home scene is limited, the algorithm can give a real-time scheme with low delay. And the algorithm model is relatively light and suitable for deployment in the home edge server for algorithm scheduling.

[0165] 3. Experiment

[0166] Experimental setup: The present method mainly relies on CPU for processing and calculation. Considering the heterogeneity of the domestic edge server, the present application respectively constructs a model on a device equipped with a 32GB memory 13th Gen Intel(R) Core(TM) i9-13900K, a 64GB memory 12th Gen Intel(R) Core(TM) i7-12700KF, a 32GB memory AMD Ryzen7 7840HS and a 32GB memory Apple M1 CPU, and obtains the same convergence effect. Finally, the present application performs all performance experiments on an edge server equipped with a 32GB memory 13th Gen Intel(R) Core(TM) i9-13900K.

[0167] Firstly, the present application discusses the convergence of HOME as Figure 2 To obtain better convergence of the method, the present application sets different dynamic learning rates to analyze the performance. Different learning rate settings are as follows:

[0168] StepLR: fixed step learning rate decay, decay factor β = 0.9.

[0169] CosLR: cosine learning rate decay.

[0170] MultiLR0.9: variable length learning rate decay, decay factor β = 0.9.

[0171] MultiLR0.5: variable length learning rate decay, decay factor β = 0.5.

[0172] 0.01: learning rate remains 0.01.

[0173] LinearLR: linear learning rate decay, decay factor β = 0.9.

[0174] Figure 2 The convergence of HOME is verified, and the overall processing time of the system decreases with the increase of the training step number. The present application compares the influence of different learning rate methods on the performance of the algorithm. From the experimental results, it can be seen that the learning rate algorithm using StepLR and MultiLR0.9 has good performance. In subsequent experiments, the learning rate is based on 0.01 using the MultiLR0.9 dynamic decay method. More user devices may occur in actual home scenarios, and the present application performs related experiments.

[0175] The present application considers different situations of the system, (1) all tasks are calculated locally (TALC algorithm); (2) all tasks are calculated on the edge server (TAEC algorithm); (3) part of the tasks is offloaded to the edge server according to the HOME algorithm and part of the tasks is calculated on the edge. The experiments of the three situations are as follows Figure 3As shown, it can be seen that the results of different device quantities are slightly different, and when the device quantity is only 5, the total processing delay of the three is relatively small. After the device quantity gradually increases, the task processing time of TAEC is obviously smaller than that of TALC, but is larger than that of HOME, which reflects the effectiveness of the method in the application in reducing the processing delay.

[0176] The application introduces edge computing into home computing to solve the problem of tight home computing power caused by home intelligence. The application constructs a home computing power scheduling model, which is more suitable for the home scene associated with sub-tasks and reduces the task processing delay of home devices. The application uses the BCD method to decompose the non-convex optimization problem into four sub-problems of offloading, computing power optimization, power allocation and bandwidth allocation. The HOME algorithm for overall optimization is proposed to realize the lightweight deployment of the home edge server to make task processing and resource allocation decisions quickly. Finally, a large number of experiments on real data prove the effectiveness and scalability of the method proposed in the application for different scenarios.

[0177] The application introduces edge computing into the home computing scene to make task processing decision and resource allocation, and realizes the idea of "thin terminal" + "edge computing". The application considers the complex home task processing scene with sub-task dependence, and focuses on the task processing time minimization problem. A home computing power optimization method (HOME) based on edge computing is designed. The application mainly solves the complex processing scene with task dependence, and the independent task scene without dependence is only a case in the modeling of the application, that is, the number of sub-tasks is 1.

[0178] The above is only a preferred embodiment of the application and is not used to limit the application. Any modification, equivalent replacement, improvement, etc. within the technical solution of the application shall be included in the protection scope of the application.

Claims

1. An efficient home computing capability optimization method based on edge computing, characterized in that, The method comprises the following steps: S1, constructing a family computing power scheduling model: the model includes a communication model and a calculation model, the communication model is used to calculate the task offloading delay, and the calculation model is used to calculate the delay generated in the calculation process of the task; S2, problem definition, which is expressed as: ； wherein, , denotes the offloading strategy, denotes that the sub-tasks are calculated by the home devices, denotes that the sub-tasks are calculated by the edge servers, denotes the transmit power of the home devices, constraint C1 denotes that each sub-task can be calculated by the edge servers or the home devices, and constraints C2 and C3 limit the CPU clock frequency ranges of the home devices and the edge servers, denotes the CPU clock frequency size of the home devices, denotes the CPU clock frequency size of the edge servers, , denote the upper limits of the CPU clock frequencies of the home devices and the edge servers respectively, and constraint C4 limits the energy consumption when the home devices calculate the tasks, denotes the upper limit of the energy consumption, constraints C5 and C6 limit the transmit power ranges of the home devices, and constraint C6 sets the minimum signal-to-noise ratio to ensure the communication conditions of the home devices, and constraint C7 sets the maximum signal-to-noise ratio to prevent the excessive energy consumption caused by the excessive transmit power under the premise of meeting the communication quality of the home devices, constraint C7 denotes the bandwidth proportion size that can be allocated to each home device, denotes the number of sub-tasks possessed by each home device, denotes the bandwidth proportion allocated to the home devices, denotes the calculation time of the s-th sub-task of the u-th home device on the home device, denotes the calculation time of the s-th sub-task of the u-th home device on the edge servers, denotes the signal-to-noise ratio from the u-th home device to the edge server e, denotes the energy consumed in the calculation process of the home devices; S3, algorithm design and complexity analysis, the BCD method is used to decompose the problem in step S2 into four sub-problems for solving, and then the HOME algorithm is designed to solve the whole problem, which specifically includes sub-problem solving, overall optimization and complexity analysis; In step S3, the sub-problem decomposition includes offloading decision optimization, computing power optimization, power allocation optimization and bandwidth allocation optimization, wherein in the bandwidth allocation optimization, the bandwidth allocation sub-problem is expressed as follows: ； The objective function for bandwidth allocation optimization is given by a monotonically decreasing function whose minimum value is taken at the maximum value of the variable x. ； in, Expressed as noise power, Indicates the first The first of the home appliances Sub-task size, The task complexity is the number of CPU cycles required to process one unit of data in the task. This represents the channel gain from the u-th home device to the edge server e. This refers to the bandwidth size. 2.The efficient home computing capability optimization method based on edge computing of claim 1, wherein, In step S1, the communication model models the channel gain as follows: ； wherein, represents the unit path loss, represents the distance between the home device and the ES, ES represents an edge server, represents the path loss exponent, is the Ricean factor; The signal that the edge server receives the offloading task of the home device can be expressed as: ； wherein, represents a transmission signal of the HE, the HE represents a home device, represents an additive white Gaussian noise as a channel between the home device and the edge server, and a noise power of the AWGN is ; The signal-to-noise ratio between the home device and the edge server is: ； According to the Shannon formula, the offloading transmission rate of the home device is calculated: 。 3.The efficient home computing capability optimization method based on edge computing of claim 2, wherein, In step S1, in the calculation model, the calculation method of the home device is as follows: The first The first The first The computation time of this task on the home device can be expressed as: ； wherein, The number of CPU cycles required to process one unit byte of the task data, i.e. the task complexity, is denoted by T. The energy consumed by the home device computation process is calculated by ； wherein, represents a CPU chip architecture related parameter of the u-th HE.

4. The efficient home computing capability optimization method based on edge computing according to claim 3, characterized in that, In step S1, in the calculation model, the edge calculation method is as follows: The offloading delay of the task is: ； The EC time is similar to the calculation of the home device, wherein EC represents edge calculation: ； The time required by the user to uninstall to the edge server is ; The energy consumption generated in the process of edge calculation is: ； Wherein, the first half of the formula represents the energy consumption of the subtask transmission of the home device, and the latter half represents the energy consumption generated by edge computing, represents the parameter related to the CPU chip architecture of the edge server.

5. The efficient home computing capability optimization method based on edge computing according to claim 4, characterized in that, In step S3, in the offloading decision optimization, the offloading strategy optimization problem is expressed as: ； wherein, represents the number of sub-tasks each home device owns; The gradient is expressed as: ； Iterative solution is performed according to the above gradient; In the computing power optimization, the CPU clock frequency of the edge server and the home device is further solved, and the variables of the fixed power allocation and the bandwidth allocation Then the computing power optimization problem can be expressed as follows: ； wherein, represents a CPU chip architecture related parameter of the u-th HE; Problem The Lagrangian function of the problem can be formulated as: ； Among them, variables and These are the non-negative Lagrange multipliers corresponding to constraints C1, C2, and C3, respectively. Based on the following KKT conditions, solve for the optimal CPU clock frequency for home devices and edge servers. and and the optimal Lagrange multiplier and ,for The KKT conditions are expressed as: ； ； ； ； ； According to the above equation set, the following can be obtained: ； ； ； ； At time, , The calculation is: ； In the power allocation optimization, the calculated offloading decision is used for simplified calculation, and then the power allocation optimization problem can be expressed as follows: ； The above function characteristics are analyzed, and the objective function is derived: ； The optimal solution of power allocation is: 。 6. The efficient home computing capability optimization method based on edge computing according to claim 5, characterized in that, In step S3, in the overall optimization, first, initialize each vector variable and numerical variable , , and , then optimize four sub-problems, fix , according to the algorithm algorithm1 relaxation changes after iteration optimization , and restore the results obtained to integer 0 and 1, then according to the optimization results of the unloading decision , , use KKT method to obtain the closed solution, and finally gradually obtain and according to the calculation method of and .

7. The efficient home computing capability optimization method based on edge computing according to claim 6, characterized in that, In step S3, in the complexity analysis, the algorithm complexity is wherein, is the total number of sub-tasks of all home devices, represents the number of edges formed by the correlation between each sub-task, and then for the optimization sub-problem of the offloading decision, the complexity is mainly related to the number of home devices, the number of sub-tasks and the number of iterations. If the algorithm converges after iterations, the complexity of the sub-problem is .

Images