CN116909717A - Task scheduling method - Google Patents

Task scheduling method Download PDF

Info

Publication number
CN116909717A
CN116909717A CN202311169958.9A CN202311169958A CN116909717A CN 116909717 A CN116909717 A CN 116909717A CN 202311169958 A CN202311169958 A CN 202311169958A CN 116909717 A CN116909717 A CN 116909717A
Authority
CN
China
Prior art keywords
task
server
servers
time
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202311169958.9A
Other languages
Chinese (zh)
Other versions
CN116909717B (en
Inventor
郭小林
寇剑
孙贞
张智勇
陈志刚
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guoneng Beijing Business Network Co ltd
Original Assignee
Guoneng Beijing Business Network Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guoneng Beijing Business Network Co ltd filed Critical Guoneng Beijing Business Network Co ltd
Priority to CN202311169958.9A priority Critical patent/CN116909717B/en
Publication of CN116909717A publication Critical patent/CN116909717A/en
Application granted granted Critical
Publication of CN116909717B publication Critical patent/CN116909717B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/48—Program initiating; Program switching, e.g. by interrupt
    • G06F9/4806—Task transfer initiation or dispatching
    • G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
    • G06F9/4881—Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2209/00—Indexing scheme relating to G06F9/00
    • G06F2209/48—Indexing scheme relating to G06F9/48
    • G06F2209/484—Precedence
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

The application belongs to the technical field of edge calculation, and particularly relates to a task scheduling method, which comprises the steps of constructing a graph neural network based on the relation among servers; constructing a task queue, and placing a new task generated in real time in the task queue; when a task is scheduled, updating the graph neural network based on the state information of each current server; based on the updated graph neural network, the decision model makes a decision, and the current task is distributed to the server; and simultaneously, the server receiving the task feeds back the current state information to the server generating the task. The automatic allocation and scheduling of tasks in the system are realized by combining the graph neural network of the server state information with the DQN algorithm, and the processing capacity, the processing time and the distance factors among the servers are considered in the graph neural network, so that the problem of transmission delay among the servers which is difficult to consider in the prior art can be effectively avoided.

Description

Task scheduling method
Technical Field
The application belongs to the technical field of edge calculation, and particularly relates to a task scheduling method.
Background
In an edge computing system, new tasks are continuously generated, and the tasks are distributed to a certain server in the system; but the deployment locations of different servers are different, the computing power is different, and the server nodes may move, so that the transmission time from the node generating the task to the different server nodes is different. For example, in a new task scenario where unmanned aerial vehicles participate in network processing and are continuously generated, each unmanned aerial vehicle carries a server for processing tasks, the computing power of different servers is different, the unmanned aerial vehicle may move, and the server carried by the unmanned aerial vehicle may temporarily close or reduce a part of computing resources due to faults.
Aiming at the situation, the traditional task scheduling method can meet the task allocation, but the problem that the transmission delay is overlarge is often caused when the task is allocated due to the fact that the transmission delay is not focused due to the distance change generated by the displacement of a server.
Disclosure of Invention
The application aims to provide a task scheduling method for solving the technical problem that the transmission delay is not considered under the condition of dynamic change of the transmission delay due to the position change of a server in an edge computing system so as to influence the efficient processing of a task.
The embodiment of the application is realized by the following technical scheme:
there is provided a task scheduling method, as shown in fig. 1, including:
constructing a graph neural network based on the relation among the servers;
constructing a task queue, and placing a new task generated in real time in the task queue;
when a task is scheduled, updating the graph neural network based on the state information of each current server;
based on the updated graph neural network, the decision model makes a decision, and the current task is distributed to the server; and simultaneously, the server receiving the task feeds back the current state information to the server generating the task.
Further, the building of the graph neural network based on the relationship between the servers specifically includes:
constructing a graph neural network based on the distance between every two servers in a range of transmissible information, wherein the edges of the graph neural network represent the transmission time between the servers connected with the graph neural network;
the nodes in the graph neural network are servers, and the attributes of the nodes comprise server numbers, distances between the current server and other servers, whether the current server can directly communicate with the other servers, the computing power of the servers, the ending time of the servers for processing the current task and the connectivity of the servers.
Further, the transmission time is proportional to the distance between the two servers and inversely proportional to the bandwidth, i.e.:
t=d/k
wherein t is transmission time, d is distance between two servers, and k is bandwidth;
the connectivity of the server is obtained by the following formula:
wherein connection_delay (j) represents connectivity of a jth server in the graph neural network, c represents the number of servers of which the jth server can directly communicate and transmit information, and t ij Representing the transmission time between the jth server and the ith server that can directly communicate the transmission information.
Further, the attributes of the tasks in the task queue comprise task numbers, occupied time, occupied space and priorities; forming a directed graph based on the attribute of the task and the partial order relation between the tasks in the task queue; the inflow and outflow of the directed graph represents the precedence relationship between tasks.
Further, the state information in the neural network is updated to be the attribute information of the server based on the state information of each current server.
Further, the decision model comprises a decision model of a learning network and a target network, the learning network has the same structure as the target network, the learning network is used for learning from new data and synchronizing the learned parameter values to the target network, and the target network is used for providing calculation of the value of the state.
Further, in the decision model, the state is a tuple consisting of the graph neural network and a task queue; the action is a tuple formed based on the task number and the server number; the prize value is: k1×occupied time-k2×occupied space-k3×priority×occupied time; wherein k1, k2 and k3 are preset coefficients, the unit of occupied time is seconds, and the unit of occupied space is megabytes.
Further, the decision model further comprises a state transfer function, the state transfer function removes the task from the task queue after executing the scheduling policy on the first task of the current task queue, and updates the ending time of the current task, which is allocated to the server of the task, to the ending time of the next task after processing the first task based on the occupation time of the task.
Further, the loss value of the decision model is the difference between the first value of the current state and the second value of the current state; the first value of the current state is obtained by the learning network according to the current state calculation, and the second value of the current state is the sum of the rewarding value and the value of the next state calculated by the target network.
Further, the occupied time includes a transmission time, a waiting time and a processing time, the transmission time is the time for a server generating a task to transmit the task to another server processing the task, the waiting time is the time for the server to process all tasks before the task is processed after the task is received, and the processing time is the time for the server to process the task.
The technical scheme of the embodiment of the application has at least the following advantages and beneficial effects:
the application constructs the graph neural network based on the state information of each server in the edge computing system, updates the state information of each server in real time based on the graph neural network, and simultaneously combines the graph neural network of the state information of each server with the DQN algorithm to realize automatic allocation and scheduling of tasks in the system, and because the processing capacity, the processing time and the distance factors among the servers are considered in the graph neural network, the problem of transmission delay among the servers which is difficult to consider in the prior art can be effectively avoided, and meanwhile, the system can automatically process the tasks with high priority by introducing the priority; and further, reasonable and efficient task scheduling is realized.
Drawings
FIG. 1 is a flow chart of a task scheduling method of the present application;
FIG. 2 is a directed graph of task attributes that can be constructed in accordance with the present application;
fig. 3 is a network configuration diagram of the DQN agent in the present application.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the present application, but not all embodiments of the present application. The components of the embodiments of the present application generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
In the edge computing system, the servers can continuously generate new tasks, and the generated tasks can be distributed to other servers in the system for processing, so that the tasks need to be reasonably distributed due to different positions and different computing capacities of the servers. However, in the existing allocation method, the situation that the server is a mobile device is rarely considered, because the deployment position of the server is dynamic, two servers cannot directly communicate or the transmission time is prolonged due to distance reasons in different time periods, and therefore, the efficiency of task allocation processing becomes low due to the fact that the difference caused by the position change of the server is difficult to be considered by using the existing task allocation algorithm.
The task scheduling method provided by the application comprises the following steps:
constructing a graph neural network based on the relation among the servers; the method comprises the following steps:
and constructing a graph neural network based on the distance between every two servers in a range of transmissible information, wherein the edge of the graph neural network is the transmission time between the servers connected by the edge.
The nodes in the graph neural network are servers, and the attributes of the nodes comprise server numbers, distances between the current server and other servers, whether the current server can directly communicate with the other servers, the computing power of the servers, the ending time of the servers for processing the current task and the connectivity of the servers.
The transmission time is proportional to the distance between the two servers and inversely proportional to the bandwidth, i.e.:
t=d/k
wherein t is transmission time, d is distance between two servers, and k is bandwidth;
the connectivity of the server is obtained by the following formula:
wherein connection_delay (j) represents connectivity of a jth server in the graph neural network, c represents the number of servers of which the jth server can directly communicate and transmit information, and t ij Representing the transmission time between the jth server and the ith server that can directly communicate the transmission information.
And constructing a task queue, and placing a new task generated in real time in the task queue.
The attributes of the tasks in the task queue comprise task numbers, occupied time, occupied space and priority; forming a directed graph based on the attribute of the task and the partial order relation between the tasks in the task queue; the inflow and outflow of the directed graph represents the precedence relationship between tasks.
When a task is scheduled, updating the graph neural network based on the state information of each current server; wherein the state information is attribute information of the node.
Based on the updated graph neural network, the decision model makes a decision, and the current task is distributed to the server; and simultaneously, the server receiving the task feeds back the current state information to the server generating the task. It should be noted that, the state information of any one server is updated, and the graph neural network is updated in real time, so as to ensure the smooth progress of task scheduling.
The decision model comprises a decision model of a learning network and a target network, the learning network has the same structure as the target network, the learning network is used for learning from new data and synchronizing the learned parameter values to the target network, and the target network is used for providing calculation of the value of the state.
In the decision model, the state is a tuple consisting of the graph neural network and a task queue; the action is a tuple formed based on the task number and the server number; the prize value is: k1×occupied time-k2×occupied space-k3×priority×occupied time; wherein k1, k2 and k3 are preset coefficients, the unit of occupied time is second, the unit of occupied space is megabyte, and the higher the priority is, the larger the value is.
The decision model also comprises a state transfer function, wherein the state transfer function removes the task from the task queue after the first task of the current task queue executes the scheduling strategy, and updates the ending time of the server for processing the current task, which is allocated to the task, to the ending time of the next task after the first task is processed based on the occupied time of the task.
The loss value of the decision model is the difference between the first value of the current state and the second value of the current state; the first value of the current state is obtained by the learning network according to the current state calculation, and the second value of the current state is the sum of the rewarding value and the value of the next state calculated by the target network. The first value and the second value are both values of finger-action pairs herein for evaluating the degree of merit of taking a particular action in a given state. This cost function may be expressed as a Q value representing the desired jackpot that can be achieved by taking an action in a certain state.
The occupation time includes a transmission time, a waiting time and a processing time, wherein the transmission time is the time for a server generating a task to transmit the task to another server processing the task, the waiting time is the time for the server to process all tasks before the task is processed after the task is received, and the processing time is the time for the server to process the task.
For easy understanding, the following description is further made for the scheme of the present application:
the task set and its previous partial order relationship are constructed as a partial order set p= (V, ≼), where V is the task set, vi ≼ vj represents the processing procedure of task vj, depending on the processing procedure of task vi, i.e. task vi needs to be processed before task vj, and the processing result of task vi needs to be known by the server computing task vj. The partial order ≼ can be represented as a directed acyclic graph.
The tasks that are continuously generated in the system form a task list with unlimited length, namely v1, v2, v3 and …, and the attributes of the tasks include time occupied by processing the tasks, occupied space, self priority and task numbers, wherein the task numbers are unique. In addition, the relationship between a certain task and other tasks in the partial order set needs to be considered.
As shown in fig. 2, the upper digit of each node represents a task number, and the lower digit represents an attribute of a task; it should be noted that there may be more than one attribute, and the figure is only schematic, so that only the transmission time attribute is displayed. The inflow and outflow in the directed graph represent precedence constraints between tasks.
With respect to servers, the application is used in an edge computing scenario where the number of servers is numerous, each server being capable of both generating and processing tasks. The task scheduling method decides how to schedule the task: the scheduling order and the server to be allocated are determined. When a task is allocated it leaves the waiting slot and the first task stored in the backlog queue is placed in the waiting slot that was just left. It is assumed that each task is handled on only a single server and takes full advantage of the computing resources of that server. The expected processing time in each server is known prior to task execution. When a task is dispatched to a server, its response time includes a latency in the latency slot and a server processing time, where the processing time is positively correlated with the space occupied by the task and negatively correlated with the processing capacity of the server.
From the above, it is understood that each time a task is assigned, it is removed from the task queue, and it is understood that the task queue is an ordered list.
In addition, the server also has attributes, which are important in the scheme of the application, and the graph neural network needs to be constructed based on the attributes of the server. The attributes of the server in the present application are defined to include the server number, the distance between the current server and other servers, whether the current server can directly communicate with other servers, the computing power of the server, the ending time of the server processing the current task, and the connectivity of the server. Wherein the value of connectivity is the number value of other servers which can be directly communicated with the current server. In the graph neural network constructed by the server attribute, the servers are nodes, the edge of each node represents two servers connected with the edge and can be directly communicated, and the attribute of the edge is information transmission time between the two servers.
After the graph neural network is constructed, because the system is required to be capable of realizing automatic and efficient task allocation, a decision model is also required to be constructed to execute optimal task scheduling based on the graph neural network. First we define a state space S. The state space S is all possible states, each state is defined as a tuple consisting of two lists, i.e. s= (L, gt), the first element is a list, it is a list containing M tasks, each task has at least 4 attributes, it is understood that the queue contains at most M tasks, for tasks not filled in the queue, the task number and other attributes take 0; i.e. l1= [ v1, v2, v3...vm ]. Each task is represented as a tuple containing all of its attribute data, e.g., l1= [ {1,6, 2,1}, {2, 8, 3, 3} ], task attributes are task number, occupied time, occupied space, and priority in order; the second element is a graph neural network, which contains attribute information of N servers and information of transmission time between them.
An action space a is redefined, which represents all possible actions. Each action is a binary group containing a task number and a server number, e.g., a= {2,1} indicates that the 2 nd task is selected and placed in the 1 st server.
Defining a prize value r, -k1×occupied time-k2×occupied space-k3×priority×occupied time; wherein k1, k2 and k3 are preset coefficients, and default values are 3, 20 and 10 respectively; the occupied time comprises three parts, namely transmission time, waiting time and processing time, wherein the transmission time is the time for a server generating a task to transmit the task to another server processing the task, the waiting time is the time for the server to process all tasks before the task is processed after the task is received, and the processing time is the time for the server to process the task.
And defining a strategy pi, and defining the behavior of the intelligent agent in the S state according to the state S (from the state space S) and the action a (from the action space A). Pi(s) represents: a policy to be taken when the state is s, which causes the state to transition from s to s'.
And defining a state transfer function, adopting a strategy pi(s) when the state is s, enabling the state to be transferred from s to s', and updating the ending time of the corresponding server for processing the current task to the ending time of the next task after the first task is processed when the corresponding task is moved out.
Based on the above definition and setting, when the state (such as location, computing power, etc.) of a certain server is changed, it is sent to the server generating the task and other servers, and the neural network information is updated. It should be noted that the update process is updated in real time.
The main flow of the task scheduling method is as follows:
placing the new task generated in real time in a task queue;
when a certain task is scheduled, the graph neural network information is updated based on the position of each current server, the distance between the current server and other servers, the computing power of the servers, whether the current server can directly communicate with the other servers, and the ending time of the current task processed by the servers. The fact that the current server can directly communicate with other servers means that when the position of the server moves, the distance between the current server and the server generating the task changes, and some servers which can directly transmit information with the server generating the task can not directly transmit information (other servers are needed to transit the information), or the current server and other servers which can not directly transmit information can directly transmit information, so that whether the attribute of the current server and other servers can directly communicate is updated.
The DQN agent makes decisions based on the information provided by the graph neural network, assigning the task to a certain server. If the task generating server and the task assigned server cannot directly transmit information, a path with the shortest transmission time between the two nodes is selected for transmission according to the transmission time between the nodes.
The server receiving the task updates its own state information and feeds back its current state information to the server generating the task.
The goal of the scheduling algorithm is to select a task from a plurality of tasks in the task queue according to the task attribute in the current task queue and the occupation condition (state value) of the server, and then to transmit the task to one server to finish the processing (action value). The process is completed by decision making from state to action with DQN agent: by training a deep learning neural network, a value (S) in a state space (S) is input, probabilities of all actions (a) in an action space (A) are output, and the maximum probability is selected as a decision of the neural network. The actions include two parts: the selected task number and the assigned server serial number.
The network structure formed by the DQN agent is shown in figure 3, and the network structure comprises two parts, wherein the first part is a graph neural network, and the second part is the DQN agent; the input of the DQN agent is divided into two parts, wherein the first part is a task queue, and the second part is a graph neural network data structure constructed by the state information relation among the servers.
The training process of the DQN agent can be as follows:
1. initializing a task queue and a server. New tasks are continually generated. Defining states, actions, prize values, state transfer functions, etc., defining learning network and target network structures, their learning processes, etc.
2. The learning network calculates the action (a) to be taken currently according to the state(s) of the current task, namely, selects a certain task to be handed to a certain server for processing. The parameters of the learning network are then updated based on the loss values calculated by the learning network and the target network.
3. Based on the current action (a), i.e. after a task is handed to a certain server, the current state is updated with a state transfer function (s is updated to s'): and the corresponding task is moved out, and the ending time of the corresponding server for processing the current task is updated to the ending time of the next task after the first task is processed.
And 4, continuously repeating the operations 2 and 3 until the learning network and the target network converge.
And 5, finishing the training process of the agent, and using the training process for task scheduling. And inputting a task list, giving a scheduling result by the agent, and printing a server condition once every 5 times of task scheduling.
Based on the above description of the scheme of the application, the application can be known to construct the graph neural network based on the state information of each server in the edge computing system, and update the state information of each server in real time based on the graph neural network, the graph neural network of the state information of each server simultaneously combines with the DQN algorithm to realize automatic allocation and scheduling of tasks in the system, and because the processing capacity, the processing time and the distance factors among the servers are considered in the graph neural network, the problem of transmission delay among the servers which is difficult to consider in the prior art can be effectively avoided, and meanwhile, because of the introduction of the priority, the system can automatically process the task with high priority preferentially; and further, reasonable and efficient task scheduling is realized.
The above is only a preferred embodiment of the present application, and is not intended to limit the present application, but various modifications and variations can be made to the present application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims (10)

1. A method for task scheduling, comprising:
constructing a graph neural network based on the relation among the servers;
constructing a task queue, and placing a new task generated in real time in the task queue;
when a task is scheduled, updating the graph neural network based on the state information of each current server;
based on the updated graph neural network, the decision model makes a decision, and the current task is distributed to the server; and simultaneously, the server receiving the task feeds back the current state information to the server generating the task.
2. The task scheduling method according to claim 1, wherein the building of the graph neural network based on the relationship between the servers specifically includes:
constructing a graph neural network based on the distance between every two servers in a range of transmissible information, wherein the edges of the graph neural network represent the transmission time between the servers connected with the graph neural network;
the nodes in the graph neural network are servers, and the attributes of the nodes comprise server numbers, distances between the current server and other servers, whether the current server can directly communicate with the other servers, the computing power of the servers, the ending time of the servers for processing the current task and the connectivity of the servers.
3. The task scheduling method according to claim 2, wherein the transmission time is proportional to the distance between two servers and inversely proportional to the bandwidth, namely:
t=d/k
wherein t is transmission time, d is distance between two servers, and k is bandwidth;
the connectivity of the server is obtained by the following formula:
wherein connection_delay (j) represents connectivity of a jth server in the graph neural network, c represents the number of servers of which the jth server can directly communicate and transmit information, and t ij Representing the transmission time between the jth server and the ith server that can directly communicate the transmission information.
4. The task scheduling method according to claim 2, wherein the attributes of the tasks in the task queue include task number, occupied time, occupied space, and priority; forming a directed graph based on the attribute of the task and the partial order relation between the tasks in the task queue; the inflow and outflow of the directed graph represents the precedence relationship between tasks.
5. The task scheduling method according to claim 1, wherein the updating of the state information in the graph neural network based on the state information of each current server is performed as attribute information of the server.
6. The task scheduling method according to claim 4, wherein the decision model includes a decision model of a learning network and a target network, the learning network being identical in structure to the target network, the learning network being configured to learn from new data and synchronize the learned parameter values to the target network, the target network being configured to provide calculation of the value of the state.
7. The task scheduling method according to claim 6, wherein in the decision model, a state is a tuple consisting of the graph neural network and a task queue; the action is a tuple formed based on the task number and the server number; the prize value is: k1×occupied time-k2×occupied space-k3×priority×occupied time; wherein k1, k2 and k3 are preset coefficients, the unit of occupied time is seconds, and the unit of occupied space is megabytes.
8. The task scheduling method according to claim 7, wherein the decision model further includes a state transfer function, the state transfer function removing the task from the task queue after the first task of the current task queue executes the scheduling policy, and updating the end time of processing the current task assigned to the server of the task to the end time of processing the next task after the first task based on the occupied time of the task.
9. The task scheduling method of claim 7, wherein the loss value of the decision model is a difference between a first value of the current state and a second value of the current state; the first value of the current state is obtained by the learning network according to the current state calculation, and the second value of the current state is the sum of the rewarding value and the value of the next state calculated by the target network.
10. The task scheduling method according to claim 4, wherein the occupied time includes a transmission time, a waiting time, and a processing time, the transmission time being a time for a server that generates a task to transmit the task to another server that processes the task, the waiting time being a time taken by the server to process all tasks before the task is processed after the task is received, the processing time being a time for the server to process the task.
CN202311169958.9A 2023-09-12 2023-09-12 Task scheduling method Active CN116909717B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311169958.9A CN116909717B (en) 2023-09-12 2023-09-12 Task scheduling method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311169958.9A CN116909717B (en) 2023-09-12 2023-09-12 Task scheduling method

Publications (2)

Publication Number Publication Date
CN116909717A true CN116909717A (en) 2023-10-20
CN116909717B CN116909717B (en) 2023-12-05

Family

ID=88358742

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311169958.9A Active CN116909717B (en) 2023-09-12 2023-09-12 Task scheduling method

Country Status (1)

Country Link
CN (1) CN116909717B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113238847A (en) * 2021-05-20 2021-08-10 西安电子科技大学 Distribution and scheduling method based on distributed network environment and capable of distributing tasks
US11206221B1 (en) * 2021-06-04 2021-12-21 National University Of Defense Technology Online task dispatching and scheduling system and method thereof
CN114706631A (en) * 2022-04-22 2022-07-05 哈尔滨工业大学 Unloading decision method and system in mobile edge calculation based on deep Q learning
CN115686779A (en) * 2022-10-14 2023-02-03 兰州交通大学 Self-adaptive edge computing task scheduling method based on DQN
WO2023124947A1 (en) * 2021-12-29 2023-07-06 苏州浪潮智能科技有限公司 Task processing method and apparatus, and related device

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113238847A (en) * 2021-05-20 2021-08-10 西安电子科技大学 Distribution and scheduling method based on distributed network environment and capable of distributing tasks
US11206221B1 (en) * 2021-06-04 2021-12-21 National University Of Defense Technology Online task dispatching and scheduling system and method thereof
WO2023124947A1 (en) * 2021-12-29 2023-07-06 苏州浪潮智能科技有限公司 Task processing method and apparatus, and related device
CN114706631A (en) * 2022-04-22 2022-07-05 哈尔滨工业大学 Unloading decision method and system in mobile edge calculation based on deep Q learning
CN115686779A (en) * 2022-10-14 2023-02-03 兰州交通大学 Self-adaptive edge computing task scheduling method based on DQN

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
高寒 等: "移动边缘计算环境中基于能耗优化的深度神经网络计算任务卸载策略", 计算机集成制造系统, no. 06, pages 167 - 175 *

Also Published As

Publication number Publication date
CN116909717B (en) 2023-12-05

Similar Documents

Publication Publication Date Title
CN111756812B (en) Energy consumption perception edge cloud cooperation dynamic unloading scheduling method
CN109669768B (en) Resource allocation and task scheduling method for edge cloud combined architecture
CN113098714B (en) Low-delay network slicing method based on reinforcement learning
CN108076158B (en) Minimum load route selection method and system based on naive Bayes classifier
CN114338504A (en) Micro-service deployment and routing method based on network edge system
CN110784366B (en) Switch migration method based on IMMAC algorithm in SDN
CN116069512B (en) Serverless efficient resource allocation method and system based on reinforcement learning
CN111988787B (en) Task network access and service placement position selection method and system
CN113645273B (en) Internet of vehicles task unloading method based on service priority
CN116451934B (en) Multi-unmanned aerial vehicle edge calculation path optimization and dependent task scheduling optimization method and system
CN114443249A (en) Container cluster resource scheduling method and system based on deep reinforcement learning
CN114281104A (en) Multi-unmanned-aerial-vehicle cooperative regulation and control method based on improved ant colony algorithm
CN110490319B (en) Distributed deep reinforcement learning method based on fusion neural network parameters
Xu et al. Living with artificial intelligence: A paradigm shift toward future network traffic control
CN116893861A (en) Multi-agent cooperative dependency task unloading method based on space-ground cooperative edge calculation
CN114691372A (en) Group intelligent control method of multimedia end edge cloud system
CN114201303A (en) Task unloading optimization method of fixed path AGV in industrial Internet of things environment
CN112213956B (en) Automatic driving simulation task scheduling method, device, equipment and readable medium
CN113747450A (en) Service deployment method and device in mobile network and electronic equipment
CN116909717B (en) Task scheduling method
CN114217944A (en) Dynamic load balancing method for neural network aiming at model parallelism
CN115484205B (en) Deterministic network routing and queue scheduling method and device
CN116367190A (en) Digital twin function virtualization method for 6G mobile network
CN116112488A (en) Fine-grained task unloading and resource allocation method for MEC network
CN112887347B (en) Dynamic migration method and device for edge calculation in industrial internet

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant