Task awareness and intelligent scheduling method for convergence network

Abstract

The invention discloses a converged network-oriented task awareness and intelligent scheduling method, and relates to the technical field of communication networks, the method is applied to a converged network, and the method comprises the following steps: constructing a converged network architecture; extracting a user equipment node of the current to-be-processed task from the computing power identifier of the current to-be-processed task as a source node; based on the converged network architecture, determining a corresponding path by using a path search algorithm when the source node is used as a starting point and each computing power center node is used as a destination node; according to the computing power identifier of the current to-be-processed task, selecting an optimal path from paths corresponding to all computing power center nodes; according to the method, the transmission time delay and the service time delay are considered when the optimal path is selected, so that the reasonability of task perception and intelligent scheduling oriented to the fusion network can be improved.

Description

Technical Field

[0001] The present application relates to the field of communication network technology, and in particular to a task perception and intelligent scheduling method for converged networks. Background Art

[0002] Computing-network convergence is a new type of network structure proposed in recent years. Due to the new characteristics and complexity of the network structure in the computing-network convergence network scenario, traditional task scheduling solutions are often not directly applicable to the computing-network convergence scenario. There is still a lack of effective methods and strategies for comprehensive consideration and optimization of factors such as network resources, computing power resources and user needs involved in the converged network. Therefore, there are many challenges and problems to be solved in the computing-network resource scheduling solution.

[0003] At present, there are few research and practical cases related to computing network resource scheduling. How to fully consider network resources and computing power resources, coordinate and schedule tasks initiated by users to computing power centers, and more effectively utilize computing network resources is an urgent problem to be solved. Therefore, it is necessary to conduct in-depth research on computing network resource fusion scheduling solutions, explore new task scheduling algorithms suitable for fusion network application scenarios, and verify and optimize them to better exert the potential of computing network fusion architecture, realize the effective utilization of heterogeneous resources and the efficient execution of user tasks. Summary of the invention

[0004] The purpose of this application is to provide a task perception and intelligent scheduling method for a converged network, which can improve the rationality of task perception and intelligent scheduling for a converged network.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] The present application provides a task perception and intelligent scheduling method for a fusion network, wherein the task perception and intelligent scheduling method for a fusion network is applied to a fusion network, wherein the fusion network includes:

[0007] Computing power analysis routing, computing power network controller, core network, multiple user devices and multiple computing power centers;

[0008] The computing power network controller and the plurality of user devices are connected to the computing power resolution routing; the computing power resolution routing is also connected to the plurality of computing power centers through the core network;

[0009] The user device is used to publish tasks;

[0010] The computing power analysis routing is used to analyze the received task when receiving a task sent by any user device, and determine the computing power identifier according to the analysis result;

[0011] The computing power network controller is used to determine the target computing power center and the optimal path of the core network based on the computing power identifier, the core network link status and the computing power of each computing power center pair;

[0012] The computing power analysis routing is used to transmit the task to the target computing power center according to the optimal path;

[0013] The task perception and intelligent scheduling method for the fusion network includes:

[0014] Constructing a converged network architecture; the converged network architecture includes: a plurality of nodes, a plurality of links, and a weight of each link; any of the nodes is a user device node, a routing node, or a computing center node; the link has a user device node or a routing node as a link start point, and a routing node or a computing center node as a link end point; the link weight represents the transmission cost of transmitting data from the link start point to the link end point of the corresponding link;

[0015] Get the first task in the pending task queue as the current pending task;

[0016] Extract the user device node of the current task to be processed from the computing power identifier of the current task to be processed as the source node;

[0017] Based on the fusion network architecture, a path search algorithm is used to determine the corresponding path when the source node is used as the starting point and each computing power center node is used as the destination node;

[0018] According to the computing power identifier of the current task to be processed, select the optimal path from the paths corresponding to all computing power center nodes;

[0019] Determine the computing power center node corresponding to the optimal path as the optimal destination node;

[0020] Computing power analysis routing sends the current pending task from the source node to the optimal destination node according to the optimal path;

[0021] The optimal destination node generates a processing result after processing the current pending task;

[0022] The computing power analysis routing returns the processing results from the optimal destination node to the source node;

[0023] Delete the current pending task from the pending task queue and return to step "get the first task in the pending task queue as the current pending task".

[0024] Optionally, the task perception and intelligent scheduling method for converged networks further includes:

[0025] When the computing power resolution routing receives a task request sent by any user device, it updates the queue of pending tasks according to the task request.

[0026] Optionally, when receiving a task request sent by any user device, the computing power resolution routing updates the pending task queue according to the task request, including:

[0027] When the computing power resolution routing receives a task request sent by any user device, it determines that the requested task is the current task to be sorted;

[0028] Extracting the task category of the current task to be sorted from the task attributes of the current task to be sorted; the task attributes include task category, storage resource amount, bandwidth requirement and task priority; the task category is a transmission task or a computing task;

[0029] When the current task to be sorted is a computing task, determining the subtask category of the current task to be sorted to be the current subtask category;

[0030] Query the CPU ratio and GPU ratio corresponding to the current subtask category in the task component table; the computing power analysis route stores the task component table; the task component table is used to express the CPU ratio and GPU ratio corresponding to different computing task subtask categories; the sum of the CPU ratio and GPU ratio corresponding to the same computing task subtask category is equal to 1;

[0031] Determine the product of the storage resource amount of the current task to be sorted and the CPU ratio corresponding to the current subtask category as the CPU demand of the current task to be sorted;

[0032] Determine the product of the storage resource amount of the current task to be sorted and the GPU ratio corresponding to the current subtask category as the GPU demand of the current task to be sorted;

[0033] Encoding the CPU requirement, the GPU requirement, and the task attribute to obtain a computing power identifier;

[0034] When the current task to be sorted is a transmission task, encoding the task attribute to obtain a computing power identifier;

[0035] Add the computing power identifier to the message header of the current task to be sorted;

[0036] Based on the priority of the current task to be sorted and the priorities of all pending tasks in the pending task queue, the current task to be sorted is inserted into the pending task queue to complete the update of the pending task queue.

[0037] Optionally, based on the fusion network architecture, a path search algorithm is used to determine a path corresponding to a source node as a starting point and each computing power center node as a destination node, including:

[0038] At the source node, the current pending tasks are divided into data packet queues;

[0039] Determine any computing power center node as the current destination node d;

[0040] Determine that the first data packet in the data packet queue is the current data packet;

[0041] Add the source node s as the first element to the path list;

[0042] Let source node s be current node c;

[0043] Let the cumulative transmission cost of the current data packet to reach the current node be t = 0;

[0044] Let the number of nodes passed h = 1;

[0045] Initialize the current data packet p = {s, d, c, t, h, f}; f indicates whether the current data packet has reached the current destination node d;

[0046] Initialize the minimum Q value to positive infinity;

[0047] Determine any neighbor node of the current node c as the current neighbor node; the neighbor node of the current node c is the link end point corresponding to the link with the current node c as the link start point in the converged network architecture;

[0048] Determine the Q value from the current node c through the current neighbor node to the current destination node d as the current neighbor Q value;

[0049] Determine whether the current neighbor Q value is less than the minimum Q value, and obtain a first determination result;

[0050] If the first judgment result is no, the current neighbor node is updated, and the process returns to step "determining the Q value of reaching the current destination node d from the current node c through the current neighbor node is the Q value corresponding to the current neighbor node";

[0051] If the first judgment result is yes, then the value of the number of nodes passed h is increased by 1;

[0052] Add the current neighbor node as the hth element in the path list;

[0053] Determine whether the number of nodes h passed is greater than the preset path length, and obtain a second determination result;

[0054] If the second judgment result is yes, then initialize the current data packet and add the current data packet to the data packet queue;

[0055] If the second judgment result is no, then determine that the link weight corresponding to the current node c and the current neighbor node is the current link weight;

[0056] Let the value of the cumulative transmission cost t increase the current link weight;

[0057] Determine whether the current neighbor node is the current destination node d, and obtain a third determination result;

[0058] If the third judgment result is no, the current neighbor node is used as the current node, the current neighbor Q value is used as the minimum Q value, and the process returns to step "determining any neighbor node of the current node c as the current neighbor node";

[0059] If the third judgment result is yes, then determining that the path list is a path corresponding to the current destination node;

[0060] The cumulative transmission cost t is determined to be the cumulative transmission cost of the path corresponding to the current destination node.

[0061] Optionally, after determining that the cumulative transmission cost t is the cumulative transmission cost of the path corresponding to the current destination node, the method further includes:

[0062] Update the current destination node and return to step "determine that the first data packet in the data packet queue is the current data packet" until all computing power center nodes are traversed to obtain the paths corresponding to all computing power center nodes.

[0063] Optionally, according to the computing power identifier of the current task to be processed, the optimal path is selected from the paths corresponding to all computing power center nodes, including:

[0064] Extract the task category of the current pending task from the computing power identifier of the current pending task;

[0065] When the current task to be processed is a transmission task, the path corresponding to the minimum cumulative transmission cost is selected from the paths corresponding to all computing power center nodes as the optimal path.

[0066] Optionally, according to the computing power identifier of the current task to be processed, the optimal path is selected from the paths corresponding to all computing power center nodes, and further includes:

[0067] When the current task to be processed is a computing task, the CPU demand and GPU demand are extracted from the computing power identifier;

[0068] Obtain a load balancing factor; the load balancing factor is used to describe the remaining resources of all computing power centers at the current moment;

[0069] Determine the path corresponding to any computing power center node as the current path;

[0070] Determine the objective function value corresponding to the current path according to the objective function, the load balancing factor, the CPU demand, the GPU demand, and the accumulated transmission cost of the current path;

[0071] Update the current path and return to step "determine the objective function value corresponding to the current path according to the objective function, the load balancing factor, the CPU demand, the GPU demand and the accumulated transmission cost of the current path" until all computing power center nodes are traversed to obtain the objective function value of the path corresponding to each computing power center node;

[0072] Determine the path corresponding to the maximum objective function value as the optimal path.

[0073] Optionally, before obtaining the first task in the pending task queue as the current pending task, the following is further included:

[0074] Construct the objective function and determine the constraints of the objective function.

[0075] Optionally, the objective function is:

[0076]

[0077] Among them, I is the objective function, λ j (i) represents the jth user task TU j Whether it is assigned to the i-th computing center CC i , when λ j (i) = 1 indicates user task TU j Assigned to the computing center CC i , when λ j (i) = 0 indicates user task TU j Not assigned to the computing center CC i ; W represents the number of user tasks, N represents the number of computing centers, α represents the service delay weight, Represents task TU j In the computing center CC i The service delay on the network, β represents the transmission delay weight, Represents task TU j The complete data is transmitted to the computing center CC i The cumulative transmission delay of , γ represents the weight of the load balancing factor, and dr represents the load balancing factor.

[0078] Optionally, the constraint condition is:

[0079]

[0080] Among them, c j Represents the jth user task TU j The CPU demand, C i Represents the i-th computing center CC i The remaining CPU resources, g j Denotes the jth user task TUj GPU demand, G i Represents the i-th computing center CC i The remaining GPU resources, s j Denotes the jth user task TU j The storage resource requirement, S i Represents the i-th computing center CC i remaining storage resources.

[0081] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0082] The present application provides a task perception and intelligent scheduling method for converged networks, designs a dynamic resource perception and user task priority scheduling mechanism, which comprehensively considers factors such as task characteristics and user attributes, estimates the task's demand for heterogeneous computing resources, and adds the parsed information to the task's computing power identifier; by introducing a deep reinforcement learning algorithm, it intelligently perceives the network link status and traffic characteristics, and autonomously learns and decides on the transmission path of data packets; it solves the problems of coordinated processing of computing tasks and transmission tasks and dynamic scheduling of computing center resources in computing-network fusion scenarios, and designs a dynamic resource scheduling algorithm that comprehensively considers task types and characteristics, network link conditions, and coordinated computing center resource surplus conditions to adapt to the diversified and dynamic communication needs in converged network scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0084] Figure 1 This is a schematic diagram of the fusion network structure in an embodiment of the present application;

[0085] Figure 2 This is a flow chart of a task perception and intelligent scheduling method for a converged network in an embodiment of the present application;

[0086] Figure 3 This is a schematic diagram of a task perception and intelligent scheduling method for a converged network in an embodiment of the present application;

[0087] Figure 4 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0088] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0089] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0090] In an exemplary embodiment, a task perception and intelligent scheduling method for a converged network is provided. The task perception and intelligent scheduling method for a converged network is applied to a converged network, such as Figure 1 , the fusion network includes: computing power resolution routing, computing power network controller, core network, multiple user devices and multiple computing power centers. The computing power network controller and multiple user devices are connected to the computing power resolution routing. The computing power resolution routing is also connected to multiple computing power centers through the core network. User devices are used to publish tasks. The computing power resolution routing is used to parse the received tasks when receiving tasks sent by any user device, and determine the computing power identification according to the analysis results. The computing power network controller is used to determine the target computing power center and the optimal path of the core network based on the computing power identification, the core network link status and the computing power of each computing power center pair. The computing power resolution routing is used to transmit the task to the target computing power center according to the optimal path.

[0091] The converged network includes: user equipment, computing power resolution routing, computing power network controller, core network and computing power center.

[0092] User devices are the initiators of tasks, including smartphones, personal computers, IoT devices, etc. Assume that there are M user devices that can initiate tasks in this model, and let U = {U1, U2, ..., U M} to represent a collection of user devices. User tasks are computing or transmission tasks that users expect to complete through the computing network. They can be divided into computing tasks and transmission tasks according to whether computing resources are required. Computing tasks refer to computing tasks that require the use of heterogeneous computing resources in the computing center, such as cloud code compilation, artificial intelligence model training, 3D modeling and rendering, etc. Transmission tasks refer to data transmission tasks that mainly occupy network bandwidth resources and have little or no CPU computing power requirements, such as high-definition video streaming, remote backup and synchronization, and large file downloads.

[0093] The computing power resolution router is a key node for users to upload tasks to the core network through various access methods such as WIFI, broadband, and mobile data. The computing power resolution router has the function of intelligent task resolution. Through the pre-trained deep learning model, it can accurately predict the user's task demand for heterogeneous computing power resources and storage resources, and convert this information into a computing power identifier to label the user's task. The computing power identifier is then sent to the computing power network controller for unified management and scheduling.

[0094] The computing network controller is a key device for computing resource scheduling. It comprehensively considers multiple factors such as the computing power identification of the user task, the current link status of the core network, and the load of the computing center. It schedules the user task to the most suitable computing center through the scheduling algorithm, thereby minimizing the total delay from task initiation to completion and maximizing the resource utilization of the computing center.

[0095] The core network is a high-speed backbone network that connects the computing power analysis routing and the computing power center. It is responsible for providing high-quality, low-latency data transmission services between the computing power center and the user. After the computing power network controller determines the specific routing path and the target computing power center of the user task, the core network can upload the task to the computing power center according to the specified path to ensure the end-to-end communication quality. Assume that there are a total of K routing nodes in the model, with R = {R1, R2, ..., R K} to represent a collection of routing nodes.

[0096] The computing center is a high-performance computer cluster deployed at the edge or core of the network. It is equipped with high-performance CPU, GPU, FPGA and other heterogeneous computing resources to meet the processing requirements of various computing tasks. At the same time, the computing center also deploys large-capacity storage devices to provide efficient data access and storage services for various transmission tasks. Assume that there are N computing centers in the model, CC = {CC1, CC2, ..., CC N} to represent the set of computing power centers.

[0097] like Figure 2 , a task perception and intelligent scheduling method for converged networks, including:

[0098] Step 101: Construct a converged network architecture. The converged network architecture includes: multiple nodes, multiple links, and the weight of each link. Any node is a user device node, a routing node, or a computing center node. The link starts with the user device node or the routing node and ends with the routing node or the computing center node. The weight of the link represents the transmission cost of transmitting data from the link start point to the link end point of the corresponding link.

[0099] Step 102: Obtain the first task in the pending task queue as the current pending task.

[0100] Step 103: extracting the user equipment node of the current task to be processed from the computing power identifier of the current task to be processed as the source node.

[0101] Step 104: Based on the fusion network architecture, a path search algorithm is used to determine the corresponding path with the source node as the starting point and each computing center node as the destination node.

[0102] Step 105: According to the computing power identifier of the current task to be processed, select the optimal path from the paths corresponding to all computing power center nodes.

[0103] Step 106: Determine the computing power center node corresponding to the optimal path as the optimal destination node.

[0104] Step 107: The computing power analysis routing sends the current pending task from the source node to the optimal destination node according to the optimal path.

[0105] Step 108: The optimal destination node generates a processing result after processing the current task to be processed.

[0106] Step 109: The computing power resolution routing returns the processing result from the optimal destination node to the source node.

[0107] Step 1010: Delete the current pending task from the pending task queue and return to step 102.

[0108] In addition, the task perception and intelligent scheduling method for converged networks also includes:

[0109] Step 1011: When the computing power resolution router receives a task request sent by any user device, it updates the queue of pending tasks according to the task request.

[0110] Step 1011 includes:

[0111] When the computing power resolution routing receives a task request sent by any user device, it determines that the requested task is the current task to be sorted.

[0112] The task category of the current task to be sorted is extracted from the task attributes of the current task to be sorted. The task attributes include task category, storage resource amount, bandwidth requirement and task priority. The task category is a transmission task or a computing task.

[0113] When the current task to be sorted is a computing task, it is determined that the subtask category of the current task to be sorted is the current subtask category.

[0114] Query the CPU ratio and GPU ratio corresponding to the current subtask category in the task component table. The task component table is stored in the computing power analysis route. The task component table is used to express the CPU ratio and GPU ratio corresponding to different computing task subtask categories. The sum of the CPU ratio and GPU ratio corresponding to the same computing task subtask category is equal to 1.

[0115] The product of the storage resource amount of the current task to be sorted and the CPU ratio corresponding to the current subtask category is determined as the CPU demand of the current task to be sorted.

[0116] The product of the storage resource amount of the current task to be sorted and the GPU ratio corresponding to the current subtask category is determined as the GPU demand of the current task to be sorted.

[0117] The CPU demand, GPU demand and task attributes are encoded to obtain the computing power identifier.

[0118] When the current task to be sorted is a transmission task, the task attributes are encoded to obtain a computing power identifier.

[0119] Add the computing power ID to the message header of the current task to be sorted.

[0120] Based on the priority of the current task to be sorted and the priorities of all pending tasks in the pending task queue, the current task to be sorted is inserted into the pending task queue to complete the update of the pending task queue.

[0121] Step 104 includes:

[0122] At the source node, the current pending tasks are divided into data packet queues.

[0123] Determine any computing power center node as the current destination node d.

[0124] Determine the first data packet in the data packet queue as the current data packet.

[0125] Add the source node s as the first element to the path list.

[0126] Let the source node s be the current node c.

[0127] Let the cumulative transmission cost of the current data packet reaching the current node be t=0.

[0128] Let the number of nodes passed be h=1.

[0129] Initialize the current data packet p = {s, d, c, t, h, f}; f indicates whether the current data packet has reached the current destination node d.

[0130] Initialize the minimum Q value to positive infinity.

[0131] Determine any neighbor node of the current node c as the current neighbor node. The neighbor node of the current node c is the link end point corresponding to the link with the current node c as the link start point in the converged network architecture.

[0132] Determine the Q value from the current node c through the current neighbor node to the current destination node d as the current neighbor Q value.

[0133] It is determined whether the current neighbor Q value is less than the minimum Q value to obtain a first determination result.

[0134] If the first judgment result is no, the current neighbor node is updated, and the process returns to step "determining the Q value of reaching the current destination node d from the current node c through the current neighbor node is the Q value corresponding to the current neighbor node".

[0135] If the first judgment result is yes, the value of the number of passed nodes h is increased by 1.

[0136] Add the current neighbor node as the hth element to the path list.

[0137] It is determined whether the number of nodes h passed is greater than the preset path length to obtain a second determination result.

[0138] If the second determination result is yes, the current data packet is initialized and added to the data packet queue.

[0139] If the second judgment result is no, then the link weight corresponding to the current node c and the current neighbor node is determined to be the current link weight.

[0140] Let the value of the cumulative transmission cost t increase the current link weight.

[0141] It is determined whether the current neighbor node is the current destination node d to obtain a third determination result.

[0142] If the third judgment result is no, the current neighbor node is used as the current node, the current neighbor Q value is used as the minimum Q value, and the process returns to the step of "determining any neighbor node of the current node c as the current neighbor node".

[0143] If the third judgment result is yes, then the path list is determined to be the path corresponding to the current destination node.

[0144] The cumulative transmission cost t is determined to be the cumulative transmission cost of the path corresponding to the current destination node.

[0145] Update the current destination node and return to step "determine that the first data packet in the data packet queue is the current data packet" until all computing power center nodes are traversed to obtain the paths corresponding to all computing power center nodes.

[0146] Step 105 includes:

[0147] The task category of the current pending task is extracted from the computing power identifier of the current pending task.

[0148] When the current task to be processed is a transmission task, the path corresponding to the minimum cumulative transmission cost is selected from the paths corresponding to all computing power center nodes as the optimal path.

[0149] When the current task to be processed is a computing task, the CPU demand and the GPU demand are extracted from the computing power identifier.

[0150] Get the load balancing factor. The load balancing factor is used to describe the remaining resources of all computing centers at the current moment.

[0151] Determine the path corresponding to any computing power center node as the current path.

[0152] According to the objective function, load balancing factor, CPU demand, GPU demand and the accumulated transmission cost of the current path, the objective function value corresponding to the current path is determined.

[0153] Update the current path and return to step "determine the objective function value corresponding to the current path based on the objective function, load balancing factor, CPU demand, GPU demand, and the cumulative transmission cost of the current path" until all computing power center nodes are traversed to obtain the objective function value of the path corresponding to each computing power center node.

[0154] Determine the path corresponding to the maximum objective function value as the optimal path.

[0155] In addition, the task perception and intelligent scheduling method for converged networks provided by the present application, before obtaining the first task in the queue of pending tasks as the current pending task, further includes:

[0156] Construct the objective function and determine the constraints of the objective function.

[0157] The objective function is:

[0158]

[0159] Among them, I is the objective function, λ j (i) represents the jth user task TU j Whether it is assigned to the i-th computing center CC i , when λ j (i) = 1 indicates user task TU j Assigned to the computing center CC i , when λ j (i) = 0 indicates user task TU j Not assigned to the computing center CC iW represents the number of user tasks, N represents the number of computing centers, and α represents the service delay weight. Represents task TU j In the computing center CC i The service delay on the network, β represents the transmission delay weight, Represents task TU j The complete data is transmitted to the computing center CC i The cumulative transmission delay of , γ represents the weight of the load balancing factor, and dr represents the load balancing factor.

[0160] The constraints are:

[0161]

[0162] Among them, c j Represents the jth user task TU j The CPU demand, C i Represents the i-th computing center CC i The remaining CPU resources, g j Denotes the jth user task TU j GPU demand, G i Represents the i-th computing center CC i The remaining GPU resources, s j Denotes the jth user task TU j The storage resource requirement, S i Represents the i-th computing center CC i remaining storage resources.

[0163] The present application also provides an application scenario, which applies the above-mentioned task perception and intelligent scheduling method for converged networks. Specifically: the task perception and intelligent scheduling method for converged networks provided in this embodiment can be applied in a communication network scheduling scenario.

[0164] In an exemplary embodiment, a task perception and intelligent scheduling system for converged networks is provided, including a task parsing module, a routing decision module and a resource adaptation module. The system works as follows: Figure 3 shown.

[0165] The task parsing module is used to automatically parse the tasks initiated by the user and add computing power identification, including:

[0166] The deep learning model analyzes user tasks, specifically:

[0167] In the task parsing module, the tasks submitted by users will carry a series of key attribute information describing the tasks, such as task type, file storage size, bandwidth requirements, and task priority. Among them, the task type is an important dimension that directly determines the characteristics of the task's demand for different computing resources. For example, the cloud code compilation task, a subtask of the computing task, usually requires a large amount of CPU resources, while the demand for GPU resources is relatively small. In the transmission task, large file downloads mainly consume storage and network bandwidth resources, and the demand for computing resources is relatively low. In order to accurately describe the resource requirements of different types of tasks, a task component table is constructed. By analyzing and mining a large amount of real task data, the computing tasks are subdivided into 10 typical subtasks, and the resource consumption characteristics of each subtask are quantitatively described. Then, after the user task is transmitted to the computing power parsing route through various access methods, the computing power parsing route can estimate its demand for CPU and GPU resources based on the tasks uploaded by the user, and realize more refined resource allocation. For transmission tasks, it is usually only necessary to occupy storage resources and allocate a small amount of CPU resources for operations such as copying and pasting data.

[0168] In addition to the task type and file storage size, the user's bandwidth requirements and task priority are also important factors affecting resource allocation. For tasks with higher bandwidth requirements, it is necessary to give priority to computing nodes with better network conditions during scheduling to ensure the efficiency and quality of data transmission. At the same time, for tasks with higher priorities, more computing resources, such as the number of CPU and GPU cores or frequency, need to be given to shorten the execution time of the task and improve the user experience.

[0169] Add computing power identification information to user tasks, specifically:

[0170] The task component table can dynamically perceive and predict the needs of user tasks for different computing resources, and then encapsulate these demand information into a computing power identifier CID. CID, as the "identity card" of the task, carries the resource demand information of the task, providing an important basis for the scheduling decision of the computing power network controller. Based on the task attributes and resource requirements described by CID, the computing power network controller comprehensively considers the real-time status of each computing power node in the computing power network, such as the remaining capacity and load of CPU, GPU, storage and network, as well as the characteristics of heterogeneous computing power resources. Through efficient scheduling algorithms and strategies, the user task is dynamically allocated to the most suitable execution computing power center, realizing the on-demand allocation and global optimization of computing power resources. Through the design of CID, this module can be seamlessly connected with other functional modules of the computing power network, providing a basis for the normal operation of the next routing decision and resource adaptation module, and then completely realizing intelligent resource management and optimization, providing a full range of computing power service guarantees. When a user initiates a task request, the task attribute information is encoded in the above format, encapsulated in the Options field of the IP datagram header, and transmitted to the computing power resolution routing along with the message. After receiving the message, the computing power resolution router extracts the computing power identifier, decodes the attribute parameters of each task, and sends them to the task resolution module to obtain the demand estimate for various heterogeneous resources and generate the computing power identifier for the computing power network controller to make task scheduling and resource allocation decisions.

[0171] The encoding scheme for each attribute field is as follows:

[0172] (1) Task type (TaskType): It is represented by 1 bit, 0 represents a transmission task, and 1 represents a computing task.

[0173] (2) CPU demand: This is represented by 16 bits in units of million cycles. The maximum demand that can be represented is 65,535 million cycles.

[0174] (3) GPU Demand: This is represented by 16 bits in units of million cycles, and can represent a maximum demand of 65,535 million cycles.

[0175] (4) Storage Demand: This value is represented by 32 bits in MB and can represent a maximum storage demand of 4 TB.

[0176] (5) Bandwidth Demand: This field is represented by 13 bits in Mbps, and can represent a maximum bandwidth demand of 8192 Mbps.

[0177] (6) Task Priority: It is represented by 2 bits, from 1 to 3, representing the priority from low to high. The larger the value, the higher the priority. 0 is not enabled.

[0178] When a user initiates a task request, the task attribute information is encoded in the above format, encapsulated into the Options field of the IP datagram header, and transmitted to the computing power resolution router along with the message. After receiving the message, the computing power resolution router extracts the computing power identifier, decodes the attribute parameters of each task, and sends them to the task resolution module to obtain the demand estimate for various heterogeneous resources and generate the computing power identifier for the computing power network controller to make task scheduling and resource allocation decisions.

[0179] The routing decision module is used to implement routing transmission path planning for computing tasks and transmission tasks, including:

[0180] Training of the path decision model, specifically:

[0181] In the computing network integration scenario, the core network is an important device in the process of connecting computing power resolution routing and computing power center. The core network can be abstracted as a weighted directed graph G(R,E) consisting of multiple routing nodes and links, where R = {r1,r2,...,r N} represents the set of all routing nodes in the core network, and each node r represents a routing node. represents the set of links between nodes, and each link e(u,v) represents a direct transmission path from node u to node v. In order to describe the transmission characteristics of different links, a weight w(u,v) is introduced for each link to represent the transmission cost of the path, such as delay, cost, etc., which is set to the waiting time in this application.

[0182] In the computing network, a task data packet p can be represented as a six-tuple:

[0183] p={s,d,c,t,h,f}

[0184] Among them, s and d represent the source node and destination node of the data packet respectively, c represents the current node, t represents the cumulative waiting time of the data packet, h represents the number of hops of the data packet (the number of nodes passed), and f represents whether the current data packet has reached the current destination node d. A data packet starts from the source node, passes through multiple intermediate nodes in sequence according to the routing decision, and finally reaches the destination node, completing end-to-end transmission.

[0185] The routing planning algorithm designed in this application is based on a reinforcement learning algorithm, which continuously learns and optimizes the transmission path of data packets to minimize end-to-end delay. The routing planning algorithm maintains a Q value table for each routing node r, which is used to evaluate the long-term cumulative cost of all possible paths from the node to the destination node. The Q value table is a two-dimensional matrix, in which the rows represent the nodes, the columns represent the next-hop nodes, and the element Q(u,v) represents the action value of selecting node v as the next hop when the data packet p is at node u. During the data packet transmission process, the routing decision module selects the next-hop node according to the ε-greedy strategy: randomly select an adjacent node with probability ε, and select the adjacent node with the smallest Q value with probability 1-ε. As learning progresses, ε gradually decreases, and the decision tends to be stable and optimal.

[0186] The main process of routing simulation is:

[0187] (1) At the beginning of each time step, new data packets of this time step are first generated and added to the data packet queue of the corresponding source node. The newly generated data packets are introduced into the network to participate in the routing decision and transmission process.

[0188] (2) Use a loop to traverse the packet queue of each node. Take out the packet at the head of each non-empty queue and process it. Select the next hop node of the packet based on the current Q value table and the greedy rate. Generate a random number between 0 and 1. If the random number is greater than the greedy rate, it means random exploration, that is, randomly select one from all available next hop nodes of the current node as the next hop. Otherwise, select the next hop node with the smallest Q value, that is, according to the current Q value table, select the path with the smallest estimated cost. This selection strategy takes into account both exploration and utilization. When the greedy rate is high, it is more inclined to select the known optimal path. When the greedy rate is low, it performs more random exploration to discover new possible paths.

[0189] (3) After selecting the next hop node, the Q value table is updated according to the Q learning algorithm. First, the immediate reward is calculated. Here, negative waiting time is used as the immediate reward to encourage the data packet to reach the destination node as soon as possible. Then, according to the Q learning update formula, the Q value of the next hop node from the current node to the destination node is updated. Through this step, the Q value table is continuously adjusted and optimized so that it can more accurately estimate the long-term cost of different paths.

[0190] (4) Determine whether the next hop node of the data packet is the destination node. If so, mark the data packet as arrived and update relevant statistical information, such as total delay. If not, add the data packet to the queue of the next hop node and update relevant attributes of the data packet, such as the current node, waiting time, number of hops, etc.

[0191] (5) Introduce a packet replication mechanism, which is a technology for handling long path packets. Its purpose is to prevent packets from being transmitted and circulated indefinitely in the network, which can lead to network congestion and performance degradation. This mechanism sets a threshold for path length. During the routing decision process at each time step, the path length of each packet is checked, that is, the number of nodes the packet passes through. If the path length of a packet exceeds the preset threshold, it is considered that the packet may be trapped in a loop routing or an overly long path problem.

[0192] When the path length of a data packet exceeds the threshold, a new data packet is created and added to the data packet list. This new data packet has the same properties as the original data packet, such as source node, destination node, waiting time, etc. The difference is that this data packet resets the path information of the new data packet and initializes its path to only include the source node, allowing it to re-explore a shorter or better path.

[0193] Use the trained path decision model to select the appropriate path for the user task. Specifically:

[0194] In the trained network topology, the optimal routing path from the user task generation node to the destination computing center node is found according to the Q value table obtained by the routing decision algorithm, and the path and transmission delay are returned. This transmission delay will be sent to the computing network controller for resource scheduling decisions. The input of the routing planning function is the user task generation node and the destination computing center node. The return value is a list representing the optimal routing path from the source node to the destination node, and the path contains all the nodes passed along the way. The main part of the routing planning function is a loop that continuously searches for and returns the optimal path until the path from the source node to the destination node is found.

[0195] At the beginning of each loop, the starting node is set as the source node, the ending node is set as the destination node, the current node is initialized as the starting node, and the path information is a list used to store the found routing path, which initially only contains the starting node.

[0196] Next is a loop, the condition of the loop is that the next hop node is not the destination node. In each loop, the currently found minimum Q value is first initialized to positive infinity, ensuring that the Q value of the first neighbor node is set to the initial value of the minimum Q value in the first iteration. Initialize the next hop node to empty, traverse all neighbor nodes of the current node, and check whether the Q value from the current node through the neighbor node to the destination node is less than the current minimum Q value. If the above conditions are met, update the minimum Q value to the Q value of the neighbor node, and update the next hop node to the neighbor node. After traversing all neighbor nodes, the next hop node is obtained and added to the path, and the transmission delay of this hop is recorded. Update the current node to the next hop node and continue to find the path for the next step.

[0197] Repeat the above steps until the current node is equal to the destination node, indicating that a complete path from the source node to the destination node has been found. The optimal path and the transmission delay of the entire path are returned as the return value of the routing planning function.

[0198] The resource adaptation module is used for the collaborative processing of computing tasks and transmission tasks and the dynamic scheduling of computing center resources in the computing-network integration scenario, including:

[0199] Design a training environment for the fusion network, specifically:

[0200] Consider a computing network composed of multiple heterogeneous computing centers, each of which has different computing resource configurations (such as CPU resources, GPU resources, and storage resources). Suppose the computing network consists of a set of computing centers CC = {CC1, CC2, ..., CC N}. For any computing power center CC i ,i∈1,2,...,N, the total amount of resources is R i It can be represented as a triple:

[0201] R i =(C i ,G i ,S i )

[0202] Among them, C i ,G i ,S i Respectively represent the computing center CC i The total amount of CPU resources, GPU resources, and storage resources.

[0203] The jth user task TU obtained by the task parsing module j , its resource demand It can be obtained through the carried computing power identifier and expressed as a four-tuple:

[0204]

[0205] Among them, c j ,g j ,s j ,p j Represents user tasks TU j The amount of CPU, GPU, and storage resources required, as well as task priority.

[0206] The computing network resource scheduling problem can be defined as: given a set of user tasks {TU i}, the optimal computing center is selected for each task, so that the service delay of the entire computing network is minimized, the load is as balanced as possible, and the computing resource requirements of the task are met.

[0207] Let the Boolean variable λ j (i) represents the user task TU j Whether it is assigned to the computing center CC i , when λ j (i) = 1 indicates user task TU j Assigned to the computing center CC i , if it is distributed to other computing centers, then λ j (i) = 0, then the optimization objective can be expressed as:

[0208]

[0209] in, Represents task TU j In the computing center CC i The service delay on Represents task TU j The complete data is transmitted to the computing center CC i , dr represents the load balancing factor between computing centers in the computing network, and α, β, and γ are all weight coefficients.

[0210] Design a transmission scheduling algorithm, specifically:

[0211] In order to solve the above optimization problem, a dynamic computing network fusion resource scheduling algorithm based on D3QN is proposed. The algorithm includes:

[0212] (1) State space S

[0213] Define states i For the computing center CC i The remaining resources, s i =(C i ,G i ,S i ), where C i ,Gi ,S i , respectively represent the computing center CC i The current remaining CPU resources, GPU resources, and storage resources of the entire system are defined as:

[0214] S=(s1,s2,...,s N )

[0215] (2) Action Space A

[0216] Define action a i To select and assign user tasks to the computing center CC i , then the action space A can be expressed as:

[0217] A=(a1,a2,...,a N )

[0218] (3) Reward function R

[0219] Through the above analysis, the reward of the transmission scheduling algorithm consists of three parts: the inverse of the service delay of executing the task after being assigned to the designated computing center, the inverse of the transmission delay of the user task uploaded to the computing center, and the load balance of the computing center (the closer the resource utilization rate of each computing center, the larger the value).

[0220] Then the reward function R can be defined as:

[0221]

[0222] The flow of the transmission scheduling algorithm is as follows:

[0223] (1) Construct the configuration information of M computing centers, including computing center ID, CPU speed, GPU speed, link bandwidth and other attributes. At the same time, generate N user tasks, including computing tasks and transmission tasks.

[0224] (2) Execute a scheduling decision in the simulation environment, update the remaining resources of the computing network, and return the reward after execution and the flag of whether it is finished. Specifically, the input parameters are the current task, task number and scheduling action (i.e. the selected computing center). First, select the computing center according to the scheduling action, use the user task priority and the characteristics of the computing center to calculate the required amount of resources, update the CPU, GPU resources and storage resources of the computing center, and simulate the resource occupancy of the task. The computing power demand of the user task is millions of cycles Mcycle, and the computing power resources allocated by the computing center are processor frequency GHz. The former is divided by the latter to obtain the service delay. The final service delay is the maximum value of the CPU service delay and the GPU service delay. When allocating resources to user tasks, the user task priority is divided into three levels (1, 2, 3), and the priority corresponds to the multiple of the basic allocated computing power resources. Each computing center can allocate different CPU and GPU basic computing power resources. Secondly, calculate the reward value generated by this scheduling action. If this scheduling action makes the remaining resources of the computing power center negative, it means that the resource constraint is violated, and a large negative reward value is returned, and the round end flag done is set to True, indicating that this round of simulation is over. Otherwise, calculate the reward value according to the reward function. The reward function combines the execution time of the user task in the computing power center, the transmission delay of the user task from the starting point to the specified computing power center, and the load balancing of the computing power center to evaluate the scheduling decision. Finally, return the reward value and the round end flag done.

[0225] (3) Calculate the reward value for load balancing between computing centers to guide the optimization direction of the scheduling strategy. This part calculates the current CPU and GPU resource utilization of each computing center, that is, the ratio of the remaining resources to the initial resources, and then calculates the mean and variance of the CPU and GPU utilization respectively, and normalizes the variance. Finally, the average value of 1 minus the normalized CPU and GPU variance is taken as the final reward value. At the beginning of scheduling, the utilization of each computing center is relatively small, so it is set that when the sum of the CPU or GPU utilization exceeds a certain threshold, the reward for the load balancing part is calculated. This reward design can make the scheduling strategy of the transmission scheduling algorithm more inclined to optimize the load balancing of the computing center.

[0226] This embodiment provides a task perception and intelligent scheduling system for converged networks, and the workflow is as follows:

[0227] S1: Establish system architecture in converged network scenario.

[0228] S2: The task parsing module is used to automatically parse the tasks initiated by the user and add computing power identification.

[0229] S3: The routing decision module is used to implement routing transmission path planning for computing tasks and transmission tasks.

[0230] S4: The resource adaptation module is used for the collaborative processing of computing tasks and transmission tasks and the dynamic scheduling of computing center resources in the computing-network integration scenario.

[0231] Through the above steps S1 to S4, an embodiment of the present invention provides a task perception and intelligent scheduling method for a converged network, which can realize intelligent perception of task types and characteristic attributes, network link status, and computing center resource occupancy in a converged network scenario, and combine this information to generate optimal task scheduling and resource allocation decisions.

[0232] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a task perception and intelligent scheduling method for a converged network is implemented.

[0233] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0234] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0235] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0236] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0237] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0238] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0239] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0240] The present application uses specific examples to illustrate the principles and implementation methods of the present application. The above examples are only used to help understand the method and core ideas of the present application. At the same time, for those skilled in the art, according to the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present application.

Claims

1. A task perception and intelligent scheduling method for converged networks, characterized in that: The task perception and intelligent scheduling method for a fusion network is applied to a fusion network, and the fusion network includes: Computing power analysis routing, computing power network controller, core network, multiple user devices and multiple computing power centers; The computing power network controller and the plurality of user devices are connected to the computing power resolution routing; the computing power resolution routing is also connected to the plurality of computing power centers through the core network; The user device is used to publish tasks; The computing power analysis routing is used to analyze the received task when receiving a task sent by any user device, and determine the computing power identifier according to the analysis result; The computing power network controller is used to determine the target computing power center and the optimal path of the core network based on the computing power identifier, the core network link status and the computing power of each computing power center pair; The computing power analysis routing is used to transmit the task to the target computing power center according to the optimal path; The task perception and intelligent scheduling method for the fusion network includes: Constructing a converged network architecture; the converged network architecture includes: a plurality of nodes, a plurality of links, and a weight of each link; any of the nodes is a user device node, a routing node, or a computing center node; the link has a user device node or a routing node as a link start point, and a routing node or a computing center node as a link end point; the link weight represents the transmission cost of transmitting data from the link start point to the link end point of the corresponding link; Get the first task in the pending task queue as the current pending task; Extract the user device node of the current task to be processed from the computing power identifier of the current task to be processed as the source node; Based on the fusion network architecture, a path search algorithm is used to determine the corresponding path when the source node is used as the starting point and each computing power center node is used as the destination node; According to the computing power identifier of the current task to be processed, select the optimal path from the paths corresponding to all computing power center nodes; Determine the computing power center node corresponding to the optimal path as the optimal destination node; Computing power analysis routing sends the current pending task from the source node to the optimal destination node according to the optimal path; The optimal destination node generates a processing result after processing the current pending task; The computing power analysis routing returns the processing results from the optimal destination node to the source node; Delete the current pending task from the pending task queue and return to step "Get the first task in the pending task queue as the current pending task".

2. The task perception and intelligent scheduling method for converged networks according to claim 1 is characterized in that: The task perception and intelligent scheduling method for the fusion network also includes: When the computing power resolution routing receives a task request sent by any user device, it updates the queue of pending tasks according to the task request.

3. The task perception and intelligent scheduling method for converged networks according to claim 2 is characterized in that: When the computing power resolution router receives a task request from any user device, it updates the pending task queue according to the task request, including: When the computing power resolution routing receives a task request sent by any user device, it determines that the requested task is the current task to be sorted; Extracting the task category of the current task to be sorted from the task attributes of the current task to be sorted; the task attributes include task category, storage resource amount, bandwidth requirement and task priority; the task category is a transmission task or a computing task; When the current task to be sorted is a computing task, determining the subtask category of the current task to be sorted to be the current subtask category; Query the CPU ratio and GPU ratio corresponding to the current subtask category in the task component table; the computing power analysis route stores the task component table; the task component table is used to express the CPU ratio and GPU ratio corresponding to different computing task subtask categories; the sum of the CPU ratio and GPU ratio corresponding to the same computing task subtask category is equal to 1; Determine the product of the storage resource amount of the current task to be sorted and the CPU ratio corresponding to the current subtask category as the CPU demand of the current task to be sorted; Determine the product of the storage resource amount of the current task to be sorted and the GPU ratio corresponding to the current subtask category as the GPU demand of the current task to be sorted; Encoding the CPU requirement, the GPU requirement, and the task attribute to obtain a computing power identifier; When the current task to be sorted is a transmission task, encoding the task attribute to obtain a computing power identifier; Add the computing power identifier to the message header of the current task to be sorted; Based on the priority of the current task to be sorted and the priorities of all pending tasks in the pending task queue, the current task to be sorted is inserted into the pending task queue to complete the update of the pending task queue.

4. The task perception and intelligent scheduling method for converged networks according to claim 3 is characterized in that: Based on the fusion network architecture, a path search algorithm is used to determine the corresponding path when the source node is used as the starting point and each computing center node is used as the destination node, including: At the source node, the current pending tasks are divided into data packet queues; Determine any computing power center node as the current destination node d; Determine that the first data packet in the data packet queue is the current data packet; Add the source node s as the first element to the path list; Let source node s be current node c; Let the cumulative transmission cost of the current data packet to reach the current node be t = 0; Let the number of nodes passed h = 1; Initialize the current data packet p = {s, d, c, t, h, f}; f indicates whether the current data packet has reached the current destination node d; Initialize the minimum Q value to positive infinity; Determine any neighbor node of the current node c as the current neighbor node; the neighbor node of the current node c is the link end point corresponding to the link with the current node c as the link start point in the converged network architecture; Determine the Q value from the current node c through the current neighbor node to the current destination node d as the current neighbor Q value; Determine whether the current neighbor Q value is less than the minimum Q value, and obtain a first determination result; If the first judgment result is no, the current neighbor node is updated, and the process returns to step "determining the Q value of reaching the current destination node d from the current node c through the current neighbor node is the Q value corresponding to the current neighbor node"; If the first judgment result is yes, then the value of the number of nodes passed h is increased by 1; Add the current neighbor node as the hth element in the path list; Determine whether the number of nodes h passed is greater than the preset path length, and obtain a second determination result; If the second judgment result is yes, then initialize the current data packet and add the current data packet to the data packet queue; If the second judgment result is no, then determine that the link weight corresponding to the current node c and the current neighbor node is the current link weight; Let the value of the cumulative transmission cost t increase the current link weight; Determine whether the current neighbor node is the current destination node d, and obtain a third determination result; If the third judgment result is no, the current neighbor node is used as the current node, the current neighbor Q value is used as the minimum Q value, and the process returns to step "determining any neighbor node of the current node c as the current neighbor node"; If the third judgment result is yes, then determining that the path list is a path corresponding to the current destination node; The cumulative transmission cost t is determined to be the cumulative transmission cost of the path corresponding to the current destination node.

5. The task perception and intelligent scheduling method for converged networks according to claim 4 is characterized in that: After determining that the cumulative transmission cost t is the cumulative transmission cost of the path corresponding to the current destination node, the method further includes: Update the current destination node and return to step "determine that the first data packet in the data packet queue is the current data packet" until all computing power center nodes are traversed to obtain the paths corresponding to all computing power center nodes.

6. The task perception and intelligent scheduling method for converged networks according to claim 4 is characterized in that: According to the computing power identifier of the current task to be processed, the optimal path is selected from the paths corresponding to all computing power center nodes, including: Extract the task category of the current pending task from the computing power identifier of the current pending task; When the current task to be processed is a transmission task, the path corresponding to the minimum cumulative transmission cost is selected from the paths corresponding to all computing power center nodes as the optimal path.

7. The task perception and intelligent scheduling method for converged networks according to claim 4 is characterized in that: According to the computing power identifier of the current task to be processed, the optimal path is selected from the paths corresponding to all computing power center nodes, which also includes: When the current task to be processed is a computing task, the CPU demand and GPU demand are extracted from the computing power identifier; Obtain a load balancing factor; the load balancing factor is used to describe the remaining resources of all computing power centers at the current moment; Determine the path corresponding to any computing power center node as the current path; Determine the objective function value corresponding to the current path according to the objective function, the load balancing factor, the CPU demand, the GPU demand, and the accumulated transmission cost of the current path; Update the current path and return to step "determine the objective function value corresponding to the current path according to the objective function, the load balancing factor, the CPU demand, the GPU demand and the accumulated transmission cost of the current path" until all computing power center nodes are traversed to obtain the objective function value of the path corresponding to each computing power center node; Determine the path corresponding to the maximum objective function value as the optimal path.

8. The task perception and intelligent scheduling method for converged networks according to claim 7 is characterized in that: Before obtaining the first task in the pending task queue as the current pending task, it also includes: Construct the objective function and determine the constraints of the objective function.

9. The task perception and intelligent scheduling method for converged networks according to claim 8, characterized in that: The objective function is: Among them, I is the objective function, λ j (i) represents the jth user task TU j Whether it is assigned to the i-th computing center CC i , when λ j (i) = 1 indicates user task TU j Assigned to the computing center CC i , when λ j (i) = 0 indicates user task TU j Not assigned to the computing center CC i ; W represents the number of user tasks, N represents the number of computing centers, α represents the service delay weight, Represents task TU j In the computing center CC i The service delay on the network, β represents the transmission delay weight, Represents task TU j The complete data is transmitted to the computing center CC i The cumulative transmission delay of , γ represents the weight of the load balancing factor, and dr represents the load balancing factor.

10. The task perception and intelligent scheduling method for converged networks according to claim 9, characterized in that: The constraints are: Among them, c j Represents the jth user task TU j The CPU demand, C i Represents the i-th computing center CC i The remaining CPU resources, g j Denotes the jth user task TU j GPU demand, G i Represents the i-th computing center CC i The remaining GPU resources, s j Denotes the jth user task TU j The storage resource requirement, S i Represents the i-th computing center CC i remaining storage resources.

Images