Intelligent computing fusion network heterogeneous resource elastic adaptation and deterministic scheduling system and method

Abstract

The application provides a kind of intelligence calculation fusion network heterogeneous resource elasticity adaptation and deterministic scheduling system and method, belongs to industrial internet of things technical field, comprising: the system of establishing distributed management and centralized control under the scene of intelligence calculation fusion network;According to the distributed management and centralized control system, the two-stage random optimization problem of different time scales is established;For large-scale resource adaptation problem, a resource elasticity adaptation scheme based on heuristic algorithm is adopted;For small-scale scheduling sub-problems, a deterministic scheduling scheme based on deep reinforcement learning is adopted;Using the resource elasticity adaptation scheme and deterministic scheduling scheme, the optimal resource adaptation and scheduling result is obtained.The application can meet the resource flexible configuration demand under high dynamicity and uncertainty environment, ensure low delay and high reliability in data scheduling process.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, specifically to a system and method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks that meets the needs of data-intensive scenarios. Background Technology

[0002] With the rapid development of artificial intelligence technology, data-intensive applications such as intelligent control of industrial robots, AI model training, and holographic communication are constantly emerging. These applications rely on large-scale model architectures to process massive amounts of data, giving rise to an exponential increase in demand for computing power. In particular, for the distributed training of large-scale AI models, it is necessary to schedule exabyte (EB) level computing power across multiple domains and ensure the stability of parameter synchronization latency, which poses a significant challenge to traditional resource management and network scheduling mechanisms.

[0003] Data-intensive applications place dual demands on networks: on the one hand, they require dynamic and coordinated adaptation of computing, storage, and network resources; on the other hand, they emphasize enhancing the network's deterministic quality of service (QoS) assurance capabilities. Specifically, these applications not only have stringent requirements for computing power but also set high standards for network reliability and low latency. However, traditional network architectures have limitations in unified management of intelligent computing resources and cross-domain collaborative scheduling, which affects the QoS assurance of data-intensive application scenarios and may hinder the large-scale application of intelligent manufacturing.

[0004] When dealing with complex and ever-changing business needs in heterogeneous network environments, further breakthroughs are still needed in the collaborative optimization scheduling and dynamic deterministic guarantee mechanism of multi-dimensional resources (including computing power, storage, and network). Summary of the Invention

[0005] The purpose of this invention is to provide a system and method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks, so as to solve at least one of the technical problems existing in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a system for elastic adaptation and deterministic scheduling of heterogeneous resources in an intelligent computing converged network, comprising: a domain server (DS) that manages resources in a computing domain, a centralized controller (CC) that acquires resource information of all devices, and time-sensitive network devices, the time-sensitive network devices including terminals, TSN switches, and a centralized network configuration (CNC); wherein, terminals in each computing domain establish a connection with the DS through the TSN, and then the terminals send all information to the domain server; the control plane collects status information to make network configuration decisions; a distributed management and centralized control mode is adopted; in centralized control, the centralized controller is responsible for distributing control lists and forwarding tables of all TSN switches, and determining the computing domain, path, and transmission time of all task flows; for distributed management, terminals report to the domain server (DS) to form a computing domain; the centralized controller (CC) determines the computing domain, path, and transmission time of the task flow based on the information collected from the DS; the data plane schedules the task flow according to the scheduling table from the CC.

[0008] As a further limitation of the first aspect of the present invention, a two-stage stochastic optimization problem with different time scales is established, including: a resource adaptation subproblem for a large time scale and a scheduling subproblem for a small time scale; for the resource adaptation subproblem at the large scale, a resource adaptation algorithm based on heuristic algorithms is used, including a cross-domain intelligent computing resource elastic adaptation mechanism framework, a description and modeling of task flow requirements and resource matching problems, and a cross-domain collaborative scheduling algorithm based on re-decision greed; the cross-domain intelligent computing resource elastic adaptation mechanism framework operates in windows after decision-making, denoted as i∈U={1,2,…,U}; within a window, the resource adaptation decision remains unchanged; in each window, DS obtains the available resources and resource requirements of task flows in adjacent computing domains in the current window, and the central controller obtains the resource status of all computing domains in the current window and the resource requirements of task flows, and makes resource adaptation decisions based on task flow resource requirements and resource constraints within the domain.

[0009] As a further limitation of the first aspect of the present invention, the overall resource adaptation algorithm of the heuristic algorithm is as follows: Sorting and initialization: Sorting tasks according to their resource requirements and deadlines; prioritizing tasks with large requirements and short tolerance times; Traversing all tasks: Starting from high priority, traversing all computational domains to find the optimal domain that can meet the resource requirements of the task; Achieving load balancing: Among the domains that meet the conditions, selecting the domain with lower resource utilization for allocation to achieve load balancing while improving resource utilization; The cross-domain collaborative scheduling algorithm based on heavy decision greed includes: multi-objective weighted sorting, dynamic candidate set reduction, fragment-aware adaptation, and a two-level reassignment mechanism.

[0010] As a further limitation of the first aspect of the present invention, multi-objective weighting includes:

[0011] Design a composite weighting function to prioritize tasks:

[0012]

[0013] Where α and β are resource type weight coefficients, and γ is the time delay weight. and This represents the resource requirements and latency requirements of the task flow u. For computationally intensive tasks, higher weights can be assigned, resources can be allocated preferentially, and different resource dimensions can be compared. This sorting mechanism allows the algorithm to prioritize critical tasks.

[0014] As a further limitation of the first aspect of the present invention, the dynamic candidate set reduction includes: for each task, the algorithm dynamically constructs a candidate domain set that satisfies the following conditions:

[0015]

[0016] in, It is a set of computational domains that meet the resource requirements of task flow u. This represents the three types of resources currently available in computation domain m. This represents the resource reservation threshold for computation domain m, used to ensure system stability. This represents the three resource quantities required by task flow u; when the three available resource quantities of computing domain m are all greater than the resource quantities required by task flow, it is an available computing domain, and a dynamic candidate set of available computing domains is constructed.

[0017] As a further limitation of the first aspect of the present invention, the fragment perception adaptation includes:

[0018] Define a multidimensional resource utilization vector and a comprehensive compactness function to achieve resource fragmentation-aware allocation:

[0019] First, calculate the normalized utilization rate of various resources in each domain:

[0020]

[0021] in, This represents the utilization rate of the three resources in computational domain m. This represents the remaining resources within the computational domain m;

[0022] Then, the overall resource compactness of computational domain m is defined, which is the average resource utilization rate of the three resources:

[0023]

[0024] Finally, the most suitable domain is selected based on the principle of minimum compactness:

[0025]

[0026] As a further limitation of the first aspect of the present invention, the secondary reallocation mechanism includes: for tasks that failed in the primary allocation, i.e., tasks that failed in the first allocation, a secondary reallocation mechanism is introduced to minimize the variance of resource utilization of all computing nodes through a secondary allocation strategy. Constraints Ensure that each set T from the first-level allocation failures failed The task flow is allocated to a single computation domain, thereby minimizing the variance of global resource distribution while ensuring feasible task scheduling.

[0027]

[0028] Among them, T failed Represents the set of failed tasks in the first-level allocation, var(Φ m ) represents the variance of resource compactness across all domains, DR i (m,u)∈{0,1} is a decision variable, representing whether to assign task flow u to computation domain m.

[0029] Secondly, the present invention provides a method for elastic adaptation and deterministic scheduling of heterogeneous resources in an intelligent computing converged network based on the system described in the first aspect. The method is characterized by comprising: terminals within each computing domain establishing connections with the DS via a TSN, and subsequently sending all information to the domain server; the control plane collecting status information to make network configuration decisions; and adopting a distributed management and centralized control mode. In centralized control, the centralized controller is responsible for distributing control lists and forwarding tables of all TSN switches, determining the computing domain, path, and transmission time of all task flows. For distributed management, terminals report to the domain server DS, forming a computing domain. The centralized controller CC determines the computing domain, path, and transmission time of the task flows based on information collected from the DS. The data plane schedules the task flows according to the scheduling table from the CC.

[0030] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks as described in the first aspect.

[0031] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks as described in the first aspect.

[0032] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks as described in the first aspect.

[0033] The beneficial effects of this invention are as follows: By constructing a system that combines distributed management and centralized control, and introducing a two-stage stochastic optimization model based on different time scales, it effectively solves key problems in existing technologies such as difficulties in multi-dimensional resource collaborative scheduling and insufficient deterministic guarantee capabilities. This improves resource utilization efficiency and task scheduling success rate, enhances the deterministic guarantee capabilities of the network, and supports unified scheduling and collaborative management in multi-domain heterogeneous network environments.

[0034] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing fusion networks according to an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of a scenario for the intelligent computing converged network heterogeneous resource elastic adaptation and deterministic scheduling system described in an embodiment of the present invention. Detailed Implementation

[0038] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0039] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0040] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0041] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0042] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0043] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0044] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0045] Example 1

[0046] This embodiment provides a method for elastic adaptation and deterministic scheduling of heterogeneous resources in an intelligent computing converged network, including: establishing a system for distributed management and centralized control in an intelligent computing converged network scenario; and establishing a two-stage stochastic optimization problem with different time scales based on the distributed management and centralized control system.

[0047] Establish a system for distributed management and centralized control in a network scenario where intelligent computing is integrated, including:

[0048] Domain servers (DS) manage resources within a computing domain name;

[0049] A central controller (CC) that can monitor all device resource information;

[0050] Time-Sensitive Networks (TSN) devices mainly include terminals, TSN switches, and Centralized Network Configuration (CNC).

[0051] Domain servers are responsible for managing the computing resources of their respective computing domains. Each computing domain is allocated computing devices or computing centers with varying performance. All domain servers are connected to a central controller;

[0052] The central server is responsible for sensing all computing tasks generated within the computing domain, the availability of resources, network topology, link capacity, and executing resource adaptation and scheduling strategies.

[0053] Any two computing domains can be connected via a TSN switch.

[0054] Time-sensitive information between the two TSN networks is scheduled by the 5G system and the connecting device relay.

[0055] Based on the distributed management and centralized control system, establish two-stage stochastic optimization problems at different time scales, including: for the resource adaptation subproblem at a large time scale, design a heuristic resource adaptation algorithm; for the scheduling subproblem at a small time scale, design a deterministic scheduling algorithm based on deep reinforcement learning (DRL).

[0056] Based on the distributed management and centralized control system, a two-stage stochastic optimization problem with different time scales is established. For the resource adaptation subproblem with a large time scale, a heuristic resource adaptation algorithm is designed, including: description and modeling of task flow requirements and resource matching problems; and design of a cross-domain collaborative scheduling algorithm based on heavy decision greed.

[0057] For the resource adaptation subproblem on a large time scale, a heuristic resource adaptation algorithm is designed. The task flow requirements and resource matching problem are described and modeled, including: optimization objective: maximize overall resource utilization; resource constraints: within each computing domain, the reserved resources plus the consumed resources cannot exceed the resource capacity; single computing domain processing constraint: each computing task can only be assigned to one computing domain for processing within a window.

[0058] The resource adaptation subproblem on a large time scale is addressed by a heuristic-based resource adaptation algorithm and a cross-domain collaborative scheduling algorithm based on heavy decision greed, which includes: multi-objective weighted sorting; dynamic candidate set reduction; fragment-aware adaptation; and a two-level reallocation mechanism.

[0059] The aforementioned system establishes a two-stage stochastic optimization problem with different time scales based on the distributed management and centralized control system. For the small time scale scheduling subproblem, a deterministic scheduling algorithm based on DRL is designed, including: modeling the deterministic scheduling optimization problem; transforming the deterministic scheduling problem into a Markov decision process; and a deterministic scheduling algorithm based on the Dual Double Deep Q Network (D3QN).

[0060] For the small-scale scheduling subproblem, a deterministic scheduling algorithm based on DRL is designed. The deterministic scheduling optimization problem is modeled as follows: Optimization objective: Maximize the task flow success scheduling rate; Path integrity constraint: Ensure that data packets pass through all nodes sequentially along the paths in the path set; Queue capacity constraint: The total number of data packets in each queue within a time slot cannot exceed the maximum capacity of the queue; Path selection uniqueness constraint: Each data packet can only choose one path and one starting time slot; Link bandwidth constraint: The total bandwidth of the same link in each time slot cannot exceed the total bandwidth of the link; Same computation domain constraint: Each task flow must be scheduled to be processed in the computation domain determined by the resource adaptation algorithm.

[0061] For the small-scale scheduling subproblem, a deterministic scheduling algorithm based on DRL is designed, transforming the deterministic scheduling problem into a Markov decision process, including: State: The central controller collects data packet information and current queue capacity information from the DS, TSN bridges, and 5GS system; Action: The central controller determines the path and time slot of data packets for people and goods based on the state; Reward: After the agent performs an action, it receives a reward to evaluate the value of taking that action in that state, in order to maximize the success rate of scheduling.

[0062] For the small time-scale scheduling subproblem, a deterministic scheduling algorithm based on DRL and a deterministic scheduling algorithm based on D3QN are designed, including: algorithm initialization; empirical sampling; network training.

[0063] Example 2

[0064] This embodiment provides a system and method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks, which can be used in emerging scenarios such as the Industrial Internet of Things. Figure 1As shown, the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks includes: establishing a system corresponding to distributed management and centralized control in the intelligent computing converged network scenario; establishing a two-stage optimization problem at different time scales based on the distributed management and centralized control system; employing a heuristic-based resource adaptation algorithm for the large-scale resource adaptation sub-problem; and adopting a deterministic scheduling sub-problem based on DRL for the small-scale scheduling sub-problem. This embodiment provides a method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks, comprehensively considering resource adaptation and reliable scheduling, appropriate computation domains, paths, and transmission time slots, thereby ensuring deterministic scheduling.

[0065] like Figure 2 As shown, the intelligent computing converged network heterogeneous resource elastic adaptation and deterministic scheduling system includes a domain server DS that manages resources in the computing domain name, a centralized controller CC that can grasp all device resource information, and time-sensitive network devices (terminals, TSN switches, and centralized network configuration CNC).

[0066] The above mechanism operates as follows: 1) Terminals within each computing domain establish connections with the DS via the TSN, and then the terminals send all information to the domain server. 2) The control plane collects status information to make network configuration decisions. A distributed management and centralized control mode is adopted. In centralized control, the CC is responsible for distributing the control lists and forwarding tables of all TSN switches, determining the computing domain, path, and transmission time of all task flows. For distributed management, terminals report to the DS, forming a computing domain. Taking the domain server DS1 as an example, there is a subset of terminals U = {u1, u2, ..., Ui}, where Ui is a subordinate device of DS1. DS1 maintains a list to manage devices. Terminal Ui only registers or deregisters on DS1, avoiding direct processing pressure on the control plane and the generation of a large amount of control data on the backbone network. 3) The CC determines the computing domain, path, and transmission time of the task flow based on the information collected from the DS. 4) The data plane schedules the task flow according to the scheduling table from the CC.

[0067] The network framework for distributed management and centralized control establishes a two-stage stochastic optimization problem with different time scales, including a resource adaptation subproblem for large time scales and a scheduling subproblem for small time scales.

[0068] For the large-scale resource adaptation subproblem, a resource adaptation algorithm based on heuristics is proposed, including a cross-domain intelligent computing resource elastic adaptation mechanism framework, description and modeling of task flow requirements and resource matching problems, and a cross-domain collaborative scheduling algorithm based on heavy decision greed.

[0069] The resource adaptation framework operates in windows after the decision is made, denoted as i∈U={1,2,…,U}. Within a window, the resource adaptation decision remains unchanged. In each window, DS obtains the available resources and resource requirements of the task flow in the adjacent computing domains of the current window. The specific functions of the central controller CC are as follows: (1) Obtain the resource status of all computing domains in the current window and the resource requirements of the task flow; (2) Make resource adaptation decisions based on the resource requirements of the task flow and the resource constraints within the domain (i.e., which computing domain's resources to use to process the task flow); (3) Evaluate the system performance based on the overall resource utilization feedback of the window.

[0070] The resources within each computing domain can be represented by a three-parameter model. in These represent the GPU, CPU, and storage resources allocated to computation and storage within window i, respectively. Therefore, the resource allocation of CC within window i can be represented by the following matrix:

[0071]

[0072] Among them, GPU i CPU i and Store i Let represent the total GPU resources, total CPU resources, and total storage resources within the M computing domains of the window, respectively. Within window i, the adaptation algorithm deployed on CC determines which DS to distribute the task flow to for processing. The resource requirements of each computing task flow are represented as the usage of the corresponding resources. Specifically, resource allocation refers to which of the three types of resources in the computing domain is used to process the task flow generated by the service. GPU resources are allocated in units of CUDA cores, CPU resources in units of processor cores, and storage resources in units of MB. In summary, the resource adaptation decision for computing tasks within window i can be represented as a binary variable DR. i (m,u), define DR i (m,u)∈{0,1} indicates whether to use resources within the computation domain m to process the computation task flow u, with 0 indicating no and 1 indicating yes. Because the types and requirements of computation tasks are constantly changing, resource allocation decisions need to be made dynamically within each window i.

[0073] To ensure the smooth processing of the task, some constraints are defined using the following equations:

[0074]

[0075] in, and This represents the resources reserved for computation domain m. and This represents the resource requirements of task flow u in window i, specifically the GPU. m CPU m and Store m This represents the total resources of computing domain m. The first three constraints ensure that the reserved resources plus the consumed resources do not exceed the resource capacity of the computing domain. The last constraint ensures that each computing task can only be assigned to one computing domain for mirrored computing within a window.

[0076] The goal of resource adaptation is to improve overall resource utilization. Resource utilization within each computing domain is expressed as:

[0077] in:

[0078]

[0079] Therefore, the objective function for resource adaptation is expressed as:

[0080]

[0081]

[0082] For convenience, denoted as ∑ m∈M UR m .

[0083] The overall resource adaptation algorithm is as follows:

[0084] Sorting and initialization: Sort the tasks according to their resource requirements and deadlines. Prioritize tasks with higher resource requirements and shorter tolerance times.

[0085] Traverse all tasks: Starting with the highest priority, traverse all computational domains to find the optimal domain that can meet the resource requirements of the task.

[0086] Achieve load balancing: Among the domains that meet the conditions, select the domain with lower resource utilization for allocation, thereby achieving load balancing while improving resource utilization.

[0087] Cross-domain collaborative scheduling algorithms based on heavy decision-making greed include: multi-objective weighted sorting, dynamic candidate set reduction, fragment-aware adaptation, and two-level reassignment mechanism.

[0088] Multi-objective weighted:

[0089] Design a composite weighting function to prioritize tasks:

[0090]

[0091] Where α and β are resource type weight coefficients, and γ is the time delay weight. and This indicates the resource requirements and latency requirements of the task flow u. It can assign higher weights to computationally intensive tasks (high GPU / CPU requirements) and prioritize resource allocation, making different resource dimensions comparable. This sorting mechanism allows the algorithm to prioritize critical tasks.

[0092] Dynamic candidate set reduction: For each task, the algorithm dynamically constructs a set of candidate domains that meet the conditions.

[0093]

[0094] in, It is a set of computational domains that meet the resource requirements of task flow u. This represents the three types of resources currently available in computation domain m. This represents the resource reservation threshold for computation domain m, used to ensure system stability. This represents the three resource quantities required by task flow u. When the available resources of computing domain m are all greater than the resource quantities required by the task flow, it is a usable computing domain. By constructing a dynamic candidate set of available computing domains, the algorithm avoids invalid resource allocation attempts, improves time efficiency, and at the same time, the reservation mechanism ensures that the system maintains basic service capabilities.

[0095] Fragment perception adaptation:

[0096] Define a multidimensional resource utilization vector and a comprehensive compactness function to achieve resource fragmentation-aware allocation:

[0097] First, calculate the normalized utilization rate of various resources in each domain:

[0098]

[0099] in, This represents the utilization rate of the three resources in computational domain m. This represents the remaining resources within the computational domain m.

[0100] Then, the overall resource compactness of computational domain m is defined, which is the average resource utilization rate of the three resources:

[0101]

[0102] Finally, the most suitable domain is selected based on the principle of minimum compactness:

[0103]

[0104] Fragment-aware adaptation strategy uses compactness function Φ mThis approach effectively addresses three types of fragmentation issues in resource allocation: balancing the utilization of heterogeneous resources to avoid waste caused by single resource bottlenecks (e.g., when one resource (e.g., GPU) in a computing domain is exhausted while other resources (e.g., CPU) are sufficient, the remaining resources cannot be fully utilized); avoiding small-capacity discrete fragmentation by concentrating the use of low-utilization domain resources while preserving the integrity of other domains to meet the demands of large tasks; and achieving global load balancing by minimizing resource distribution deviations between domains. Multi-dimensional fragmentation management improves resource utilization efficiency in heterogeneous environments and reduces task allocation failure rates.

[0105] Secondary reallocation mechanism:

[0106] For tasks that fail to be assigned in the first stage, i.e., tasks that fail to be assigned in the first round, the algorithm introduces a second-stage reassignment mechanism:

[0107]

[0108] Among them, T failed Represents the set of failed tasks in the first-level allocation, var(Φ m ) represents the variance of resource compactness across all domains, DR i (m,u)∈{0,1} are decision variables, representing whether to assign task flow u to computation domain m. The core purpose of this formula is to minimize the variance of resource utilization of all computation nodes through a quadratic allocation strategy. Constraints Ensure that each set T from the first-level allocation failures failed The task flow is allocated to a single computing domain, thereby minimizing the variance of global resource distribution while ensuring feasible task scheduling.

[0109] The two-level reassignment mechanism improves algorithm performance through four key features: it employs a global perspective for reordering, making decisions based on the latest resource state; it utilizes relaxation conditions for exploration, traversing from the domain with the lowest resource utilization to the next low-utilization domain, optimizing overall resource balance; and it constructs a decision matrix to ensure each task is assigned to at most one domain, while minimizing variance. This hierarchical allocation strategy improves task allocation success rate, optimizes resource balance, and enhances overall system performance. The re-decision mechanism combines the resource utilization of the computational domain to decide on computational nodes, avoiding the local optima problem inherent in traditional greedy algorithms.

[0110] This paper proposes a deterministic scheduling scheme based on DRL for small-scale scheduling subproblems, including: the overall framework of deterministic scheduling for wired and wireless networks, the description and transformation of packet demand and resource matching problems, and the design of a deterministic scheduling algorithm based on D3QN.

[0111] The overall framework for deterministic scheduling of wired and wireless networks includes a network model, a latency model, and a wired and wireless converged scheduling model.

[0112] Network Model:

[0113] In the scheduling phase, a converged wireless and wired network framework supporting deterministic scheduling is considered. Each DS is connected via a 5G system and a TSN system, and deterministic scheduling between computing domains is achieved through deterministic mechanisms and constraints. The TSN system connects to the 5G system via two interfaces, DS-TT and NW-TT. The proposed scheduling framework operates in scheduling windows after a scheduling decision. Within a time window t∈V={1,2,3,…,V}, the scheduling decision remains unchanged. In each window t, the specific functions of CC are as follows: (1) In the current scheduling window, obtain the DS address that generates the task flow, the destination DS address of the task flow obtained from the resource adaptation part, and the latency requirements of the data packets; (2) Based on the destination DS address of the task flow and the acceptable latency of each data packet, make scheduling decisions for the data packets; (3) Evaluate the system performance based on the overall successful scheduling rate feedback of the scheduling window.

[0114] Delay model:

[0115] The scheduling process operates on a window-by-window basis, where the window size is the size of the time slot. For efficient operation, each task flow is subject to a corresponding end-to-end latency. Specifically, each divisible service is divided into U independent task flows. After the first resource adaptation phase determines the destination address of the task flow, the V data packets of the specific task flow are scheduled. The total latency of the V data packets of task flow u must be less than the end-to-end latency of task flow u. Due to constraints, this paper ignores the internal processing latency of data packets within the TSN switch. and transmission delay Considering only the propagation delay, the delay of data packet v is... It consists of 5 parts, defined as follows:

[0116] Propagation delay from TSN terminal to TSN-GM The propagation delay of data packet v from the TSN terminal device to the TSN bridge system.

[0117] Propagation delay from TSN-GM to DS-TT port The propagation delay of data packet v from the TSN bridge system to the DS-TT port.

[0118] Overall latency from entering 5GS to exiting 5GS This article considers the latency from when a data packet enters the 5G system to when it leaves the 5G system, and includes a total of 5 parts, among which d DS-TT-UEd represents the latency from DS-TT forwarding to the UE. air This represents the air interface latency from UE to gNB. This represents the processing latency of the computation task within the gNB. d represents the latency from gNB to UPF. UPF-NW-TT This represents the latency of the computing task being forwarded from the UPF to the NW-TT. The above describes the total latency of data packet v within the 5G system, denoted as...

[0119] Propagation delay from NW-TT port to TSN-GM Propagation delay from the NW-TT port to the TSN bridge system.

[0120] Transmission from TSN-GM to TSN terminal The propagation delay of data packet v from the TSN bridge system to the TSN terminal device.

[0121] In cases where the resource granularity of TSN and 5G is mismatched, to achieve joint scheduling of time slot resources and controllable latency calculation in the 5G-TSN converged network, a mini-time slot is used as the time scheduling unit for the 5G network. The duration of the mini-time slot is very short, and its boundary is kept consistent with the time synchronization period of the TSN network to achieve seamless integration between the two networks. Secondly, when applying the CQF mechanism to the 5G-TSN converged network, the key is to design a unified time slot length to ensure that messages sent by the upstream switch are received by the downstream switch in the same time slot. Messages received by the switch queue must be sent out in the next time slot. For the network architecture used in this paper, there are three scheduling scenarios between any two hops: TSN to TSN, TSN to 5G, and 5G to TSN. Therefore, there are three ways to calculate the time slot size. The maximum value among these three scenarios is taken as the time slot size to ensure controllable end-to-end latency.

[0122] T = maxf0{d TSN-5G ,d 5G-TSN ,d TSN-TSN}

[0123] Right now:

[0124]

[0125] Propagation delay and It depends only on the data volume and link bandwidth. Therefore, if the current network link bandwidth is... The propagation delay can then be expressed as Where size u Size is the data size of task flow u, where V is the number of data packets into which task flow u is divided.u ~V is the size of a data packet.

[0126] In this embodiment, to achieve end-to-end deterministic scheduling performance, the latency constraint for entering and leaving the 5GS is: This constraint can be achieved through appropriate configuration and optimization of the 5G system, such as network configuration optimization (by optimizing the configuration parameters of the 5G network, such as adjusting the processing latency, transmission latency, and queue capacity of the base station (gNB), the latency within the 5G system can be reduced), resource reservation (in the 5G system, reserving sufficient resources for critical services, such as bandwidth and processing capacity, can ensure that the latency requirements of these services are met), and network management (through network management methods, such as dynamically adjusting network resource allocation, to ensure that the latency within the 5G system remains within a predetermined range). In this paper, 5GS is regarded as a logical TSN bridge. Based on this, a CQF mechanism is used to implement traffic shaping. That is, two CQF queues schedule data by periodically alternating between sending and receiving data packets. Each queue can only be in one state within a time slot. The queue in the receiving state cannot send data packets, and the queue in the sending state cannot receive data packets.

[0127] Therefore, d min = (h-1)×T∽(h+1)×T=d max This represents the overall latency range, where T is the timeslot length and h is the number of hops for packet forwarding. During the scheduling phase, a DRL-based algorithm is designed to determine the scheduling path and sending queue for packet v. A path list is pre-defined, with each path consisting of multiple nodes. The algorithm selects a path for each packet and allocates queues to each node along the path. Within each TSN-GM, there are multiple TSN switches, denoted as TS = {TS1, TS2, ..., TS} at the sending end. N The receiver is denoted as TE = {TE1, TE2, ..., TE}. N When data packets pass through different sending and receiving TSN switches, they form different scheduling paths. The set of paths is represented by P, denoted as P = {DS, TS, 5GS, TE, DE}. For example, Path1 = {DS1, TS1, TS2, 5GS, TE3, DE2} or Path2 = {DS1, TS1, 5GS, TE4, DE2}, where DS1 is the source sending terminal and DE2 is the receiving terminal. When path 1 is selected, the number of hops is 5, and when path 2 is selected, the number of hops is 4. At the same time, selecting different queues to send data packets v will result in different waiting delays.

[0128] Specifically, if the data packet selects the receive queue, the waiting delay is the remaining time of the current receive time slot; if the data packet selects the send queue, it needs to wait for the remaining time of the current send time slot, as well as the next complete receive time slot, without considering the case where the data packet arrives at the time slot boundary, as follows:

[0129]

[0130] Where J represents the selection of the queue, Q t and Q r These represent the transmit and receive queues for the current time slot, respectively. This means when J equals Q t hour When J is not equal to Q t hour t represents the arrival time of the data packet, and s represents the start time of the current time slot. If the data packet is in the receive queue Q... r If the data packet is scheduled, the waiting delay within the TSN switch is (T-(ts)). t If the data packet is scheduled, the waiting time in the TSN switch is (T-(ts)+T).

[0131] The end-to-end total latency of task flow u is:

[0132]

[0133] Therefore, the total latency of computing service k is:

[0134]

[0135] The completion of a data packet needs to take into account latency constraints, namely:

[0136]

[0137] Right now:

[0138]

[0139] in, h represents the maximum end-to-end latency that a 5G-TSN network may experience. v This represents the number of hops that data packet v has traversed. Given the end-to-end delay constraint for data packets v in task flow u, the delay constraints for the entire computing service are:

[0140]

[0141] Wired and wireless converged scheduling model:

[0142] The service is divided into U task flows, and each task flow is further divided into V data packets for scheduling. Each path is treated as a whole, consisting of multiple nodes. Action decisions are made based on the latency and resource characteristics of each data packet, combined with the current environment. Thus, the processing time (time slot) and queue size of each node on the path are predetermined by the scheduling algorithm. The decision variable for the scheduling part is defined as DT. t (path, queue), where path ∈ {0, 15} and queue ∈ {0, 1}, indicates that data packet v is scheduled using path path and queue queue in the j-th time slot. For example, the algorithm selects path path v. Where e0 is the source node, If it is the destination node, then the transmission of data packet v along the path path requires l p A series of consecutive time slots, the q-th node e on the path q The data packet is processed in the (j+q)th time slot, where j represents the time slot where packet scheduling begins. An indicator variable X is defined. t (u,v,j,p) indicates whether data packet v of task flow u is scheduled using path p in time slot j. Several important constraints are defined below:

[0143] 1) Path integrity constraints:

[0144] To ensure that data packets pass through all nodes sequentially along the paths in the path set, we have:

[0145]

[0146] This is achieved through the following constraints:

[0147]

[0148] Where, δ ql,p It is an indicator function, indicating that if node q l If a node belongs to the l-th node of path p, the value is 1; otherwise, it is 0. The above formula indicates that each node can only process one data packet in a specific time slot, avoiding multiple data packets competing for the same node's queue resources simultaneously.

[0149] 2) Queue capacity constraint:

[0150] With the sending time determined, the resources occupied by each queue in the current scheduling period can be calculated:

[0151]

[0152] To ensure successful data packet transmission, the total number of data packets in each queue within a time slot cannot exceed the maximum capacity of the queue.

[0153]

[0154] here, Let be the maximum capacity of the transmit queue within time slot j. The capacity constraint for the receive queue is similar. We have already assumed that the size of a time slot is the same as the size of a queue, meaning that each time slot can only schedule data from one queue, or process only one data packet per time slot. Therefore, the size of the data packet does not need to be considered when calculating the queue capacity constraint.

[0155] 3) Path selection uniqueness:

[0156] Considering only the case where each data packet can only choose one path and one starting timeslot for scheduling:

[0157]

[0158] Among them, X t (u,v,j,p)=1 indicates that data packet v of task flow u is scheduled using path p in time slot j.

[0159] 4) Link bandwidth constraints:

[0160] The total bandwidth occupied by the same link in the (j+l)th time slot cannot exceed have:

[0161]

[0162] Where, p l This represents the link of path p.

[0163] 5) Two-stage association:

[0164] During the resource adaptation phase, the subsequent scheduling phase needs to be considered to ensure successful task scheduling within latency constraints. Simultaneously, the scheduling phase also needs to consider the output of the resource adaptation phase to ensure tasks are assigned to the correct servers for computation. To ensure that the second-phase computation tasks are scheduled to servers already selected by the resource adaptation decision in the first phase, a ternary variable is introduced. In the i-th window, when the v-th data packet of the u-th task flow is assigned to the m-th DS and the path p with destination node m is selected... m hour Otherwise, it is 0. Therefore, the constraint can be expressed as:

[0165]

[0166] The three constraints above together ensure that Only when task flow u is assigned to the m-th computation domain (DR) i (m,u)=1) and using path p m Scheduling (X)t The value is 1 only when (u,v,j,p) is true; otherwise, it is 0.

[0167] The description and transformation of the data packet demand and resource matching problem includes: objective function and Markov decision process.

[0168] Objective function:

[0169] This embodiment focuses on improving the successful scheduling rate during the scheduling phase, defining an indicator function. Whether scheduling is successful depends on whether the latency requirement is met:

[0170]

[0171] Where 1 is the set characteristic function, taking the value 1 when the condition is met and 0 otherwise. This represents the upper limit of the delay for the v-th data packet. Not exceeding its delay requirements The value is 1 if the condition is met, and 0 otherwise. Therefore, the success rate of a computing service is expressed as:

[0172]

[0173] The formula means that the successful scheduling rate equals the number of task flows that meet the latency requirements divided by the total number of task flows in the service being calculated. Here, it considers that the total end-to-end latency of the V data packets of a task flow u meets the end-to-end latency constraint of task flow u. Therefore, the successful scheduling rate of all services is expressed as:

[0174]

[0175] Therefore, the objective function in the scheduling part is expressed as:

[0176] P1:max DT (SR),

[0177]

[0178] Markov Decision Process: In the scheduling part, since the scheduling progress of each data packet v of the future computation task flow u depends only on the current remaining data packet quantity and the remaining resource status of each node, and is independent of the history of previous data packets, this is a Markov Decision Process (MDP) problem.

[0179] In each scheduling window, the agent can observe the current network state s. t And make time slot and path allocation decisions for each data packet. t, where t∈[1,|V|], and |V| is the total number of packets in each training round. This means that in MDP, the time index t represents a packet index in the set of packets that need to be scheduled. In order to learn an optimal packet scheduling strategy, it is necessary to accurately assign the state s to the appropriate state. t Mapping to decision a t The corresponding reward r is fed back from the network environment. t And accordingly, the state is transformed into a new state s. t+1 It defines the three core elements of MDP: state, action, and reward, as follows:

[0180] Status: In scheduling window t, agent CC collects data packet information from the DS, TSN bridge, and 5GS system. and the current queue capacity information The state s under consideration t ∈S can be defined as:

[0181]

[0182] in, This includes the latency requirements for data packets. It can be described by the following formula:

[0183]

[0184] here Indicates time t, link l m In time interval T n The remaining capacity at that time.

[0185] Actions: Based on observed states, the agent can formulate real-time scheduling strategies to determine the path and time slot of data packets v in task flow u, thereby meeting overall service latency requirements. In this paper, the action space is described by a tuple (path, queue). Specifically, the action space A is divided into two dimensions. The first dimension, actions, is the path P = {DS, TS, 5GS, TE, DE}, where DS represents the list of source computation domains, DE represents the list of destination computation domains, and TS = {TS1, TS2, ..., TS}. n} represents the devices in the source TSN bridge system, 5GS represents the 5G system, and TE = {TE1, TE2, ..., TE...} n} represents the device in the destination TSN bridge system. The scheduling part makes decisions based on the intermediate nodes TS and TE. The second-dimensional action is T = {T1, T2, ..., T}. n} indicates when these data packets are sent in the 5G-TSN converged network. Therefore, policy π * S→A can be described as:

[0186]

[0187] Here, [ε≥ε0-N] is equivalent to an indicator function, which is 1 when the condition ε≥ε0-N is satisfied and 0 when it is not satisfied. ε0∈(0,1) is the initial probability value, which is a random number. N is regenerated in each step t as the decay factor for each learning step. Each action element is an integer value, reshaped into the following range [0,CQF]. num ·Path num -1]. The action indicates which queue and path packet v should be scheduled in. Note that each packet can only select one queue and one path for skipping.

[0188] Reward: Once the agent takes action a t It will receive a reward for evaluating the state. t Take the following action r t How good is it? With the goal of maximizing the successful scheduling rate, the reward function of the TSched algorithm can be expressed as:

[0189]

[0190] Here, [success] and [failed] are indicator functions, representing a value of 1 when the condition is met and 0 when it is not. When the current data packet v is successfully scheduled, the agent will receive a positive reward, i.e. Here R succ It is a positive constant value. This indicates that the data packet was successfully scheduled after the scheduling decision. Conversely, the agent will receive a negative constant value R. fail This is used to punish inappropriate decisions. Clearly, the reward is related to the successful scheduling rate in the current round. The obtained reward is used to optimize the policy DT during the training phase until the algorithm converges.

[0191] In MDP, the agent's goal is to find the optimal queue path scheduling policy DT that maximizes the cumulative discount reward, i.e.:

[0192]

[0193] Where γ∈[0,1] is the discount factor. Strategy π DT This specifies how the central controller schedules queue paths based on the observed state in each scheduling cycle. The entire formula seeks an optimal policy π. DT This makes the current state s t Action pair (s) t ,a t Starting with this strategy, the expected value of the future cumulative discount reward vector obtained is maximized.

[0194] The design of a deterministic scheduling algorithm based on D3QN includes: algorithm initialization, empirical sampling, and neural network training.

[0195] Algorithm Initialization: Once the training phase begins, the agent's online and target network parameters are initialized and then updated incrementally. Note that only one data packet is scheduled per learning step, therefore the size of the learning step is equal to the total number of data packets to be scheduled. Furthermore, an experience replay buffer B is instantiated with a capacity of D to store sampled experience that can be used to update the algorithm parameters. Then, the considered network environment is reset to generate the initial state s. 1 .

[0196] Experience Acquisition: After the initialization phase, the agent's learning process can begin by sampling interaction experiences from the network environment and storing them in the replay buffer B. This can be represented as a five-element tuple {s}. b ,a b ,r b ,s b+1 ,done}. Specifically, first, a strategy is selected based on the exploration rate ε, either a random strategy or an action is selected using Dueling DQN. Then, the decision a is executed. t This allocates resources to the current data packet and grants an immediate reward. Then, the network state transitions to a new state s. t+1 Furthermore, it obtains the available resources for the next packet in the considered integrated wireless and wired TSN network, and uses the done∈{True,False} label to indicate whether node and queue resources have been exhausted so far. The corresponding experience tuples are stored in the replay buffer B for algorithm training. When the buffer exceeds its capacity, the oldest tuple is replaced with a new experience tuple.

[0197] Neural Network Training: As mentioned earlier, both the online and target networks are trained using sampling experience. In each algorithm step, the agent randomly selects a D from the replay buffer. b Small batches of experience are used to update the parameters of both networks. Double DQN is used to calculate the target value, and the online network is calculated using the formula... Q(s b+1 ,a b+1 ;θ) Select action, target network through To evaluate the action, the mean squared error is used to calculate the loss function L(θ) = MSE(Q(s)). b ,a b ;θ),Q D3QN Then, the loss function is minimized using the gradient descent method, i.e.:

[0198]

[0199] Secondly, to ensure the stability of the training process, a soft update method is adopted, which slowly updates the parameters θ of the target network by tracking the online network. - ,Right now:

[0200] θ - =∈θ+(1-∈)θ -

[0201] Where ∈ (0,1) is the update ratio of the target network.

[0202] Example 3

[0203] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When these instructions are executed by a processor, they implement the above-described method for elastic adaptation and deterministic scheduling of heterogeneous resources in an intelligent computing converged network. This method includes: terminals within each computing domain establishing connections with the DS via a TSN, and subsequently sending all information to the domain server; the control plane collecting status information to make network configuration decisions; and adopting a distributed management and centralized control mode. In centralized control, the centralized controller is responsible for distributing control lists and forwarding tables of all TSN switches, determining the computing domain, path, and transmission time of all task flows. For distributed management, terminals report to the domain server DS, forming a computing domain; the centralized controller CC determines the computing domain, path, and transmission time of the task flows based on information collected from the DS; and the data plane schedules the task flows according to the scheduling table from the CC.

[0204] Example 4

[0205] This embodiment 4 provides a computer device, including a memory and a processor. The processor and the memory communicate with each other. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the above-described method for elastic adaptation and deterministic scheduling of heterogeneous resources in a converged intelligent computing network. The method includes: terminals in each computing domain establish a connection with the DS through a TSN, and then the terminals send all information to the domain server; the control plane collects status information to make network configuration decisions; a distributed management and centralized control mode is adopted; in centralized control, the centralized controller is responsible for distributing the control lists and forwarding tables of all TSN switches and determining the computing domain, path, and transmission time of all task flows; for distributed management, the terminals report to the domain server DS to form a computing domain; the centralized controller CC determines the computing domain, path, and transmission time of the task flow based on the information collected from the DS; the data plane schedules the task flow according to the scheduling table from the CC.

[0206] Example 5

[0207] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the above-described method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks. The method includes: terminals in each computing domain establish a connection with the DS through TSN, and then the terminals send all information to the domain server; the control plane collects status information to make network configuration decisions; a distributed management and centralized control mode is adopted; in centralized control, the centralized controller is responsible for distributing the control lists and forwarding tables of all TSN switches and determining the computing domain, path, and transmission time of all task flows; for distributed management, the terminals report to the domain server DS to form a computing domain; the centralized controller CC determines the computing domain, path, and transmission time of the task flow based on the information collected from the DS; the data plane schedules the task flow according to the scheduling table from the CC.

[0208] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0209] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0210] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0211] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0212] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A system for elastic adaptation and deterministic scheduling of heterogeneous resources in an intelligent computing fusion network, characterized in that, include: The system comprises a domain server (DS) that manages resources within a computing domain, a centralized controller (CC) that acquires resource information from all devices, and time-sensitive network devices (TSNs), including terminals, TSN switches, and a centralized network configuration (CNC). Terminals within each computing domain establish connections with the DS via TSNs, and then send all information to the domain server. The control plane collects status information and makes network configuration decisions. A distributed management and centralized control model is employed. In centralized control, the central controller distributes control lists and forwarding tables from all TSN switches, determining the computing domain, path, and transmission time for all task flows. For distributed management, terminals report to the domain server (DS), forming a computing... The computational domain; the centralized controller (CC) determines the computational domain, path, and transmission time of the task flow based on information collected from the data storage (DS); the data plane schedules the task flow according to the scheduling table from the CC; among them, a two-stage stochastic optimization problem with different time scales is established based on the system of distributed management and centralized control, including: a resource adaptation sub-problem for large time scales and a scheduling sub-problem for small scales; for the resource adaptation sub-problem at the large scale, a resource adaptation algorithm based on heuristic algorithms is used, including a cross-domain intelligent computing resource elastic adaptation mechanism framework, a description and modeling of task flow requirements and resource matching problems, and a cross-domain collaborative scheduling algorithm based on re-decision greed; the cross-domain intelligent computing resource elastic adaptation mechanism framework runs in a window after decision-making, denoted as Within a window, the resource adaptation decision remains unchanged. In each window, the DS obtains the available resources and resource requirements of the task flow within adjacent computing domains. The central controller obtains the resource status of all computing domains and the resource requirements of the task flow within the current window. Based on the task flow resource requirements and the resource constraints within the domain, it makes resource adaptation decisions. The cross-domain collaborative scheduling algorithm based on re-decision greedy scheduling includes: multi-objective weighted sorting, dynamic candidate set reduction, fragment-aware adaptation, and a two-level reallocation mechanism. Fragment-aware adaptation includes: Define a multidimensional resource utilization vector and a comprehensive compactness function to achieve resource fragmentation-aware allocation: First, calculate the normalized utilization rate of various resources in each domain: ； in, This represents the utilization rate of the three resources in computational domain m. This represents the remaining resources within the computational domain m; Then, the overall resource compactness of computational domain m is defined, which is the average resource utilization rate of the three resources: ； Finally, the most suitable domain is selected based on the principle of minimum compactness: ； The two-level reallocation mechanism includes: for tasks that failed in the first allocation (i.e., tasks that failed in the first allocation), a second-level reallocation mechanism is introduced to minimize the variance of resource utilization across all computing nodes through a secondary allocation strategy. Constraints Ensure that each set of failed assignments from the first level is not affected. The task flow is allocated to a single computation domain, thereby minimizing the variance of global resource distribution while ensuring feasible task scheduling. ； in, This represents the set of tasks that failed in the first-level allocation. This represents the variance of resource compactness across all domains. Let be the decision variable, indicating whether to assign task flow u to computation domain m.

2. The intelligent computing converged network heterogeneous resource elastic adaptation and deterministic scheduling system according to claim 1, characterized in that, The overall resource adaptation algorithm of the heuristic algorithm is as follows: Sorting and initialization: Sort the tasks according to their resource requirements and deadlines; prioritize tasks with large requirements and short tolerance times; Traverse all tasks: Starting from the highest priority, traverse all computational domains to find the optimal domain that can meet the resource requirements of the task; Achieve load balancing: Among the domains that meet the conditions, select the domain with lower resource utilization for allocation, thereby achieving load balancing while improving resource utilization.

3. The intelligent computing convergence network heterogeneous resource elastic adaptation and deterministic scheduling system according to claim 2, characterized in that, Multi-objective weighted summaries include: Design a composite weighting function to prioritize tasks: ， in, and For resource type weighting coefficients, For delay weighting, This indicates the resource requirements and latency requirements of the task flow u. It assigns higher weights to computationally intensive tasks, prioritizes resource allocation, and makes different resource dimensions comparable. This sorting mechanism enables the algorithm to prioritize critical tasks.

4. The intelligent computing converged network heterogeneous resource elastic adaptation and deterministic scheduling system according to claim 3, characterized in that, Dynamic candidate set reduction includes: for each task, the algorithm dynamically constructs a set of candidate domains that meet certain conditions. ； in, It satisfies the task flow The set of computational domains for resource requirements. This represents the three types of resources currently available in computation domain m. This represents the resource reservation threshold for computation domain m, used to ensure system stability. This represents the three resource quantities required by task flow u; when the three available resource quantities of computing domain m are all greater than the resource quantities required by task flow, it is an available computing domain, and a dynamic candidate set of available computing domains is constructed.

5. A method for elastic adaptation and deterministic scheduling of heterogeneous resources in an intelligent computing fusion network based on the system described in any one of claims 1-4, characterized in that, include: Terminals within each computing domain establish a connection with the DS via the TSN, and then the terminal sends all information to the domain server; The control plane collects status information to make network configuration decisions; it adopts a distributed management and centralized control mode; in centralized control, the central controller is responsible for distributing control lists and forwarding tables of all TSN switches and determining the computation domain, path, and transmission time of all task flows; for distributed management, the terminals report to the domain server DS to form a computation domain; the central controller CC determines the computation domain, path, and transmission time of the task flow based on the information collected from the DS; the data plane schedules the task flow according to the scheduling table from the CC.

6. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks as described in claim 5.

7. An electronic device, characterized in that, include: The device comprises a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions that implement the method for elastic adaptation and deterministic scheduling of heterogeneous resources in intelligent computing converged networks as described in claim 5.

Images