A method for task offloading and resource allocation of IoT devices based on photon counting

Abstract

The application belongs to the technical field of optical communication task processing, and particularly relates to a task offloading and resource allocation method for Internet of Things equipment based on photon counting. The application scene comprises a base station, a mobile edge computing server and N Internet of Things equipment; the base station is equipped with N LDs and N PDs, the server is arranged on the base station, data communication is facilitated, the Internet of Things equipment all contain LDs and PDs as data transceiving pairs, and each device is internally provided with a computing power unit; the method comprises the following specific steps: constructing a data model of the mobile edge computing server and the equipment; constructing a task offloading system throughput optimization problem model of the mobile edge computing network; according to the data amount generated by each device at each moment and all queue backlog amounts, a task offloading and resource allocation algorithm based on MDPP-SCA is adopted to make a task offloading and resource allocation strategy. The application can maximize the system throughput and reduce the equipment power consumption, and is suitable for low-power-consumption and high real-time requirement Internet of Things application scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical communication task processing, and particularly relates to a method for task offloading and resource allocation of Internet of Things devices. BACKGROUND

[0002] With the rapid development of Internet of Things (IoT), a large number of intelligent devices work collaboratively through complex network systems to generate massive data to support various application scenarios. However, due to the limitations of the computing power, storage capacity and energy consumption of IoT devices, the traditional mode of transmitting data to the cloud for processing has gradually become difficult to meet the requirements of real-time and low power consumption. Under this background, Mobile Edge Computing (MEC) as a new type of computing architecture has become an important technical solution to solve the above problems. By deploying computing and storage resources near the data source (such as base stations), MEC can offload computing tasks from IoT devices to edge servers, thereby reducing device burden, improving computing efficiency and reducing network delay.

[0003] Although MEC technology has made significant progress in task offloading and resource allocation, two major bottlenecks are generally faced in existing research: high energy consumption of wireless communication and latency optimization of task offloading. In the Internet of Things scenario, traditional wireless communication technologies (such as RF communication) face serious channel interference, high power consumption and limited rate problems, which directly affect the overall system performance. In addition, during the task offloading process, the simultaneous optimization of uplink communication, computing resource allocation and downlink communication needs to be considered, especially in the case of multiple devices running in parallel, the constraints of task return time are often ignored, which greatly reduces the feasibility of the scheme in practical application.

[0004] In recent years, the rise of photon counting wireless optical communication technology provides a new solution to the above problems. Photon counting technology can realize wireless data transmission with high precision and low power consumption by using photodetectors (PD) to capture the number of single photons. Compared with traditional wireless communication, photon counting technology has the advantages of low energy consumption and strong anti-interference, and has been widely studied.

[0005] On this basis, applying photon counting wireless optical communication technology to the MEC task offloading system can not only improve communication efficiency, but also maximize system throughput and reduce overall power consumption through the collaborative optimization of communication power and computing resources. However, there are still few studies on task offloading and resource allocation based on photon counting, especially in the case of considering the multi-queue structure of Internet of Things devices and dynamic task generation, how to efficiently schedule resources to meet multiple constraint conditions is still a difficult problem to be solved. SUMMARY

[0006] The application aims to provide a photon counting-based Internet of Things device task offloading and resource allocation method with high system throughput, low device power consumption and strong real-time performance.

[0007] The photon counting-based Internet of Things device task offloading and resource allocation method provided by the application has an application scenario of an outdoor mobile edge computing network composed of one base station, one mobile edge computing server and N Internet of Things devices, the base station is equipped with N laser diodes (LD) and N photodetectors (PD), the mobile edge computing server is arranged on the base station to facilitate data communication between the base station and the server, and each of the Internet of Things devices contains an LD and a PD as a data transceiving pair, each of the Internet of Things devices is internally provided with a computing unit and has a certain computing capacity. The specific steps of the method are as follows:

[0008] S1, constructing a data model of the mobile edge computing server and the Internet of Things device, including a task data model, a queue model and a communication model;

[0009] S2, constructing a task offloading system throughput optimization problem model of the above mobile edge computing network;

[0010] S3, according to the data amount generated by each Internet of Things device at each moment and the queue backlog amount of all queues, a task offloading and resource allocation algorithm based on MDPP-SCA [1, 2] is used to make a task offloading and resource allocation strategy.

[0011] Further, the data model in step S1 is used to represent the task data amount generated at each moment, and each task data is composed of three parts, namely: is the task data amount (unit: bit), which is a random variable subject to 0-1 distribution, and is represented as v is the ratio of the task return data amount to the task data amount; p is the CPU computing amount (unit: cycle / bit) required for unit data amount; and μ represents the data amount (unit: bit) of the task data P task represents the probability of generating a new task.

[0012] Further, the queue model in step S1 refers to modeling the data backlog of the server and the device at each moment according to the data structure of the queue, which includes four queue structures.

[0013] (1) the local computing queue Q lcThe local computing queue is used to store task data that needs to be calculated locally. When the task selection is performed locally, the local queue fills the task data generated by the frame. The dequeued data is controlled by the computing power of the local device. When local computing is performed, the data in the local queue is obtained. Therefore, Q lc The queue data length at time t can be represented as:

[0014]

[0015] wherein, represents the queue length of the local computing queue at time t, x t represents the offloading strategy, represents the CPU frequency of the local device, and p represents the computing density of the task, represents the computing time of the local device, and D represents the frame length.

[0016] (2) Local offloading queue Q up The local offloading queue is used to store data that needs to be offloaded to the edge server. The input data is generated by the task. The dequeued data is the transmission data at the uplink communication time. Therefore, Q up The queue data length at time t can be represented as:

[0017]

[0018] wherein, is the queue length of the local offloading queue at time t, is the uplink communication rate, represents the uplink communication time.

[0019] (3) Edge computing queue Q ec The edge computing queue is used to store task data offloaded by the Internet of Things device to the edge side. The input data is the amount of data uploaded by the user side, so the input data needs to be less than or equal to the output amount of the offloading queue. The dequeued data is the amount of data calculated by the edge end for each frame. Therefore, Q ec The queue data length at time t can be represented as:

[0020]

[0021] wherein, represents the queue length of the edge computing queue at time t, represents the CPU computing frequency of the edge server at the current time.

[0022] (4) Task return queue Q down The offloading queue is used to store the computing results of the task, so the input data is the result data returned by the edge device after completing the task calculation, and the dequeued data is the communication data of the downlink time slot. Qdown The queue data length at time t can be expressed as:

[0023]

[0024] wherein, represents the queue length of the edge queue at time t, is the downlink communication rate, represents the downlink communication time.

[0025] Further, the communication model S1 includes two parts of communication scheme construction and lower bound derivation of approximate reachable rate of photon counting wireless optical communication.

[0026] In order to improve the communication efficiency of the network and reduce the power consumption of the communication, the communication technology of photon counting wireless optical communication is used, and the communication scheme is as follows: a time division duplex (TDD) communication scheme is used, that is, the communication time is divided into uplink time slots and downlink time slots All devices can only send signals to the base station in the uplink time slots, and receive signals in the downlink time slots. The communication between the device and the base station adopts the communication mode of photon counting wireless optical communication to improve the communication rate between the devices. This is a receiver model in wireless optical communication, which can calculate the exact number of photons reaching the photodetector in a symbol time slot through the interaction of photons and photosensitive elements. In a symbol time slot τ, the number of received signal photons Y is a discrete Poisson random variable, which has a probability mass function:

[0027]

[0028] wherein Y is the number of received photons, y=0, 1,..., ∞ is a non-negative integer; S is the data symbol sent by the device, s∈[0, 1]; λ is the average photon count of the received signal, which satisfies the expression:

[0029]

[0030] wherein η represents the quantum efficiency of photon counting, h represents the channel gain, represents the Planck constant, v represents the frequency of light, n b represents the average count of photons produced by background radiation in each symbol time slot τ, U represents the conversion factor of LD, A represents the current modulation parameter, I d represents the direct current bias, which places the output signal to produce a negative value; represents the expected number of emitted photons, The representative bias intensity is used to represent the intensity of the bias, and the WDMA is used to send information in consideration of the interference between devices in the communication process, and each device is allocated a channel of different wavelength, and the application uses an infrared waveband light source, the reference wavelength is 1550nm, and the channel interval is 100GHz.

[0031] The lower bound of the approximate reachable rate of the photon counting wireless optical communication is derived by first analyzing the reachable rate of the photon counting wireless optical communication, and the approximate reachable rate and its lower bound are obtained by Gaussian approximation and entropy power inequality, and the specific derivation steps are as follows:

[0032] S11, the reachable rate is derived according to the definition of mutual information;

[0033]

[0034] S12, the Gaussian distribution is used to replace the Poisson distribution, and the received photon number is converted into a standard normal distribution form which is easy to express;

[0035] When the mean of the Poisson distribution tends to infinity, the PMF of the Poisson distribution is close to the PDF of the Gaussian distribution with equal mean and variance, that is,

[0036]

[0037] The received photon number Y can be approximated as a Gaussian variable with equal mean and variance The received photon number Y can be approximated as a Gaussian variable with equal mean and variance The received photon number Y can be approximated as a Gaussian variable with equal mean and variance

[0038]

[0039] The reachable rate R can be approximated as:

[0040] R≈R (GA) =h(Y (GA) )-h(Y (GA) |X)=h(X+T)-h(T), (10)

[0041] S13, the lower bound of the approximate reachable rate is obtained by using the entropy power inequality.

[0042] According to the entropy power inequality, for independent random variables X and Y, e 2h(X+Y) ≥e 2h(x) +e 2h(Y) At this time, the lower bound of the approximate reachable rate R (GA) can be obtained as:

[0043]

[0044] Considering that X is uniformly distributed, we have:

[0045] h(X) = -∫ x f X (x)lnf x (x)dx = ln(2|ηn s h|), (12)

[0046] Since the information entropy of a random variable following Gaussian distribution is the largest when the variances are equal, the differential entropy h(T) reaches the upper bound when the random variable T is Gaussian, and we have:

[0047]

[0048] The final lower bound of the rate is obtained as:

[0049]

[0050] Further, the throughput optimization problem model in step S2 refers to maximizing the long-term dequeuing data volume of all Internet of Things device queues Q lc and Q up under the condition of maintaining the stability of all queues, which can be expressed in the following form:

[0051]

[0052] subject to C1:

[0053] C2:

[0054] C3:

[0055] C4:

[0056] C5:

[0057] C6:x nt ∈{0,1},

[0058] C7:

[0059] C8:

[0060] C9:

[0061] C10:

[0062] For simplicity of expression, let be the set of local computing, uplink communication and downlink communication time slots, ​is the set of local computing power, communication power and base station communication power of IoT devices. Constraint C1 indicates that the storage space of IoT devices and servers is limited, and the corresponding queue length is limited, which cannot store unlimited data; C2-C4 represents the communication power p of IoT devices and base stations. up 、p down and local computing power p lc Limited, C5 represents the long-term average power consumption of IoT devices must be lower than P ave , extending the service life of the equipment; C6 represents whether the IoT device needs to offload data at each moment, 1 represents offloading data, and 0 represents that the task is calculated locally; C7-C9 represent the time slot allocation of communication and local computing time. The IoT device cannot offload data and perform task calculations at the same time, but can receive data from the base station during task calculations; the base station can perform uplink and downlink communications and task calculations at the same time; C10 represents the limited computing power of the edge server.

[0063] Furthermore, in order to obtain the optimal offloading strategy and allocation effect, the steps of the MDPP-SCA-based task offloading and resource allocation algorithm described in S3 are as follows:

[0064] S31, use constraint C5 as virtual queue Y n , get the queue backlog of the virtual queue, merge C5 and C1 conditions to get the new queue set

[0065] S32. Using Lyapunov optimization, the throughput optimization problem is decoupled into a deterministic planning problem at each moment. The decoupled optimization problem has the following form:

[0066]

[0067] S33. Decompose the decoupled optimization problem into two resource allocation sub-problems: task offloading and edge computing optimization problem, and communication and local computing resource allocation problem;

[0068] S34, solving the problem of task offloading and edge computing power optimization;

[0069]

[0070] This is a linear programming problem, and after differentiation we can get a closed-form solution:

[0071]

[0072] S35, using continuous convex approximation (SCA) to solve the communication and local computing resource allocation problem;

[0073]

[0074] According to the non-negativity of t The problem is decomposed into three sub-problems: uplink power allocation, downlink power allocation and local computing resource allocation. The local computing resource allocation problem has a solution:

[0075]

[0076] The uplink power allocation problem has the following form:

[0077]

[0078] Introduce auxiliary variable Where:

[0079]

[0080] Using first-order Taylor expansion, the optimization problem can be converted into a convex problem, and the optimal solution is obtained using the CVX toolkit.

[0081] The downlink power allocation problem has the following form:

[0082]

[0083] Where:

[0084]

[0085] Introduce auxiliary variable Using first-order Taylor expansion, the optimization problem can be converted into a convex problem, and the optimal solution is obtained using the CVX toolkit.

[0086] So far, the communication power and local computing resources have been allocated, and the allocation strategy of time slot

[0087] is obtained by classifying discussion:

[0088]

[0089] Otherwise:

[0090]

[0091] Where, there are the following marks:

[0092]

[0093] And all 1n The sum of The sum of 2n All The above steps can obtain all variables t , τt , f t ec , x t optimal solution, so as to obtain the maximum system throughput.

[0094] S36, obtaining the maximum throughput of the system, and updating the queue backlog of the next time.

[0095] The beneficial effects of the above scheme are: according to the channel state of each time, the task amount and the queue backlog state can be generated, the optimal resource allocation strategy can be obtained, and the throughput of the device is maximized. BRIEF DESCRIPTION OF DRAWINGS

[0096] Figure 1 The application scenario system model of the Internet of Things device, the base station and the MEC in the embodiment 1 of the application.

[0097] Figure 2 The flow chart of the task offloading resource allocation framework in the embodiment 1 of the application.

[0098] Figure 3 The algorithm flow chart of the task offloading and resource allocation algorithm based on MDPP-SCA in the embodiment 1 of the application.

[0099] Figure 4 The simulation effect diagram in the embodiment 1 of the application. DETAILED DESCRIPTION

[0100] The specific embodiments of the application are described below to facilitate those skilled in the art to understand the application. However, it should be clear that the application is not limited to the scope of the specific embodiments, and it is obvious to those skilled in the art that various changes are within the spirit and scope of the application defined and determined by the appended claims, and all applications utilizing the concept of the application are within the scope of protection.

[0101] As mentioned in the background, there are few works considering the application of communication scheme using photon counting to task offloading, and in considering task offloading, few works consider the return time of the task, i.e. the downlink communication between the base station and the Internet of Things device, and the application considers this time, which is more consistent with the application conditions in the real scene.

[0102] Embodiment 1:

[0103] First, the parameters used in the example are given, see Table 1.

[0104] Table 1

[0105]

[0106]

[0107] Referring to Figure 1 , is the application scenario system model of the system, and the embodiment is based on an outdoor mobile edge computing network including a base station, a mobile edge computing server and a plurality of Internet of Things devices. The base station is equipped with a laser transmitter (LD) and a photodetector (PD) and is connected to the edge computing server. Each Internet of Things device is also equipped with an LD and a PD for data transmission and reception, while a computing unit is built-in to support certain local computing capability.

[0108] Referring to Figure 2 , the application provides a task offloading scheme based on photon counting, which comprises steps S1, S2 and S3, and each step will be described in detail below in combination with specific embodiments.

[0109] Step S1: constructing a data model of the server and the Internet of Things device, including a task data model, a queue model and a communication model;

[0110] Task data model

[0111] Each task data consists of three parts: is the task data volume (unit: bit), which is a random variable subject to 0-1 distribution, and is expressed as v is the ratio of the task return data volume to the task data volume; and ρ is the CPU computing volume (unit: cycle / bit) required for unit data volume.

[0112] Queue model

[0113] Local computing queue: storing task data executed locally;

[0114] Local offloading queue: storing task data to be offloaded to the edge computing server;

[0115] Edge computing queue: storing task data uploaded to the edge server;

[0116] Task return queue: storing the task result returned after computing.

[0117] Communication model

[0118] The photon counting wireless optical communication mode is adopted, and a time division duplex (TDD) scheme is used to divide uplink and downlink time slots; the wavelength division multiplexing (WDMA) technology is introduced to allocate different wavelength channels; the Poisson distribution is used to model the number of signal photons, and the Gaussian approximation and entropy power inequality are used to derive the lower bound of the approximate achievable rate, and the specific lower bound of the Gaussian approximation achievable rate is

[0119]

[0120] Further, the channel attenuation h = h n (l) h n (a) h n (pe) h n (aoa) , h n (1) = exp(-Z n (Θ) is the atmospheric scattering loss, Z n represents the distance between the base station and the device, and Θ represents the atmospheric attenuation coefficient; h n (a) is the signal attenuation caused by atmospheric turbulence, and h n (pe) and h n (aoa) is the error caused by laser alignment, and the alignment error is ignored in the present example.

[0121] The present embodiment models h n (a) using the Gmma-Gmma distribution, and the PDF of h n (a) is

[0122]

[0123] where α n and β n represent the large-scale and small-scale vortex coefficients, respectively, Γ(·) is the Gamma function, and K n (·) is the second-order modified Bessel function of the n-th order.

[0124] Step S2: Construct a task offloading system throughput optimization problem model of the above mobile edge computing network;

[0125] The objective is to maximize the long-term task throughput of the Internet of Things device while maintaining queue stability. The specific optimization problem is as follows:

[0126]

[0127] subject to C1:

[0128] C2:

[0129] C3:

[0130] C4:

[0131] C5:

[0132] C6: x nt ∈ {0, 1},

[0133] C7:

[0134] C8:

[0135] C9:

[0136] C10:

[0137] Step S3: According to the data amount generated by the Internet of Things device at each time frame τ, the queue backlog of all queues, the task offloading and resource allocation algorithm based on MDPP-SCA is used to make a task offloading and resource allocation strategy, which is specifically divided into the following steps:

[0138] S31, convert the long-term power consumption constraint into a virtual queue backlog;

[0139] S32, combine the Lyapunov optimization method. Use the Lyapunov drift plus penalty framework to decouple the optimization problem into a deterministic planning problem.

[0140]

[0141] subject to C1~C4, C6~C10.

[0142] S33, decompose the decoupled optimization problem into two resource allocation sub-problems, task offloading and edge computing optimization problem, communication and local computing resource allocation problem;

[0143] S34, solve the task offloading and edge computing optimization problem;

[0144]

[0145] subject to C6, C10.

[0146] This is a linear programming problem, and after derivation, a closed-form solution can be obtained,

[0147]

[0148] S35, use successive convex approximation (SCA) to solve the communication and local computing resource allocation problem to obtain the maximum throughput of the Internet of Things device.

[0149]

[0150] subject to C1~C4, C7~C9.

[0151] According to the non-negativity of t The problem is decomposed into three sub-problems: uplink power allocation, downlink power allocation and local computing resource allocation. The local computing resource allocation problem has a solution

[0152]

[0153] The uplink power allocation problem has the following form

[0154]

[0155] s.t. C3.

[0156] Introducing auxiliary variables where Using the first-order Taylor expansion, the optimization problem can be converted into a convex problem, and the optimal solution can be obtained using the CVX toolkit.

[0157] The downlink power allocation problem has the following form

[0158]

[0159] subject to C4

[0160] where, Introducing auxiliary variables Using the first-order Taylor expansion, the optimization problem can be converted into a convex problem, and the optimal solution can be obtained using the CVX toolkit.

[0161] So far, the communication power and local computing resources have been allocated, and the allocation strategy of time slot τ can be discussed in categories. If

[0162] Otherwise,

[0163] where we have the following notation:

[0164]

[0165]

[0166] and the sum of all λ 1n less than 0 is denoted as and the sum of all λ 2n less than 0 is denoted as The optimal solution of all variables p t , τ t , f t ec , x t can be obtained, and the maximum system throughput is obtained.

[0167] By the above method, the embodiment can effectively solve the problem of unbalanced task offloading and resource allocation in the prior art, significantly improve the computing throughput of the Internet of Things system, and reduce the long-term power consumption of the device. Figure 4 is the implementation effect of the method, and the throughput effect of 10 Internet of Things devices after experiencing 8000 frames by using the above task offloading and resource allocation algorithm based on MDPP-SCA. In order to verify the effectiveness of the proposed algorithm, the example provides a benchmark comparison algorithm, SSC (Simple Static Configuration) algorithm: the algorithm does not make decisions based on the environment, and performs the following actions every frame:

[0168]

[0169] From Figure 4 As can be seen from (a), by using the task offloading and resource allocation algorithm proposed in the patent, the throughput of the terminal device can meet the generation amount of tasks. However, by using the SSC algorithm, the throughput of the terminal cannot achieve this effect. In Figure 4 As can be seen from (b), by using the task offloading and resource allocation algorithm based on MDPP-SCA, the queue length of the Internet of Things device is still stable after 8000 frames, and there is no problem of memory overload, but the SSC algorithm will have this problem. Figure 4 (c) reflects the energy consumption of the Internet of Things device, and in this example, the power consumption limit is set to 1 Watt. It can be seen that the task offloading and resource allocation algorithm based on MDPP-SCA can meet this requirement. The simulation can show that the present application can effectively solve the problem of unbalanced task offloading and resource allocation in the prior art.

[0170] The above technical details and algorithm implementation step flow are only used as a demonstration example to better illustrate the method proposed by the present application, and cannot be understood as a limitation of the present application. Other researchers in the field can deform and recombine within the scope of the present application, and these deformations and combinations are still within the protection scope of the present application.

[0171] Reference

[0172] [1] Neely M. Stochastic network optimization with application to communication and queueing systems [M]. Springer Nature, 2022.

[0173] [2] Marks B R, Wright G P. A general inner approximation algorithm for nonconvex mathematical programs[J]. Operations research, 1978, 26(4): 681-683.

Claims

1. A method for task offloading and resource allocation of IoT devices based on photon counting, characterized in that: The application scenario includes an outdoor mobile edge computing network consisting of a base station, a mobile edge computing server, and N IoT devices. The base station is equipped with N laser emitters (LDs) and N photodetectors (PDs). The mobile edge computing server is deployed on the base station to facilitate data communication between the base station and the mobile edge computing server. The IoT devices each include an LD and a PD as a data transceiver pair, and each IoT device has a built-in computing power unit. The specific steps of the method are: S1. Build a data model for the mobile edge computing server and IoT devices, including a task data model, a queue model, and a communication model; where: The task data model is used to represent the amount of task data generated at each moment, where each task data consists of three parts, namely: is the task data, unit bit, which is a random variable obeying the 0-1 distribution, expressed as υ is the ratio of the task return data volume to the task data volume; ρ is the CPU computing power required per unit data volume, unit cycle / bit; P task Indicates the probability of generating a new task; The queue model is to model the data accumulated at each moment by the mobile edge computing server and IoT device according to the queue data structure, including: Local computing queue: stores task data executed locally; Local offload queue: stores task data to be offloaded to the edge computing server; Edge computing queue: stores task data uploaded to the edge server; Task return queue: stores the task results returned after the calculation is completed; The communication model adopts a photon counting wireless optical communication method and uses a time division duplex (TDD) scheme to divide the uplink and downlink time slots; introduces wavelength division multiplexing (WDMA) technology to allocate different wavelength channels; uses Poisson distribution to model the number of signal photons, and derives the lower bound of the approximate achievable rate through Gaussian approximation and entropy power inequality; S2. Construct the throughput optimization problem model of the task offloading system of the mobile edge computing network mentioned above; specifically, keep all queues Under stable conditions, maximize the queue Q of all IoT devices lc and Q up The amount of long-term dequeued data; S3. Based on the amount of data generated by IoT devices at each moment and the queue backlog of all queues, the task offloading and resource allocation algorithm based on MDPP-SCA is used to formulate task offloading and resource allocation strategies.

2. The method according to claim 1, wherein: In the queue model described in step S1: (1) Local computing queue Q lc , used to store task data that needs to be calculated locally. When a task is selected for execution locally, the local queue is filled with the task data generated by the frame; the dequeued data is controlled by the computing power scheduling of the local device. When performing local calculations, the data in the local queue is obtained; so Q lc The queue data length at time t is expressed as: in, represents the queue length of the local computing queue at time t, x t Represents the uninstallation strategy, represents the local CPU frequency, ρ represents the computing density of the task, represents the computation time of the local device, and Δ represents the frame length; (2) Local unloading queue Q up , used to store data that needs to be offloaded to the edge server; the input data source is the data generated by the task; the dequeued data is the transmission data at the time of uplink communication; so Q up The queue data length at time t is expressed as: in, Represents the queue length of the local unloading queue at time t, is the uplink communication rate, Represents the uplink communication time; (3) Edge computing queue Q ec , used to store the task data unloaded by IoT devices to the edge side. The enqueued data is the amount of data uploaded by the user side. The enqueued data is less than or equal to the dequeued data of the unloading queue; the dequeued data is the amount of data calculated by the edge end for each frame. Therefore, Q ec The queue data length at time t is expressed as: in, is the queue length of the edge computing queue at time t, Represents the CPU computing frequency of the edge server at the current moment; (4) Task returns to queue Q down , used to store the calculation results of the task, so the queued data is the result data returned by the edge device after completing the task calculation, and the dequeued data is the communication data of the downlink time slot; Q dwon The queue data length at time t is expressed as: in, represents the queue length of the edge queue at time t, is the downlink communication rate, Indicates the downlink communication time.

3. The method according to claim 2, characterized in that The communication model described in step S1 includes two parts: construction of a communication scheme and derivation of the lower bound of the approximate achievable rate of photon counting wireless optical communication; wherein: (1) The communication scheme adopts a time division duplex (TDD) communication scheme, that is, the communication time is divided into uplink time slots and downlink time slots All devices can only send signals to the base station during uplink time slots and receive signals during downlink time slots. Communication between devices and the base station uses photon counting wireless optical communication to increase the communication rate between devices. This is a receiver model in wireless optical communication. Through the interaction between photons and photosensors, the receiving end's photodetector can calculate the exact number of photons reaching the photodetector within a symbol time slot. The number of received signal photons Y within a symbol time slot τ is a discrete Poisson random variable with a probability mass function: Where Y is the number of received photons, y = 0, 1, ..., ∞ is a non-negative integer; S is the data symbol sent by the device, s∈[0, 1]; λ is the average photon count of the received signal, satisfying the expression: Where η represents the quantum efficiency of photon counting, h represents the channel gain, represents Planck's constant, v represents the frequency of light, and n b represents the average photon count generated by background radiation within each symbol time slot τ, U represents the conversion factor of LD, A represents the current modulation parameter, I d Represents DC bias, placing the output signal produces a negative value; represents the expected number of emitted photons, Represents bias strength. Considering the interference between devices during communication, WDMA is used to send information, allocating channels of different wavelengths to each device. (2) Derivation of the lower bound of the approximate achievable rate of photon counting wireless optical communication: First, the achievable rate of photon counting wireless optical communication is analyzed, and the approximate achievable rate and its lower bound are obtained through Gaussian approximation and entropy power inequality. The specific derivation steps are as follows: S11. Derivation of the achievable rate based on the mutual information definition; S12. Use Gaussian distribution instead of Poisson distribution to convert the number of received photons into a standard normal distribution form that is easy to express; As the mean of the Poisson distribution approaches infinity, the PMF of the Poisson distribution approaches the PDF of a Gaussian distribution with equal mean and variance, i.e.; The number of received photons Y can be approximated as a Gaussian variable with equal mean and variance Using standard normal random variables Convert to standard normal distribution form: The achievable rate R is approximately: R≈R (GA) =h(Y (GA) )-h(Y (GA) |X)=h(X+T)-h(T), (10) S13, using the entropy power inequality to obtain the lower bound of the approximate achievable rate; According to the entropy power inequality, for independent random variables X and Y, with e 2h(X+Y) ≥e 2h(x) +e 2h(Y) ; At this time, the approximate achievable rate R is obtained (GA) The lower bound of : Considering that X is uniformly distributed, then: h(X)=-∫ x f x (x)ln f x (z)dx=ln(2|ηn,h|), (12) Since the information entropy of random variables that obey Gaussian distribution is the largest when the variance is equal, when the random variable T is a Gaussian variable, the differential entropy h(T) reaches the upper bound, which is: The final rate lower bound is:

4. The method according to claim 3, characterized in that The throughput optimization problem model described in step S2 is specifically expressed as follows: C6:x nt ∈{0,1}, remember is the set of local computing, uplink communication and downlink communication time slots, A collection of local computing power, communication power, and base station communication power for IoT devices; Constraint C1 indicates that the storage space of IoT devices and servers is limited, and the corresponding queue length is limited, which cannot store unlimited data; C2-C4 represents the communication power p between IoT devices and base stations up 、p down and local computing power p lc Limited, C5 represents the long-term average power consumption of IoT devices must be lower than P ave , extending the service life of the equipment; C6 represents whether the IoT device needs to offload data at each moment, 1 represents offloading data, and 0 represents that the task is calculated locally; C7-C9 represent the time slot allocation of communication and local computing time. The IoT device cannot offload data and perform task calculations at the same time, but can receive data from the base station during task calculations; the base station can perform uplink and downlink communications and task calculations at the same time; C10 represents the limited computing power of the edge server.

5. The method according to claim 4, characterized in that The specific steps of the MDPP-SCA-based task offloading and resource allocation algorithm described in step S3 are as follows: S31, use constraint C5 as virtual queue Y n , get the queue backlog of the virtual queue, merge C5 and C1 conditions to get the new queue set S32. Using Lyapunov optimization, the throughput optimization problem is decoupled into a deterministic planning problem at each moment. The decoupled optimization problem has the following form: S33. Decompose the decoupled optimization problem into two resource allocation sub-problems: task offloading and edge computing optimization problem, and communication and local computing resource allocation problem; S34, solving the problem of task offloading and edge computing power optimization; This is a linear programming problem, and the closed-form solution is obtained after differentiation: S35, using continuous convex approximation (SCA) to solve the communication and local computing resource allocation problem; According to the time slot τ t The non-negativity of can be used to decompose the problem into three sub-problems: uplink power allocation, downlink power allocation, and local computing resource allocation. The local computing resource allocation problem has a solution: The uplink power allocation problem has the following form: Introduce auxiliary variable α instead in, The optimization problem is transformed into a convex problem using the first-order Taylor expansion, and the optimal solution is obtained using the CVX toolkit; The downlink power allocation problem has the following form: in: Introduce auxiliary variable β instead Using the first-order Taylor expansion, the optimization problem can be transformed into a convex problem, and the optimal solution can be obtained using the CVX toolkit; At this point, the communication power and local computing resources have been allocated. Further classification and discussion will yield the allocation strategy for time slot τ: If: otherwise: Among them, there are the following tags: All λ less than 0 1n The sum is recorded as All λ less than 0 2n The sum is recorded as So we get all the variables p t ,τ t ,f t ec ,x t The optimal solution of , thus maximizing the system throughput; S36. Obtain the maximum throughput of the system and update the queue backlog at the next moment.

Images