Deep neural network segmentation and task unloading combined scheduling method and device

Abstract

The invention relates to a deep neural network segmentation and task unloading joint scheduling method and device, and the method comprises the steps: constructing a dynamic optimization model based on a Lyapunov function, and constructing a Lyapunov optimization framework containing adjustable parameters according to the dynamic optimization model; determining a deep neural network segmentation strategy meeting a preset optimal condition, and dynamically selecting a mixed mode of all-local execution, all-server execution or partial local and partial server execution of the task by using the computing power, server load and network state of any terminal device; and constructing an optimization objective function containing real-time energy consumption and a queue state, and jointly scheduling a deep neural network segmentation and task unloading strategy under the condition that the optimization objective function meets the calculation capability constraint of any server, so as to determine a final deep neural network segmentation and task unloading joint scheduling result. Therefore, the problems of high task processing energy consumption, large delay and the like in an edge computing system are solved.

Description

Technical Field

[0001] The present application relates to the field of MEC (Mobile Edge Computing) and artificial intelligence task scheduling technology, and in particular to a method for joint scheduling of deep neural network segmentation and task offloading. Background Art

[0002] With the rapid development of 5G technology, the data processing requirements of intelligent terminal devices in multimedia applications such as live video streaming, AR (Augmented Reality), and VR (Virtual Reality) are increasing dramatically. Deep Neural Networks (DNNs), the core technology for implementing these applications, face an increasingly prominent conflict between their high computational complexity and the limited resources of terminal devices.

[0003] While traditional mobile edge computing (MEC) systems can alleviate terminal pressure through task offloading, existing research suffers from the following key flaws: 1. Inadequate task structure adaptation: Most current methods simplify computational tasks into monolithic offloading, failing to effectively leverage the hierarchical nature of deep neural networks. For example, typical models such as ResNet-18 contain a divisible modular structure (e.g., five independently processable blocks), but existing technologies lack a dynamic partitioning mechanism to address this characteristic. 2. Lack of multi-server coordination: Real edge environments often have multiple heterogeneous servers, while existing solutions are mostly designed for single-server scenarios, resulting in suboptimal task allocation when server resources dynamically change. 3. Imbalance between energy efficiency and latency: Traditional methods, such as pure local computing (a Raspberry Pi 2 only supports 1.47 GFLOPS) or full offloading strategies, struggle to optimize energy consumption while maintaining latency constraints. 4. Inadequate dynamic adaptability: Existing algorithms fail to account for the time-varying nature of channel conditions, resulting in significant performance fluctuations in scheduling within short 200ms time slots. Summary of the Invention

[0004] This application provides a method for jointly scheduling deep neural network segmentation and task offloading to solve problems such as high energy consumption and large delay in task processing in edge computing systems.

[0005] The first aspect of the present application provides a method for joint scheduling of deep neural network segmentation and task offloading, comprising the following steps: constructing a dynamic optimization model based on a Lyapunov function to construct a Lyapunov optimization framework containing adjustable parameters according to the dynamic optimization model; based on the Lyapunov optimization framework, determining a deep neural network segmentation strategy that meets preset optimal conditions, and based on the deep neural network segmentation strategy, utilizing the computing power, server load and network status of any terminal device, dynamically selecting a hybrid mode of full local execution, full server execution or partial local and partial server execution of the task; based on the hybrid mode, constructing an optimization objective function containing real-time energy consumption and queue status, and jointly scheduling the deep neural network segmentation and task offloading strategies when the optimization objective function meets the computing power constraint of any server to determine the final joint scheduling result of deep neural network segmentation and task offloading.

[0006] Optionally, in one embodiment of the present application, the dynamic optimization model based on the Lyapunov function is constructed to construct a Lyapunov optimization framework containing adjustable parameters according to the dynamic optimization model, including: based on the virtual queue of the task, converting the delay constraint of the task into a queue stability target; constructing the dynamic optimization model based on the Lyapunov function according to the queue stability target, so as to construct the Lyapunov optimization framework according to the dynamic optimization model.

[0007] Optionally, in one embodiment of the present application, the calculation formula of the virtual queue is:

[0008]

[0009] Among them, Q i (t) represents the virtual queue of terminal i at time t, Q i (t+1) represents the virtual queue of the terminal i at time t+1, N i (t+1) represents the number of tasks in the queue of terminal i at the beginning of time slot t+1, is the queue constraint threshold of the terminal i.

[0010] Optionally, in one embodiment of the present application, based on the deep neural network segmentation strategy, the computing power, server load and network status of any terminal device are utilized to dynamically select a hybrid mode of full local execution, full server execution or partial local and partial server execution of the task, including: partitioning the deep neural network into tasks to generate the deep neural network segmentation strategy; based on the computing power, server load and network status of any terminal device, if the segmentation value in the deep neural network segmentation strategy meets a preset minimum value, not processing the task; if the segmentation value reaches a preset maximum value, dynamically selecting a hybrid mode of full local execution of the task; if the segmentation value is zero, dynamically selecting a hybrid mode of full server execution of the task, and if the segmentation value is within a preset interval value, dynamically selecting a hybrid mode of partial local and partial server execution of the task.

[0011] Optionally, in one embodiment of the present application, the joint scheduling of deep neural network segmentation and task offloading strategy includes: based on the optimization objective function, modeling server resource allocation as a multi-knapsack problem, and calculating the multi-knapsack problem to determine the set of unloading users, corresponding servers and splitting points that meet preset optimal conditions; based on the set of unloading users, corresponding servers and splitting points that meet the preset optimal conditions, jointly scheduling the deep neural network segmentation and task offloading strategy according to the task queue status, server load conditions and channel conditions.

[0012] The second aspect of the present application provides a deep neural network segmentation and task offloading joint scheduling device, including: a construction module for constructing a dynamic optimization model based on the Lyapunov function, so as to construct a Lyapunov optimization framework containing adjustable parameters according to the dynamic optimization model; a dynamic selection module for determining a deep neural network segmentation strategy that meets preset optimal conditions based on the Lyapunov optimization framework, and based on the deep neural network segmentation strategy, dynamically selects a hybrid mode of full local execution, full server execution or partial local and partial server execution of the task by utilizing the computing power, server load and network status of any terminal device; a joint scheduling module for constructing an optimization objective function containing real-time energy consumption and queue status based on the hybrid mode, and jointly scheduling the deep neural network segmentation and task offloading strategies when the optimization objective function meets the computing power constraint of any server to determine the final deep neural network segmentation and task offloading joint scheduling result.

[0013] Optionally, in one embodiment of the present application, the construction module includes: a conversion unit, used to convert the delay constraint of the task into a queue stability target based on the virtual queue of the task; a construction unit, used to construct the dynamic optimization model based on the Lyapunov function according to the queue stability target, so as to construct the Lyapunov optimization framework according to the dynamic optimization model.

[0014] Optionally, in one embodiment of the present application, the calculation formula of the virtual queue is:

[0015]

[0016] Among them, Q i (t) represents the virtual queue of terminal i at time t, Q i (t+1) represents the virtual queue of the terminal i at time t+1, N i (t+1) represents the number of tasks in the queue of terminal i at the beginning of time slot t+1, is the queue constraint threshold of the terminal i.

[0017] Optionally, in one embodiment of the present application, the dynamic selection module includes: a partitioning unit for partitioning the deep neural network into tasks to generate the deep neural network segmentation strategy; a non-processing unit for not processing the task when the segmentation value in the deep neural network segmentation strategy meets a preset minimum value based on the computing power of any terminal device, the server load and the network status; a first dynamic selection unit for dynamically selecting a hybrid mode of fully local execution of the task when the segmentation value reaches a preset maximum value; a second dynamic selection unit for dynamically selecting a hybrid mode of fully server execution of the task when the segmentation value is zero, and dynamically selecting a hybrid mode of partially local and partially server execution of the task when the segmentation value is within a preset interval value.

[0018] Optionally, in one embodiment of the present application, the joint scheduling module includes: a determination unit, used to model server resource allocation as a multi-knapsack problem based on the optimization objective function, and calculate the multi-knapsack problem to determine the set of unloading users, corresponding servers and split points that meet preset optimal conditions; a joint scheduling unit, used to jointly schedule the deep neural network splitting and task offloading strategy according to the task queue status, server load conditions and channel conditions based on the set of unloading users, corresponding servers and split points that meet the preset optimal conditions.

[0019] The third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the joint scheduling method for deep neural network segmentation and task offloading as described in the above embodiment.

[0020] The fourth aspect of the present application provides a computer-readable storage medium, which stores a computer program. When the program is executed by a processor, it implements the above-mentioned deep neural network segmentation and task offloading joint scheduling method.

[0021] The fifth aspect of the present application provides a computer program product, which stores a computer program that, when executed by a processor, implements the above-mentioned deep neural network segmentation and task offloading joint scheduling method.

[0022] The embodiment of the present application can integrate DNN segmentation theory and Lyapunov optimization framework to construct a dynamic decision model that includes multi-dimensional parameters such as task queue status, server load, channel conditions, etc., and solves the problem of limited terminal device resources in multi-server edge computing scenarios by optimizing the segmentation strategy and server selection mechanism of DNN tasks. While ensuring the real-time performance of task processing, the system energy consumption is significantly reduced. This technology can be widely used in real-time multimedia data processing in smart terminals, IoT devices and 5G / 6G networks, such as video analysis, augmented reality (AR) and other application scenarios with low latency and high energy efficiency requirements. Thus, it solves the problems of high energy consumption and large latency in task processing in edge computing systems.

[0023] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0025] Figure 1 This is a flowchart of a method for joint scheduling of deep neural network segmentation and task offloading according to an embodiment of the present application;

[0026] Figure 2 Detailed parameter diagram of a method for joint scheduling of deep neural network segmentation and task offloading according to one embodiment of the present application;

[0027] Figure 3 Graphs showing average queue length and energy consumption under different Lyapunov weights V according to one embodiment of the present application;

[0028] Figure 4This is a comparison chart of energy consumption under different numbers of users according to an embodiment of the present application;

[0029] Figure 5 A system block diagram of a batch task calculation optimization scheduling algorithm according to one embodiment of the present application;

[0030] Figure 6 Schematic diagram of the structure of a deep neural network segmentation and task offloading joint scheduling device provided according to an embodiment of the present application;

[0031] Figure 7 A schematic diagram of the structure of an electronic device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0033] The following describes the method and device for joint scheduling of deep neural network segmentation and task offloading according to the embodiment of the present application with reference to the accompanying drawings. In response to the problems of high energy consumption and large delay in task processing in the edge computing system mentioned in the above background technology, the present application provides a method for joint scheduling of deep neural network segmentation and task offloading. In this method, the DNN segmentation theory and the Lyapunov optimization framework can be integrated to construct a dynamic decision model containing multi-dimensional parameters such as task queue status, server load, and channel conditions. In response to the problem of limited terminal device resources in multi-server edge computing scenarios, by optimizing the segmentation strategy and server selection mechanism of DNN tasks, the system energy consumption is significantly reduced while ensuring the real-time performance of task processing. In this way, the problems of high energy consumption and large delay in task processing in the edge computing system are solved.

[0034] Specifically, Figure 1 A flowchart of a method for joint scheduling of deep neural network segmentation and task offloading provided in an embodiment of the present application.

[0035] like Figure 1 As shown, the deep neural network segmentation and task offloading joint scheduling method includes the following steps:

[0036] In step S101 , a dynamic optimization model based on the Lyapunov function is constructed, so as to construct a Lyapunov optimization framework including adjustable parameters according to the dynamic optimization model.

[0037] It is understood that embodiments of the present application can construct a dynamic optimization model based on the Lyapunov function to construct a Lyapunov optimization framework containing adjustable parameters based on the dynamic optimization model. The framework includes an adjustable parameter V, which is used to control the trade-off between energy consumption optimization and queue stability (latency guarantee). When the parameter V increases, the system tends to reduce energy consumption; when V decreases, it pays more attention to ensuring task processing latency.

[0038] This embodiment of the application utilizes the Lyapunov optimization framework to dynamically adjust task partitioning strategies and server selection. While ensuring real-time task processing, it significantly reduces overall energy consumption for both the terminal device and the system compared to fully local computing or full offload solutions. Lyapunov optimization theory rigorously demonstrates the algorithm's performance boundaries, ensuring that the optimization results for system energy consumption and task latency approach theoretical optimal values.

[0039] Among them, the embodiments of the present application can establish a performance guarantee mechanism: actual verification proves that this optimization framework can achieve an O(1 / V)-O(V) performance boundary: the difference between the average system energy consumption and the theoretical optimal value does not exceed O(1 / V), while ensuring that the average virtual queue length (corresponding to task latency) does not exceed O(V). By adjusting the parameter V, energy consumption performance and latency requirements can be flexibly balanced.

[0040] Optionally, in one embodiment of the present application, a dynamic optimization model based on a Lyapunov function is constructed to construct a Lyapunov optimization framework containing adjustable parameters based on the dynamic optimization model, including: a task-based virtual queue, converting the task's delay constraint into a queue stability target; constructing a dynamic optimization model based on a Lyapunov function according to the queue stability target, to construct a Lyapunov optimization framework based on the dynamic optimization model.

[0041] It can be understood that the queue stability target in the embodiment of the present application may be a queue stability issue.

[0042] During the actual execution process, the embodiment of the present application can convert the delay constraint of the task into a queue stability problem based on the virtual queue of the task, and construct a dynamic optimization model based on the Lyapunov function according to the queue stability, so as to construct a Lyapunov optimization framework according to the dynamic optimization model.

[0043] The embodiment of the present application can ensure that the average processing delay of the task always meets the preset threshold requirements based on the stability control mechanism of the virtual queue, effectively solving the low-latency requirements in the edge computing scenario.

[0044] In step S102, based on the Lyapunov optimization framework, a deep neural network segmentation strategy that meets the preset optimal conditions is determined, and based on the deep neural network segmentation strategy, the computing power, server load and network status of any terminal device are utilized to dynamically select a hybrid mode of full local execution, full server execution or partial local and partial server execution of the task.

[0045] It can be understood that the deep neural network segmentation strategy that meets the preset optimal conditions in the embodiment of the present application can be the optimal deep neural network segmentation solution.

[0046] In the actual execution process, the embodiment of the present application can combine the dynamic task splitting decision mechanism with the multi-server resource scheduling algorithm, determine the optimal deep neural network splitting scheme based on the Lyapunov optimization framework, and based on the optimal deep neural network splitting scheme, dynamically select the full local execution, full server execution, or a hybrid mode of partial local and partial server execution of the task by utilizing the computing power, server load, and network status of any terminal device. At the same time, a centralized optimization method is adopted to solve the optimal decision based on Lyapunov drift plus penalty in each time slot. By constructing an optimization objective function that includes real-time energy consumption and queue status, the task splitting point and server selection are jointly optimized under the premise of satisfying the computing power constraints of each server.

[0047] The embodiments of the present application can support the segmentation and scheduling of multiple deep neural network models, and can be widely used in edge computing scenarios such as video analysis and augmented reality that have strict requirements on energy efficiency and latency.

[0048] It should be noted that the preset optimal conditions can be set by those skilled in the art according to actual conditions and are not specifically limited here.

[0049] Optionally, in one embodiment of the present application, based on the deep neural network segmentation strategy, the computing power, server load and network status of any terminal device are utilized to dynamically select a hybrid mode of full local execution, full server execution or partial local and partial server execution of the task, including: partitioning the deep neural network into tasks to generate a deep neural network segmentation strategy; based on the computing power, server load and network status of any terminal device, if the segmentation value in the deep neural network segmentation strategy meets a preset minimum value, the task is not processed; if the segmentation value reaches a preset maximum value, a hybrid mode of full local execution of the task is dynamically selected; if the segmentation value is zero, a hybrid mode of full server execution of the task is dynamically selected, and if the segmentation value is within a preset interval value, a hybrid mode of partial local and partial server execution of the task is dynamically selected.

[0050] In the actual implementation process, the embodiment of the present application can consider a system consisting of I terminal users and J servers, represented by i∈{1,2,…,I} and j∈{1,2,…,J} respectively. The present application adopts a discrete time model, dividing time into t∈{1,2,…,T} time slots, with the duration of each time slot being τ. At the beginning of each time slot t, each user i starts with probability θ i (θ i ∈[0,1]) receives the DNN computing task, denoted as g i (t) = 1; with probability 1-θ i Do not accept the task, recorded as g i (t) = 0. For user i, each time slot t processes at most one DNN task in the queue (following the first-in-first-out principle). If a task is processed, record b i (t) = 1; otherwise, b i (t) = 0. i (t) represents the number of tasks in the queue of user i at the beginning of time slot t, and its evolution equation is:

[0051] N i (t+1)=N i (t)-b i (t)+g i (t), (1)

[0052] Among them, b i (t) indicates whether the task is processed, g i (t) indicates whether there is a new task arriving in this time slot (if it arrives, it will be processed in the next time slot at the earliest).

[0053] The DNN used by user i can be divided into P i Based on DNN partitioning, the user’s actions are defined in two dimensions. The first dimension involves task partitioning c i (t)∈{-1,0,1,2,…,P i}, which means the following:

[0054] If c i (t) = -1, that is, when the segmentation value in the deep neural network segmentation strategy meets the preset minimum value, the user does not process any task in this time slot.

[0055] If c i (t)=0, that is, when the split value is zero, the task is completely offloaded to the server.

[0056] If c i (t) = P i , that is, when the split value reaches the preset maximum value, the task is completely processed locally.

[0057] Otherwise, c i (t) represents the block {1,2,…,c i (t)} is processed locally, while the intermediate data is compressed and transmitted to the server, which processes the remaining blocks {c i (t)+1,…,P i}.

[0058] Based on C i (t), we can derive the following data volume for the DNN task: The computational load (in CPU cycles) of the local computation is given by:

[0059]

[0060] The size of the intermediate data to be transmitted (in bits) is given by:

[0061]

[0062] The computational load (in GPU cycles) of the server computation is given by:

[0063]

[0064] The second dimension of action involves the uninstallation choice, which is expressed as

[0065] o i (t)=[o i [1](t),o i [2](t),…,o i [j](t),…,o i [J](t)],

[0066] Among them, i [j](t)=1 means that in time slot t, user i selects server j as the offloading server, and o i [j](t)=0 means that in time slot t, user i does not select server j for offloading.

[0067] In a given time slot, user i can only select at most one server j for offloading, and its expression is:

[0068]

[0069] Assume that at most one task can be completed in a time slot. In time slot t, based on the two-dimensional action of user i [c i (t),o i (t)] and the corresponding channel gain, the local computing energy consumption can be determined and edge transmission energy consumption

[0070] Let the time used for local computation in time slot t be τ loc , the time used for transmission is τ trans . They satisfy:

[0071] τ=τ loc +τ trans (6)

[0072] The computation load completed in time slot t is The computational energy consumption of the task is given by the following formula:

[0073]

[0074] Here, γ is a constant related to the chip.

[0075] Users need to perform task offloading in a time-varying channel. Let the bandwidth allocated to user i be W i , the channel gain between user i and server j in time slot t is h i,j (t), which obeys a discrete distribution of finite states Although the system does not know The specific form of h can be observed at the beginning of each time slot t i,j (t). According to Shannon's formula, if o i [j](t) = 1, then the achievable transmission rate between user i and server j in time slot t is:

[0076] r i,j (t) = W i log2(1+p i (t)h i,j (t)). (8)

[0077] Therefore, the relationship is expressed as:

[0078]

[0079] From the above two equations, the transmission energy consumption in time slot t can be derived as follows:

[0080]

[0081] Easy to optimize τ loc and τ trans The total energy consumption of user i in time slot t is given by:

[0082]

[0083] Considering that the edge server also needs to serve other applications, the maximum number of GPU cycles that server j can process in each time slot is is dynamically restricted. This constraint can be expressed as:

[0084]

[0085] In step S103, based on the hybrid mode, an optimization objective function including real-time energy consumption and queue status is constructed, and when the optimization objective function meets the computing capacity constraints of any server, the deep neural network segmentation and task offloading strategies are jointly scheduled to determine the final deep neural network segmentation and task offloading joint scheduling result.

[0086] As a possible implementation method, the embodiment of the present application can be based on a hybrid mode and adopt a centralized optimization method to solve the optimal decision based on Lyapunov drift plus penalty in each time slot, construct an optimization objective function that includes real-time energy consumption and queue status, and jointly schedule deep neural network segmentation and task offloading strategies when the optimization objective function meets the computing power constraints of any server to determine the final deep neural network segmentation and task offloading joint scheduling results. Through innovative optimization theory, the joint optimization of system energy consumption and processing delay is realized, solving the problems of high energy consumption and large delay in task processing in edge computing systems.

[0087] This application aims to optimize the action scheduling of each user i and minimize the average energy consumption of the terminal under the conditions of meeting the task delay and server computing capacity constraints. The problem can be expressed as:

[0088]

[0089] It is worth noting that the first constraint on the average queue length can be directly explained by Little's law. Specifically, at a fixed task arrival rate θ i The average queue length is proportional to the average task residence time. Therefore, the first constraint can be viewed as an equivalent expression of the latency constraint, where the task latency is implicitly controlled by the average queue length.

[0090] In one embodiment of the present application, in order to solve the optimization problem, the following virtual queue is first introduced:

[0091]

[0092] Among them, Q i (t) represents the virtual queue of terminal i at time t, Q i (t+1) represents the virtual queue of terminal i at time t+1, N i (t+1) represents the number of tasks in the queue of terminal i at the beginning of time slot t+1, is the queue constraint threshold of terminal i.

[0093] The Lyapunov function of the virtual queue is defined as:

[0094]

[0095] Let S(t) represent the state of the entire observable system at the beginning of time slot t, and its Lyapunov drift is:

[0096]

[0097] According to the Drift-Plus-Penalty Theorem, the original problem can be transformed into:

[0098]

[0099] Where V is a trade-off parameter that balances the optimization of the objective function with the satisfaction of the constraints. This application scales D(S(t)) and finds its upper bound.

[0100]

[0101] After considering:

[0102]

[0103] and

[0104]

[0105] Afterwards, this application obtains the following upper bound:

[0106]

[0107] Finally, we get:

[0108]

[0109] Here, Δ is a constant that is independent of S(t).

[0110] Therefore, the original optimization problem can be further transformed into an immediate optimization problem for each time slot t:

[0111]

[0112] definition:

[0113]

[0114] In time slot t, if no task is processed in this slot, then If user i chooses to process completely locally, it will not occupy the server computing resources. Can be calculated. If user i chooses to offload to server j at split point c, the corresponding Since it will occupy server computing resources, it can be regarded as a special case of the multi-knapsack problem, which has a mature solution. The item weighs If you do not choose to uninstall after solving, then Select Compute locally or Do not process the task.

[0115] The embodiments of the present application can achieve dynamic load balancing among multiple servers by jointly optimizing task splitting points and server selection, avoiding idle or overloaded resources and maximizing the utilization efficiency of edge computing resources. In response to the problem of limited terminal device resources in multi-server edge computing scenarios, by optimizing the DNN task splitting strategy and server selection mechanism, the system energy consumption can be significantly reduced while ensuring the real-time performance of task processing. This technology can be widely used in real-time multimedia data processing in smart terminals, IoT devices and 5G / 6G networks, such as video analysis, augmented reality (AR) and other application scenarios with low latency and high energy efficiency requirements.

[0116] Optionally, in one embodiment of the present application, the deep neural network segmentation and task offloading strategy are jointly scheduled, including: based on the optimization objective function, modeling the server resource allocation as a multi-knapsack problem, and calculating the multi-knapsack problem to determine the set of unloading users, corresponding servers and splitting points that meet the preset optimal conditions; based on the set of unloading users, corresponding servers and splitting points that meet the preset optimal conditions, the deep neural network segmentation and task offloading strategy are jointly scheduled according to the task queue status, server load conditions and channel conditions.

[0117] Specifically, the embodiment of the present application can model server resource allocation as a multi-knapsack problem based on the optimization objective function, and calculate the multi-knapsack problem to determine the optimal set of unloading users, corresponding servers and split points. Based on the optimal set of unloading users, corresponding servers and split points, the deep neural network splitting and task unloading strategies are jointly scheduled according to the task queue status, server load conditions and channel conditions. The embodiment of the present application is particularly suitable for edge computing scenarios with variable channels, a large number of servers and limited energy. It provides an efficient and stable solution for collaborative reasoning of the Internet of Things and smart terminal devices, and has important application value. Based on the dynamic adjustment mechanism of real-time channel status and server load, it can adapt to changes in network conditions and computing resources to maintain stable and efficient operation of the system in complex environments.

[0118] Specifically, according to the previous derivation, the algorithm flow of this application is as follows: ## Scheduling algorithm flow based on Lyapunov optimization **S1. Initialization**

[0119] At time slot t=0:

[0120] Set N for all users i i (0)=0

[0121] Initialize the virtual queue Q i (t) = 0

[0122] **S2. Time slot t state observation**

[0123] Get all user i's:

[0124] Current queue status N i (t)

[0125] Virtual queue state Q i (t)

[0126] Channel gain h i,j (t)

[0127] Get all edge servers j:

[0128] Maximum number of GPU cycles

[0129] **S3. Cost function calculation**

[0130] Calculate the Lyapunov cost of the three treatments:

[0131] Do not process tasks:

[0132] Fully local processing:

[0133] Offload to server j for processing at split point c:

[0134] **S4. Solving the multi-knapsack problem**

[0135] Model server resource allocation as a multi-knapsack problem:

[0136] Item value:

[0137] Item weight:

[0138] Determine the optimal:

[0139] Uninstalling a User Collection

[0140] Corresponding server j

[0141] Split point c

[0142] Set u for the selected device i i (t) = off

[0143] **S5. Local Decision Processing**

[0144] For users who have not uninstalled:

[0145] pass choose:

[0146] Local computing

[0147] Or do not process the task

[0148] **S6. Status Update**

[0149] Execution scheduling decisions

[0150] renew:

[0151] Queue status N i (t+1)

[0152] Virtual Queue Q i (t+1)

[0153] The time slot counter increments: t←t+1

[0154] Jump to step S2 to continue execution

[0155] Specifically, it can be combined Figures 2 to 4 As shown, the working principle of the deep neural network segmentation and task offloading joint scheduling method in the embodiment of the present application is described in detail with a specific embodiment.

[0156] 1) Implementation example design and setup

[0157] The proposed algorithm is numerically simulated using MATLAB to verify its performance. The total number of time slots is set to T = 100, the duration of each time slot is τ = 200ms, and the number of servers is J = 3. The task arrival rate of each user is set to θ i ~U(0.7,1). The user average queue length threshold is defined as Each user has a fixed bandwidth W i =1MHz.

[0158] h i,j Obey the distribution:

[0159]

[0160] correspond γ=1.38×10 -23 The total computing capacity of server j in time slot t is ~U(4.74×10 8 ,5.93×10 8 ) cycles.

[0161] Use ResNet-18 architecture for image recognition in multi-terminal multi-user clusters. ResNet-18 can be divided into P i =5 blocks, such as Figure 2 shown.

[0162] 2) Implementation results

[0163] This application first verifies the energy consumption and task queue length performance of the proposed algorithm under different V values, with the number of users set to I = 50. Five random seeds are taken, and each seed uniformly selects 20 V values ​​between 0.5 and 30. Figure 3 As shown in Figure 3, under the constraint of average task queue stability, the system exhibits an O(V)-O(1 / V) trade-off relationship between energy consumption and queue length with respect to the Lyapunov parameter V.

[0164] Then, the proposed algorithm is compared with two scenarios: local computing only and complete offloading. The average energy consumption of the three algorithms is compared when the number of users I changes from 20 to 90, where the average task queue length is adjusted by adjusting V. is equal to 4. The result is Figure 4 As shown in the figure. On the one hand, as the number of users increases, the energy consumption of the proposed algorithm increases due to limited server resources, but it is still significantly lower than the case of local computing only. For 90 users, the energy consumption of the proposed algorithm is only 34.5% of that of the case of local computing only. On the other hand, when the number of users is less than 50, full offloading consumes more energy than the proposed algorithm because local computing is more energy-efficient when the channel conditions are poor. When the number of users exceeds 50, full offloading cannot reduce the average queue length to below 4, so this part of the curve is not shown. In summary, compared with the two baseline algorithms, the proposed algorithm can flexibly select the segmentation and offloading strategies according to the channel status, achieving lower energy consumption while handling a larger task computing load.

[0165] 3) Summary

[0166] This application proposes a DNN segmentation and task offloading algorithm under a dynamic multi-user-multi-server MEC framework for multimedia applications. Simulation results verify that the proposed algorithm can effectively optimize energy consumption and task queue length when processing computationally intensive multimedia tasks. Compared with the baseline method, the algorithm achieves significant energy consumption reduction while dynamically adapting to different system conditions. The research results highlight the scalability and energy efficiency of the proposed scheme, making it a promising strategy for enhancing multimedia data processing in MEC environments. Among them, the system block diagram of the batch task calculation optimization scheduling algorithm is shown in the figure below. Figure 5 shown.

[0167] The joint scheduling method for deep neural network segmentation and task offloading proposed in the embodiments of this application integrates DNN segmentation theory with the Lyapunov optimization framework to construct a dynamic decision-making model that includes multi-dimensional parameters such as task queue status, server load, and channel conditions. This addresses the issue of limited terminal device resources in multi-server edge computing scenarios by optimizing the DNN task segmentation strategy and server selection mechanism, significantly reducing system energy consumption while ensuring real-time task processing. This solves the problems of high task processing energy consumption and large latency in edge computing systems.

[0168] Next, a deep neural network segmentation and task offloading joint scheduling device proposed according to an embodiment of the present application is described with reference to the accompanying drawings.

[0169] Figure 6 It is a structural diagram of the deep neural network segmentation and task offloading joint scheduling device of an embodiment of the present application.

[0170] like Figure 6 As shown, the deep neural network segmentation and task offloading joint scheduling device 10 includes: a construction module 100, a dynamic selection module 200 and a joint scheduling module 300.

[0171] Specifically, the construction module 100 is used to construct a dynamic optimization model based on the Lyapunov function, so as to construct a Lyapunov optimization framework containing adjustable parameters according to the dynamic optimization model.

[0172] The dynamic selection module 200 is used to determine the deep neural network segmentation strategy that meets the preset optimal conditions based on the Lyapunov optimization framework, and based on the deep neural network segmentation strategy, utilize the computing power of any terminal device, server load and network status to dynamically select the full local execution, full server execution or a mixed mode of partial local and partial server execution of the task.

[0173] The joint scheduling module 300 is used to construct an optimization objective function including real-time energy consumption and queue status based on a hybrid mode, and to jointly schedule deep neural network segmentation and task offloading strategies when the optimization objective function satisfies the computing capacity constraints of any server to determine the final deep neural network segmentation and task offloading joint scheduling results.

[0174] Optionally, in one embodiment of the present application, the building block 100 includes: a conversion unit and a building unit.

[0175] Among them, the conversion unit is used for the task-based virtual queue to convert the task delay constraint into the queue stability target.

[0176] A construction unit is provided for constructing a dynamic optimization model based on a Lyapunov function according to a queue stability objective, so as to construct a Lyapunov optimization framework according to the dynamic optimization model.

[0177] Optionally, in one embodiment of the present application, the calculation formula of the virtual queue is:

[0178]

[0179] Among them, Q i (t) represents the virtual queue of terminal i at time t, Q i (t+1) represents the virtual queue of terminal i at time t+1, N i (t+1) represents the number of tasks in the queue of terminal i at the beginning of time slot t+1, is the queue constraint threshold of terminal i.

[0180] Optionally, in one embodiment of the present application, the dynamic selection module 200 includes: a partitioning unit, a non-processing unit, a first dynamic selection unit, and a second dynamic selection unit.

[0181] Among them, the partitioning unit is used to partition the deep neural network into tasks to generate a deep neural network segmentation strategy.

[0182] The non-processing unit is used to not process the task if the segmentation value in the deep neural network segmentation strategy meets the preset minimum value based on the computing power of any terminal device, server load and network status.

[0183] The first dynamic selection unit is configured to dynamically select a hybrid mode of fully local execution of the task when the segmentation value reaches a preset maximum value.

[0184] The second dynamic selection unit is used to dynamically select a hybrid mode of full server execution of the task when the split value is zero, and dynamically select a hybrid mode of partial local and partial server execution of the task when the split value is in a preset interval value.

[0185] Optionally, in one embodiment of the present application, the joint scheduling module 300 includes: a determination unit and a joint scheduling unit.

[0186] The determination unit is used to model server resource allocation as a multi-knapsack problem based on the optimization objective function, and calculate the multi-knapsack problem to determine the set of unloaded users, corresponding servers and split points that meet the preset optimal conditions.

[0187] The joint scheduling unit is used to jointly schedule the deep neural network segmentation and task offloading strategy based on the set of offloading users, corresponding servers and split points that meet the preset optimal conditions, according to the task queue status, server load and channel conditions.

[0188] It should be noted that the above explanation of the embodiment of the deep neural network segmentation and task offloading joint scheduling method is also applicable to the deep neural network segmentation and task offloading joint scheduling device of this embodiment, and will not be repeated here.

[0189] The joint scheduling device for deep neural network segmentation and task offloading proposed in the embodiments of this application integrates DNN segmentation theory with the Lyapunov optimization framework to construct a dynamic decision-making model that includes multi-dimensional parameters such as task queue status, server load, and channel conditions. This addresses the issue of limited terminal device resources in multi-server edge computing scenarios by optimizing the DNN task segmentation strategy and server selection mechanism, significantly reducing system energy consumption while ensuring real-time task processing. This solves the problems of high task processing energy consumption and large latency in edge computing systems.

[0190] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include:

[0191] Memory 701 , processor 702 , and computer programs stored in the memory 701 and executable on the processor 702 .

[0192] When the processor 702 executes the program, the deep neural network segmentation and task offloading joint scheduling method provided in the above embodiment is implemented.

[0193] Furthermore, the electronic device further includes:

[0194] The communication interface 703 is used for communication between the memory 701 and the processor 702 .

[0195] The memory 701 is used to store computer programs that can be run on the processor 702 .

[0196] The memory 701 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0197] If the memory 701, processor 702, and communication interface 703 are implemented independently, the communication interface 703, memory 701, and processor 702 can be interconnected via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 7 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0198] Optionally, in a specific implementation, if the memory 701, the processor 702 and the communication interface 703 are integrated on a chip, the memory 701, the processor 702 and the communication interface 703 can communicate with each other through an internal interface.

[0199] The processor 702 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0200] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned deep neural network segmentation and task offloading joint scheduling method.

[0201] An embodiment of the present application also provides a computer program product, which stores a computer program, and when the program is executed by a processor, implements the above-mentioned deep neural network segmentation and task offloading joint scheduling method.

[0202] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0203] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0204] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing a custom logical function or process step, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed in a different order than shown or discussed, including performing functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application pertain.

[0205] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or N wires (electronic devices), a portable computer disk cartridge (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program can be obtained electronically by optically scanning the paper or other medium and then editing, interpreting or processing it in other suitable ways as necessary, and then storing it in a computer memory.

[0206] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0207] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0208] In addition, the functional units in the various embodiments of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into a module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0209] The storage medium mentioned above may be a read-only memory, a magnetic disk, or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application. Persons skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A joint scheduling method for deep neural network segmentation and task offloading, characterized in that: The following steps are involved: Constructing a dynamic optimization model based on a Lyapunov function, and constructing a Lyapunov optimization framework including adjustable parameters according to the dynamic optimization model; Based on the Lyapunov optimization framework, a deep neural network segmentation strategy that meets preset optimal conditions is determined, and based on the deep neural network segmentation strategy, the computing power of any terminal device, server load and network status are utilized to dynamically select a full local execution mode, a full server execution mode or a hybrid mode of partial local and partial server execution of the task; Based on the hybrid mode, an optimization objective function including real-time energy consumption and queue status is constructed, and when the optimization objective function satisfies the computing power constraint of any server, the deep neural network segmentation and task offloading strategies are jointly scheduled to determine the final deep neural network segmentation and task offloading joint scheduling result.

2. The method according to claim 1, characterized in that The method of constructing a dynamic optimization model based on a Lyapunov function, and constructing a Lyapunov optimization framework including adjustable parameters according to the dynamic optimization model, comprises: Based on the virtual queue of the task, converting the delay constraint of the task into a queue stability target; The dynamic optimization model based on the Lyapunov function is constructed according to the queue stability target, so as to construct the Lyapunov optimization framework according to the dynamic optimization model.

3. The method according to claim 2, characterized in that The calculation formula of the virtual queue is: Among them, Q i (t) represents the virtual queue of terminal i at time t, Q i (t+1) represents the virtual queue of the terminal i at time t+1, N i (t+1) represents the number of tasks in the queue of terminal i at the beginning of time slot t+1, is the queue constraint threshold of the terminal i.

4. The method according to claim 1, wherein The method dynamically selects a hybrid mode of full local execution, full server execution, or partial local and partial server execution of tasks based on the deep neural network segmentation strategy and utilizing the computing power of any terminal device, server load, and network status, including: Partitioning the deep neural network into tasks to generate the deep neural network segmentation strategy; Based on the computing power of any terminal device, the server load and the network status, if the segmentation value in the deep neural network segmentation strategy meets a preset minimum value, not processing the task; When the split value reaches a preset maximum value, dynamically selecting a hybrid mode of fully local execution of the task; When the split value is zero, a hybrid mode of full server execution of the task is dynamically selected, and when the split value is within a preset interval value, a hybrid mode of partial local and partial server execution of the task is dynamically selected.

5. The method according to claim 1, wherein The joint scheduling deep neural network segmentation and task offloading strategy includes: Based on the optimization objective function, server resource allocation is modeled as a multi-knapsack problem, and the multi-knapsack problem is calculated to determine a set of offloaded users, corresponding servers, and split points that meet preset optimal conditions; Based on the set of offloading users, corresponding servers and split points that meet the preset optimal conditions, the deep neural network splitting and task offloading strategy is jointly scheduled according to the task queue status, server load and channel conditions.

6. A deep neural network segmentation and task offloading joint scheduling device, characterized in that: include: A construction module, configured to construct a dynamic optimization model based on a Lyapunov function, so as to construct a Lyapunov optimization framework including adjustable parameters according to the dynamic optimization model; A dynamic selection module is configured to determine a deep neural network segmentation strategy that satisfies preset optimal conditions based on the Lyapunov optimization framework, and dynamically select a task execution mode of all local execution, all server execution, or a hybrid mode of partial local and partial server execution based on the deep neural network segmentation strategy and utilizing the computing power of any terminal device, server load, and network status; The joint scheduling module is used to construct an optimization objective function including real-time energy consumption and queue status based on the hybrid mode, and to jointly schedule deep neural network segmentation and task offloading strategies when the optimization objective function meets the computing capacity constraints of any server to determine the final deep neural network segmentation and task offloading joint scheduling result.

7. The device according to claim 6, characterized in that The building blocks include: a conversion unit, configured to convert the delay constraint of the task into a queue stability target based on the virtual queue of the task; A construction unit is used to construct the dynamic optimization model based on the Lyapunov function according to the queue stability target, so as to construct the Lyapunov optimization framework according to the dynamic optimization model.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for joint scheduling of deep neural network segmentation and task offloading as described in any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the deep neural network segmentation and task offloading joint scheduling method as described in any one of claims 1 to 5.

10. A computer program product comprising a computer program, characterized in that The computer program is executed to implement the deep neural network segmentation and task offloading joint scheduling method as described in any one of claims 1 to 5.

Images