A deterministic transmission scheduling system for computing-network integration and its working method

Abstract

The present invention provides a deterministic transmission scheduling system for computing-network integration and its working method. The system consists of five modules: a system coordination module, a service analysis module, a computing-network perception module, a transmission scheduling algorithm module, and an algorithm deployment module. The system coordination module is responsible for calling each module to achieve efficient collaboration between modules. The service analysis module is responsible for acquiring and analyzing computing task information. The computing-network perception module is responsible for sensing and collecting computing-network status. The transmission scheduling algorithm module is responsible for executing the trained deterministic transmission scheduling algorithm and providing a transmission scheduling plan. The algorithm deployment module distributes the transmission scheduling plan to switches and routers in the network so that they are scheduled according to the plan. The system and method of the present invention can integrate communication resources and computing resources to support the scheduling and transmission of deterministic computing tasks.

Description

Technical Field

[0001] The present invention relates to the field of communication network technology, and in particular to a deterministic transmission scheduling system for computing-network integration and a working method thereof. Background Art

[0002] With the rapid development of emerging technologies such as 5G and mobile internet, the demand for computing resources from new services and applications is increasing. Computing resources are shifting towards users, and how to efficiently utilize them has become a focus of attention. Computing-network convergence has emerged in this context; this technology deeply integrates computing and networking, enabling collaborative management and allocation of computing resources to enhance user experience. With the advent of a new round of technological and industrial revolutions, the in-depth integration of artificial intelligence and the new industrial revolution has shaped new data-driven business models, creating an urgent need for deterministic transmission and real-time processing of computing tasks. The network is the primary medium for information transmission and an indispensable component of the new industrial ecosystem.

[0003] However, the current Internet is designed based on IP networks. IP networks utilize a non-connection-oriented technology. This means that at the network transmission level, the data sender only needs to know the receiving end's IP address to forward the data packet to the destination address through routers. The advantage of this design is that it allows for excellent network scalability, and the simple design of the network layer enables the network to operate in a "lightly loaded" mode. However, over time, the types of Internet services have continued to increase, and users' demand for quality of service (QoS) has also increased. This has led to the "best effort" (BE) network transmission mode beginning to reveal some shortcomings, such as latency uncertainty, lack of QoS guarantees, inadequate congestion control, and uneven link capacity allocation. Because the network layer does not guarantee the transmission time of data packets, end-to-end latency can fluctuate significantly, which is unacceptable for applications that require real-time or low-latency communication, such as voice and video communications, online gaming, and remote surgery.

[0004] To address the above challenges, there is an urgent need to design a deterministic computing-network fusion system to support a large number of deterministic computing tasks. Summary of the Invention

[0005] The embodiments of the present invention provide a deterministic transmission scheduling system and a working method thereof for computing-network integration, so as to realize effective transmission scheduling of computing tasks in the network.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0007] According to one aspect of the present invention, a deterministic transmission scheduling system for computing-network integration is provided, comprising: a system coordination module, a service analysis module, a computing-network perception module, a transmission scheduling algorithm module, and an algorithm deployment module;

[0008] The system coordination module is used to connect with the service analysis module, the computing network perception module, the transmission scheduling algorithm module and the algorithm deployment module, call each module, and transmit the computing task information uploaded by the service analysis module and the computing network status information uploaded by the computing network perception module to the transmission scheduling algorithm module;

[0009] The service parsing module is used to obtain and parse the parameter information of the computing task; the parameters include task serial number, source IP, destination IP, cycle, task type, pre-calculation data volume, calculation task volume, post-calculation data volume and maximum allowed delay, and upload the computing task parameter information to the system coordination module;

[0010] The computing network perception module is used to perceive and collect the computing network status, which includes the remaining computing resources of each computing power node, the remaining capacity resources of each link, the propagation delay between connected nodes, and the complete topology of the network; and upload the computing network status information to the system coordination module;

[0011] The transmission scheduling algorithm module is used to execute the trained deterministic transmission scheduling algorithm based on the computing task information and computing network status information transmitted by the system coordination module, and output a transmission scheduling plan, which includes: the computing power node responsible for the calculation, the service path from the source node to the computing power node and from the computing power node to the destination node, and the transmission time slot;

[0012] The algorithm deployment module is used to send the transmission scheduling scheme output by the transmission scheduling algorithm module to switches and routers in the network, so that computing tasks are scheduled according to the transmission scheduling scheme.

[0013] Preferably, the computing network perception module is used to send a perception command to each node, instructing each node to send perception information to it. The perception information includes: the IP address of this node, the IP address list of adjacent nodes with remaining computing power of this node, the delay list between adjacent nodes and the remaining link capacity list between adjacent nodes. After receiving the perception information of all nodes, the complete computing network status information is integrated and uploaded to the collaborative scheduling module.

[0014] Preferably, the transmission scheduling module deploys an end-to-end transmission scheduling method based on D3QN.

[0015] Preferably, deterministic transmission mechanisms represented by Round Robin Specific Queuing Forwarding (CSQF) and Round Robin Queuing Forwarding (CQF) are deployed in the switches and routers.

[0016] Preferably, a network is set up to include an end layer, an edge layer and a cloud layer, the end layer includes a user terminal that initiates a computing task, the edge layer includes heterogeneous edge computing nodes, transmission nodes and heterogeneous networks, the deterministic transmission scheduling system for computing-network integration is deployed in the cloud layer, the service parsing module in the deterministic transmission scheduling system for computing-network integration is used as a task sensing server, the computing-network perception module is used as a computing-network perception server, the transmission scheduling algorithm module is used as an algorithm training server, the algorithm deployment module is used as an algorithm deployment server, and the system coordination module is used as a cloud computing server. The cloud computing server processes computing tasks, the task perception server collects information on computing tasks from the TSN switch of the end layer, the computing-network perception server collects computing power resources and network resources, the algorithm training server collects information in the task perception server and the computing-network perception server as status, takes the scheduling plan as an action, executes the D3QN algorithm, performs neural network training, and obtains a transmission scheduling plan. The algorithm deployment server sends the transmission scheduling plan to the switches and routers in the network.

[0017] According to another aspect of the present invention, a working method based on the deterministic transmission scheduling system for computing-network integration is provided, comprising:

[0018] The system coordination module transmits the computing task information uploaded by the service analysis module and the computing network status information uploaded by the computing network perception module to the transmission scheduling algorithm module;

[0019] The service parsing module obtains and parses the parameter information of the computing task; the parameters include task number, source IP, destination IP, cycle, task type, pre-calculation data volume, calculation task volume, post-calculation data volume and maximum allowed delay, and uploads the computing task parameter information to the system coordination module;

[0020] The computing network perception module senses and collects the computing network status, which includes the remaining computing resources of each computing power node, the remaining capacity resources of each link, the propagation delay between connected nodes, and the complete topology of the network; and uploads the computing network status information to the system coordination module;

[0021] The transmission scheduling algorithm module executes the trained deterministic transmission scheduling algorithm based on the computing task information and computing network status information transmitted by the system coordination module, and outputs a transmission scheduling plan. The transmission scheduling plan includes: the computing power node responsible for the calculation, the service path from the source node to the computing power node and from the computing power node to the destination node, and the transmission time slot;

[0022] The algorithm deployment module sends the transmission scheduling scheme output by the transmission scheduling algorithm module to switches and routers in the network, so that the computing tasks are scheduled according to the transmission scheduling scheme.

[0023] Preferably, the method further comprises:

[0024] The system coordination module models the network topology, G = {V, E}, G represents the network topology graph, assuming that there are n computing nodes in the network topology, V = {V1, V2, V3, ..., V n}, V represents the set of computing nodes; Vi represents the i-th computing node, i = 1, 2, 3, ..., n, C represents the set of computing power nodes, Ci represents the computing power of the i-th computing node, computing nodes are connected by data links, E∈{(x, y)|x, y∈V}, E represents the set of data links, assuming that there are m data links in the network topology, E={E1, E2, E3, ...E m}, E j represents the jth data link, j = 1, 2, 3, ..., n, each data link has a delay and capacity, D = {D1, D2, D3, ..., D m}, D represents the set of data link propagation delays, D j represents the propagation delay of the jth data link, j = 1, 2, 3, ..., m, B = {B1, B2, B3, ..., B m} represents the set of data link bandwidths, B j represents the link bandwidth of the jth data link, Cap = {Cap1, Cap2, Cap3, ..., Cap m} represents the set of data link capacities, Cap j represents the link capacity of the jth data link;

[0025] The system coordination module calculates the task characteristics, models the task, and uses F to describe the set of tasks. Assume there are n tasks in total, and F = {F1, F2, F3, ... F n}, F k Represents the kth computing task, using an octet (src k , dst k ,inf k ,com k , outf k , prd k , type k ) to describe the computing task, src k , dst k Represents the source node and destination node of the kth computing task; inf k Indicates the data volume of the kth computing task; com k Indicates the computing resources required for the kth computing task; outf k Indicates the amount of data after the calculation result of the k-th calculation task; prdk Indicates the cycle of the k-th computing task; Indicates the maximum delay allowed for the kth computing task; type k Represents the type of the k-th computing task;

[0026] For any computing task, the transmission scheduling algorithm module selects a suitable sending time and a path P from the source node to the destination node connected by links. The path P and the computing task have the same starting point and end point, so that the computing task f k For example, the following conditions must be met:

[0027] P=(E1,E2,…,E n )

[0028] E1.src = src k

[0029] E i .dst=E i+1 .src,

[0030] E n .dst=dst k ,

[0031]

[0032] in, Indicates that the computing task is in the computing power node V i Calculate on, Indicates that the computing task is not on the computing power node V i The calculation is performed on the power node V. The above formula shows that the calculation task is only performed on the power node V. j Calculate on;

[0033] Set four constraints: end-to-end latency constraint, conflict-free constraint, single computing node processing constraint, and capacity constraint;

[0034] Define the variable S k To represent the SDF calculation task f k Whether it is successfully scheduled: When the SDF computing task is successfully scheduled, S k =1, otherwise S k =0, use W i To describe the benefits of the weakly deterministic computing task WDF, if the delay of WDF meets the delay requirements of SDF, that is, Then W k =1, let parameter a represent the multiple of the allowed WDF delay exceeding the maximum delay; satisfy 1≤a, if but:

[0035]

[0036] Otherwise it is 0. The objective function is defined as follows:

[0037]

[0038] Set the states, actions, and rewards of a deterministic transmission scheduling algorithm based on the dueling dual deep Q network D3QN;

[0039] An action is described by a tuple (time, path, node). "Time" consists of a series of time slots that will be used to allocate computing tasks. It represents the time slot at which the computing task will be sent to the network. "Path" consists of all paths in the network from the starting point to the end point of the computing task. It represents the path that the computing task will take through the computing power network to reach the destination node. "Node" consists of all computing power nodes in the alternative path. It represents the node at which the computing task will be calculated.

[0040] DRSR uses the E-greedy algorithm, so the strategy π*:S→A is described as:

[0041]

[0042] Among them, A represents all actions, a t Represents the action performed by the agent at step t.

[0043] The state space consists of three parts: the remaining capacity of the link in different time slots, the remaining computing power of the computing node in different time slots, and the computing tasks that need to be scheduled at step t;

[0044]

[0045]

[0046] Among them, S t represents the state space at step t; represents the remaining capacity of all links in different time slots at step t; Represents the remaining computing power of all computing nodes in different time slots at step t; Represents the computing task that needs to be scheduled at step t; represents step t, time slot T i Downlink E j Remaining capacity; represents step t, time slot T i Lower computing power node V j Chain E j The remaining computing power;

[0047] The reward function is given by the following formula:

[0048]

[0049] The final solution obtained by the D3QN-based deterministic transmission scheduling algorithm for computing-network integration includes the time slot when the computing task is sent into the network, the transmission path from the source node to the destination node, and the computing power node responsible for the calculation in the transmission path.

[0050] It can be seen from the technical solutions provided by the above-mentioned embodiments of the present invention that the system and method of the present invention can integrate communication resources and computing resources to support the scheduling and transmission of deterministic computing tasks.

[0051] Additional aspects and advantages of the present invention will be set forth in part in the following description, will become apparent from the following description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 A schematic diagram of the structure of a deterministic transmission scheduling system for computing-network integration provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of an end-edge cloud network application scenario for the deterministic transmission scheduling system for computing-network integration of the present invention;

[0055] Figure 3 This is a flow chart of the deterministic transmission scheduling algorithm for computing-network integration provided in Example 1 of the present invention;

[0056] Figure 4 This is a flow chart of the working method of the deterministic scheduling system for computing-network integration provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0057] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0058] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0059] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention pertains. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0060] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0061] Example 1:

[0062] The structure of a deterministic transmission scheduling system for computing-network integration provided by an embodiment of the present invention is as follows: Figure 1 As shown, it includes: system coordination module, service analysis module, computing network perception module, transmission scheduling algorithm module and algorithm deployment module.

[0063] The system coordination module is used to connect with the service analysis module, computing network perception module, transmission scheduling algorithm module and algorithm deployment module, call each module, and realize efficient collaboration between modules; transmit the computing task information uploaded by the service analysis module and the computing network status information uploaded by the computing network perception module to the transmission scheduling algorithm module.

[0064] The service parsing module is used to obtain and parse the parameter information of the computing task; the parameters include task serial number, source IP, destination IP, cycle, task type, amount of data before calculation, amount of computing task, amount of data after calculation and maximum allowed delay; upload the computing task information to the system coordination module.

[0065] The computing network perception module is used to perceive and collect the computing network status, which includes the remaining computing resources of each computing power node, the remaining capacity resources of each link, the propagation delay between connected nodes, and the complete topology map of the network; and upload the computing network status information to the system coordination module.

[0066] The transmission scheduling algorithm module is used to execute the trained deterministic transmission scheduling algorithm based on the computing task information and computing network status information transmitted from the system coordination module, and output a transmission scheduling plan, which includes: the computing power node responsible for the calculation, the service path from the source node to the computing power node and from the computing power node to the destination node, and the sending time slot.

[0067] The algorithm deployment module is used to send the transmission scheduling plan output by the transmission scheduling algorithm module to the switches and routers in the network, so that the computing tasks are scheduled according to the transmission scheduling plan.

[0068] In some instances, the computing network perception module sends a perception command to each node, instructing each node to send perception information to it. This perception information includes: the node's IP address, a list of IP addresses of neighboring nodes with remaining computing power, a list of delays between neighboring nodes, and a list of remaining link capacities between neighboring nodes. If the computing network perception module does not receive perception information from a node, it will continue to send perception commands to that node. After receiving the perception information from all nodes, the computing network perception module integrates the complete computing network status information and uploads the computing network status information to the collaborative scheduling module.

[0069] In some examples, the transmission scheduling module deploys an end-to-end transmission scheduling method based on D3QN.

[0070] In some examples, the network adopts a deterministic transmission technology represented by the CSQF mechanism to ensure network determinism.

[0071] In embodiments of the present invention, computing tasks include deterministic computing tasks with certain latency requirements in scenarios such as the Industrial Internet, the Internet of Vehicles, and virtual reality. The network is a heterogeneous network, including but not limited to the Internet, satellite networks, and data center networks. Computing nodes such as CPUs (Central Processing Units), GPUs (Graphics Processing Units), FPGAs (Field Programmable Gate Arrays), and edge service servers provide computing services to complete the computing tasks.

[0072] The embodiments of the present invention also deploy deterministic transmission mechanisms represented by cyclic specified queuing and forwarding (CSQF) and cyclic queuing and forwarding (CQF) in switches and routers.

[0073] More specifically, the above-mentioned deterministic transmission scheduling system for computing-network integration targets network scenarios such as Figure 2 As shown, the network includes the end layer, edge layer and cloud layer.

[0074] The end layer primarily consists of user terminals that initiate computing tasks and some local computing resources with limited computing power. User terminals upload deterministic computing tasks with latency requirements for scenarios such as the Industrial Internet, the Internet of Vehicles, and VR / AR. Within a specific area of the end layer, a TSN (Time-Sensitive Networking) switch is located. The TSN switch is responsible for unified management of all user terminals within the area. User terminals are allowed to register or deregister only on the TSN switch in their area, and clock synchronization within the same area is also required.

[0075] The edge layer primarily consists of heterogeneous edge computing nodes, transmission nodes, and heterogeneous networks. Edge computing nodes are high-performance servers that process user computing tasks. They are deployed on heterogeneous networks, including but not limited to the internet, satellite networks, and data center networks. Efficient data transmission between these networks is guaranteed. Transmission nodes can only transmit data packets according to algorithms and cannot perform computational tasks. However, some edge computing nodes also transmit data packets.

[0076] The aforementioned deterministic transmission scheduling system for computing-network convergence is also deployed at the cloud layer. The service analysis module in this deterministic transmission scheduling system for computing-network convergence serves as a task sensing server, the computing-network perception module serves as a computing-network perception server, the transmission scheduling algorithm module serves as an algorithm training server, the algorithm deployment module serves as an algorithm deployment server, and the system coordination module serves as a cloud computing server. Cloud computing servers have strong computing capabilities and can handle computationally intensive tasks. The task sensing server collects computing task information from the TSN switches at the edge layer. The computing-network perception server is responsible for collecting all computing power and network resources. The algorithm training server collects information from the task sensing server and computing-network perception server as status, uses the scheduling plan (transmission time, computing nodes, service path) as an action, executes the D3QN algorithm, performs neural network training, and obtains the transmission scheduling plan. The algorithm deployment server distributes the transmission scheduling plan to the switches and routers in the network.

[0077] The embodiment of the present invention models the network topology based on the network characteristics. G = {V, E}. Where G represents the network topology. Assume that there are n computing nodes in the network topology, V = {V1, V2, V3, ..., V n}. Where V represents the set of computing power nodes; V i represents the i-th computing power node, i = 1, 2, 3, ..., n. C represents the set of computing power nodes. i Represents the computing power of the i-th computing node. The computing nodes are connected by data links, E∈{(x, y)|x, y∈V}. Where E represents the set of data links. Assume that there are m data links in the network topology. E={E1,E2,E3,…E m}. Among them, E j represents the jth data link, j = 1, 2, 3, ..., n. Each data link has a delay and a capacity. D = {D1, D2, D3, ..., D m}, D represents the set of data link propagation delays. j represents the propagation delay of the jth data link, j = 1, 2, 3, ..., m. B = {B1, B2, B3, ..., B m} represents the set of data link bandwidths. j Represents the link bandwidth of the jth data link. Cap = {Cap1, Cap2, Cap3, ..., Cap m} represents the set of data link capacities. j Represents the link capacity of the jth data link.

[0078] The embodiment of the present invention models the computing tasks according to their characteristics. The embodiment of the present invention uses F to describe the set of computing tasks. Assume that there are n computing tasks in total, and F = {F1, F2, F3, ... F n}. Among them, F k Represents the kth computing task. The embodiment of the present invention adopts the octet (src k , dst k ,inf k ,com k , outf k , prd k , type k ) to describe the computing task. k , dst k Represents the source node and destination node of the kth computing task; inf k Indicates the data volume of the kth computing task; com k Indicates the computing resources required for the kth computing task; outfk Indicates the amount of data after the calculation result of the k-th calculation task; prd k Indicates the cycle of the k-th computing task; Indicates the maximum delay allowed for the kth computing task; type k Represents the type of the k-th computing task.

[0079] Different computing tasks have different periods. To facilitate scheduling, this embodiment of the present invention defines a superperiod—the lowest common multiple of all computing task periods. The computing task information is identical across all superperiods. Assuming the network topology remains unchanged, the scheduling method for computing tasks within one superperiod can be generalized to all superperiods. Therefore, this embodiment of the present invention only considers the scheduling of computing tasks within one superperiod. In real-world scenarios, different computing nodes may not be time-synchronized, but time offsets can be measured and perceived by the controller. Therefore, in our experiments, we assume that all computing nodes are time-synchronized. Furthermore, considering that different computing tasks have different latency requirements, we categorize deterministic computing tasks into strongly deterministic and weakly deterministic ones. Strongly deterministic computing tasks are targeted at scenarios such as autonomous driving, telemedicine, and motion control, and have strict requirements on latency and jitter. Exceeding these latency and jitter limits poses security risks, and therefore cannot tolerate network instability and requires strict determinism. Weakly deterministic computing tasks are targeted at user-oriented scenarios such as cloud gaming, AR / VR, and the metaverse, and have strict latency and jitter requirements but can tolerate network instability to a certain extent.

[0080] This embodiment of the present invention models the latency of computing tasks based on their characteristics. Each task can be scheduled for computation on three types of nodes: end computing nodes at the end layer, edge computing nodes at the edge layer, and cloud computing nodes at the cloud layer. This embodiment of the present invention assumes that all tasks are transmitted and computed immediately without queuing.

[0081] End computing nodes are closest to the task initiator and the user terminal, requiring little or no propagation latency. However, they lack computing power and are suitable for tasks requiring minimal computation. Furthermore, the amount of data transmitted after computation is typically much smaller than the amount before, significantly reducing network overhead. If the computing nodes, task initiator, and task recipient are on the same end-to-end LAN, the computing power of the end layer is fully utilized.

[0082] When the algorithm deployment server selects local computing and the source node and computing node are the same, there is no need for propagation delay between the task sender and the computing node, only the computing delay needs to be considered. The computing delay of the computing task is T local .

[0083] T local =comk / C srck

[0084] in, Represents the computing task f k The computing power of the source node.

[0085] When the algorithm deployment server selects local computing and the source node and the destination node are directly connected, the computation delay, transmission delay, and propagation delay need to be considered. Assume that the link between the source node and the destination node is E k The transmission delay is Dk.trans, and the propagation delay is D k , total delay T sum is the sum of computation delay, transmission delay and propagation delay.

[0086] D k .trans=outf k |R k =outf k / [B k *lb(1+P*h k / No)]

[0087] T sum =T local +D k .trans+D k

[0088] Among them, R k Represents link E k The transmission rate; P represents the transmission power of the sending device, in watts. hk represents the channel gain of the link Ek. N o Represents the noise power in the link. lb means the logarithm with base 2.

[0089] The cloud nodes in the cloud layer have strong computing capabilities, but are far away from the task initiator and require a long propagation delay. They are suitable for computing tasks that are not sensitive to latency, such as large model training. The task needs to be transmitted from the source node to the computing node, and then transmitted from the computing node to the destination node after the computing task is completed. Assume that the computing node is node V com From the computing task source node src k To the computing node V com The path is path src，com ={src k , V1,…V com}; From the calculation node V com To the destination node dst of the computing task k The path is path com，dst ={V com , V2,…dst k}. When the source node uploads a task, a single node (node V j ) has a propagation delay of The calculated delay is: T compute ; When returning a result, a single node (node V a )’s propagation delay is: The total delay is T sum .

[0090]

[0091] T compute =com k |C i

[0092]

[0093]

[0094] T sum =D src，com .uptrans+D com，dst .downtrans+D k .prop+T compute

[0095] Among them, D src，com .uptrans represents the transmission delay of all nodes in the path of uploading the task to the computing node; D com，dst .downtrans represents the transmission delay of all nodes in the path of transmitting the result from the computing node to the destination node; D k .prop represents the total propagation delay from the source node to the destination node.

[0096] Edge computing nodes, located at the edge layer, have a computing power and propagation latency that fall somewhere in between. They are suitable for most tasks that have certain latency requirements and a certain amount of computational effort. Latency calculation is similar to that of cloud nodes.

[0097] The embodiment of the present invention defines the task scheduling and transmission problem based on network modeling, computing task modeling, and computing task delay modeling, and establishes constraint conditions and objective functions.

[0098] Specifically, for any computing task, the transmission scheduling algorithm module needs to select a suitable sending time and a path P connected by links from the source node to the destination node, and there must be only one computing node on the path to process the task to ensure that the computing task meets the deterministic transmission conditions. The path must have the same starting point and end point as the computing task. k For example, the following conditions must be met:

[0099] P=(E1,E2,…,E n )

[0100] E1.src = src k

[0101] E i .dst=E i+1 .src ，

[0102] E n .dst=dst k，

[0103]

[0104] in, Indicates that the computing task is in the computing power node V i Calculate on, Indicates that the computing task is not on the computing power node V i The above formula shows that the computing task is only performed on the computing power node V j Calculate on.

[0105] More specifically, the embodiment of the present invention adopts four constraints, namely end-to-end delay constraint, conflict-free constraint, single computing node processing constraint and capacity constraint.

[0106] End-to-end delay constraint: Any strongly deterministic computing task must meet the determinism requirement, that is, the total delay of the computing task must be less than the maximum delay.

[0107]

[0108] Among them, SDF represents the type of computing task, which is a strongly deterministic computing task.

[0109] The delay of a computing task consists of sending delay, propagation delay, processing delay, queuing delay and computing delay. k For example, suppose the computing node that performs the calculation is edge E J The destination node; the sending delay is D k .trans, the propagation delay is D k .prop; processing delay and queuing delay are D k .process and D k .queue; calculation delay is D k .com; the total delay is D k .sum.

[0110]

[0111] D k.sum=D k .trans+D k .prop+D k .queue+D k .com

[0112] When the network topology information is known, the processing delay is a constant value and is much smaller than other delays. Therefore, the present invention ignores the processing delay when calculating the delay.

[0113] No-conflict constraint: For any two computing tasks, messages cannot be sent to the same data link at the same time, otherwise a link conflict will occur.

[0114] Single computing node processing constraint: For any computing task, only one computing node can be selected for calculation.

[0115] Capacity constraint: At any moment, the amount of data scheduled to any link should not exceed the link capacity. Indicates the calculation task f at time t k The data to be sent is scheduled to link E i superior, Indicates the calculation task f at time t k The calculated result is dispatched to link E i Above, T stands for all moments.

[0116]

[0117] More specifically, the optimization goal of the embodiment of the present invention is to ensure that all strongly deterministic computing tasks are successfully scheduled while trying to ensure the latency requirements of weakly deterministic computing tasks. Define the variable S i To indicate whether the SDF computing task is successfully scheduled: When the SDF computing task is successfully scheduled, S i =1, otherwise S i =0. The present invention uses W i To describe the benefits of weak deterministic computing tasks (WDF). If the latency of WDF meets the latency requirements of SDF, that is, Then W i =1, let parameter a represent the multiple of the allowed WDF delay exceeding the maximum delay; satisfy 1≤a, if but

[0118]

[0119] Otherwise it is 0. The objective function is defined as follows:

[0120]

[0121]

[0122] Specifically, according to the requirements of the deterministic tasks, an embodiment of the present invention provides a deterministic transmission scheduling algorithm based on the Dueling Double Deep Q Network (D3QN), which clarifies the network status, actions and rewards.

[0123] This embodiment of the present invention uses the tuple (time, path, node) to describe an action. "Time" consists of a series of time slots that will be allocated to the computing task, representing the time slot at which the computing task will be sent to the network. "Path" consists of all paths in the network from the starting point to the end point of the computing task, representing the path the computing task will take through the computing network to reach the destination node. "Node" consists of all computing nodes in the alternative path, representing the node at which the computing task will be calculated.

[0124] DRSR uses the E-greedy algorithm, so the strategy π*:S→A is described as:

[0125]

[0126] Among them, A represents all actions, a t Represents the action performed by the agent at step t.

[0127] The state space consists of three parts: the remaining capacity of the link in different time slots, the remaining computing power of the computing node in different time slots, and the computing tasks that need to be scheduled at step t.

[0128]

[0129]

[0130] Among them, S t represents the state space at step t; represents the remaining capacity of all links in different time slots at step t; Represents the remaining computing power of all computing nodes in different time slots at step t; Represents the computing task that needs to be scheduled at step t; represents step t, time slot T i Downlink E j Remaining capacity; represents step t, time slot T i Lower computing power node V j Chain E j The remaining computing power.

[0131] The reward function is given by the following formula:

[0132]

[0133] More specifically, if Figure 3 As shown in the figure, the steps of the deterministic transmission scheduling algorithm for computing-network integration based on D3QN are as follows:

[0134] 1: Deploy the learning agent

[0135] Deploy a learning agent in the cloud to perceive the current network computing resources and computing task information that needs to be scheduled and processed.

[0136] 2: The learning agent selects appropriate sending slots and computing nodes for computing tasks.

[0137] The learning agent uses a deep neural network to select appropriate sending time slots and computing nodes for computing tasks based on the current remaining computing network resources and the computing task information that needs to be scheduled and processed.

[0138] The learning agent selects the appropriate service path for the computing task.

[0139] The learning agent uses another deep neural network to select a suitable service path for the computing task based on the current remaining computing network resources, the computing task information that needs to be scheduled and processed, the sending time slot of the computing task, and the computing node.

[0140] 3: Get rewards and update the network

[0141] After the learning agent performs an action, it receives a reward based on the changes in computing network resources and updates two deep neural networks based on the reward.

[0142] Example 2:

[0143] This embodiment is used to provide a working method of a deterministic transmission scheduling system for computing-network integration as described in Example 1. The processing flow of this method is as follows: Figure 4 As shown, the processing steps include the following:

[0144] S1: The service parsing module obtains and parses the parameter information of the computing task; the parameters include task number, source IP, destination IP, cycle, task type, amount of data before calculation, amount of computing task, amount of data after calculation and maximum allowed delay.

[0145] S2: The computing network perception module perceives and collects the computing network status, including the remaining computing resources of each computing power node, the remaining capacity resources of each link, the propagation delay between connected nodes, and the complete topology map of the network.

[0146] S3: After receiving the parameter information of the computing task and the current computing network status, the transmission scheduling algorithm module allocates computing nodes, transmission time slots and arranges service paths for the computing task according to the deterministic transmission scheduling method.

[0147] S4: The algorithm deployment module sends the transmission scheduling plan to the switches and routers in the network so that the computing tasks are scheduled according to the plan.

[0148] Among them, the transmission scheduling algorithm module takes the current remaining resources of the computing network, network topology, and computing task information that needs to be scheduled and processed as input, and uses a deterministic transmission scheduling scheme based on D3QN, specifically including: computing nodes, sending time slots, and service paths of computing tasks.

[0149] In summary, the embodiments of the present invention provide a deterministic transmission scheduling system and its working method for computing-network integration, including: a system coordination module responsible for calling each module to achieve efficient collaborative work between modules; a service parsing module obtains and parses parameter information of computing tasks; the parameters include task serial number, source IP, destination IP, cycle, task type, amount of data before calculation, amount of computing task, amount of data after calculation and maximum permitted delay; a computing network perception module is responsible for perceiving and collecting the computing network status, including the remaining computing resources of each computing power node, the remaining capacity resources of each link, the propagation delay between connected nodes and the complete topology map of the network; a transmission scheduling algorithm module is responsible for executing the trained deterministic transmission scheduling algorithm and providing a transmission scheduling plan; the algorithm deployment module sends the transmission scheduling plan to the switches and routers in the network so that the computing tasks are scheduled according to the plan.

[0150] The method and system of the present invention take into account various business needs and network environments, demonstrating strong adaptability and flexibility. While maintaining the quality of service for strongly deterministic flows, this method strives to meet the quality of service for weakly deterministic flows, enabling more efficient utilization of computing network resources. Through intelligent management and scheduling, the system of the present invention reduces network congestion and resource waste, improving overall reliability.

[0151] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0152] From the above description of the embodiments, it can be seen that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus the necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0153] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0154] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A deterministic transmission scheduling system for computing-network integration, characterized by: Including system coordination module, service analysis module, computing network perception module, transmission scheduling algorithm module and algorithm deployment module; The system coordination module is used to connect with the service analysis module, the computing network perception module, the transmission scheduling algorithm module and the algorithm deployment module, call each module, and transmit the computing task information uploaded by the service analysis module and the computing network status information uploaded by the computing network perception module to the transmission scheduling algorithm module; The service parsing module is used to obtain and parse the parameter information of the computing task; the parameters include task serial number, source IP, destination IP, cycle, task type, pre-calculation data volume, calculation task volume, post-calculation data volume and maximum allowed delay, and upload the computing task parameter information to the system coordination module; The computing network perception module is used to perceive and collect the computing network status, which includes the remaining computing resources of each computing power node, the remaining capacity resources of each link, the propagation delay between connected nodes, and the complete topology of the network; Upload computing network status information to the system coordination module; The transmission scheduling algorithm module is used to execute the trained deterministic transmission scheduling algorithm based on the computing task information and computing network status information transmitted by the system coordination module, and output a transmission scheduling plan, which includes: the computing power node responsible for the calculation, the service path from the source node to the computing power node and from the computing power node to the destination node, and the transmission time slot; The algorithm deployment module is used to send the transmission scheduling scheme output by the transmission scheduling algorithm module to switches and routers in the network, so that computing tasks are scheduled according to the transmission scheduling scheme; The system coordination module models the network topology, G={V, E}, where G represents the network topology. Assuming there are n computing nodes in the network topology, , V represents the set of computing power nodes; Representative computing nodes, i=1, 2, 3, ..., n, C represents the set of computing power nodes. Represents the computing power of the i-th computing node. The computing nodes are connected through data links. , E represents the set of data links. Assume that there are m data links in the network topology. , represents the jth data link, j = 1, 2, 3, ..., m, each data link has a delay and capacity, , D represents the set of data link propagation delays, represents the propagation delay of the jth data link, j=1, 2, 3, ..., m, represents the set of data link bandwidths, represents the link bandwidth of the jth data link, represents the set of data link capacities, represents the link capacity of the jth data link; The system coordination module calculates the task characteristics, models the task, and uses F to describe the set of tasks. Assume that there are n tasks in total, and , Represents the kth computing task, using an octet ( ) to describe the computational task, Represents the source node and destination node of the kth computing task; Indicates the data volume of the k-th computing task; Indicates the computing resources required for the k-th computing task; Indicates the amount of data after the calculation result of the k-th calculation task; represents the cycle of the k-th computing task; Indicates the maximum delay allowed for the k-th computing task; Represents the type of the k-th computing task; For any computing task, the transmission scheduling algorithm module selects a suitable sending time and a path P from the source node to the destination node connected by links. The path P and the computing task have the same starting point and end point. For example, the following conditions must be met: in, Represents a computing task At the computing power node Calculate on, Indicates that the computing task is not on the computing power node Calculation is performed on the computing power node. Calculate on; Set four constraints: end-to-end latency constraint, conflict-free constraint, single computing node processing constraint, and capacity constraint; Defining variables To indicate whether the strongly deterministic computing task SDF is successfully scheduled: when the SDF computing task is successfully scheduled =1, otherwise 0, use To describe the benefits of the weakly deterministic computing task WDF, if the delay of WDF meets the delay requirements of SDF, that is, ,but =1, let parameter a represent the multiple of the maximum delay allowed for WDF; satisfy 1 a, if ,but: otherwise is 0, the objective function is defined as follows: Set the states, actions, and rewards of a deterministic transmission scheduling algorithm based on the dueling dual deep Q network D3QN; Actions are described using a tuple (time, path, node). Time consists of a series of time slots that will be used to allocate computing tasks, indicating the time slot at which the computing task will be sent to the network. Path consists of all paths in the network from the starting point to the end point of the computing task, indicating the path the computing task will take through the computing network to reach the destination node. Node consists of all computing nodes in the alternative path, indicating the node at which the computing task will be calculated. Using the E-greedy algorithm, the strategy π*:S→A is described as: Among them, A represents all actions, represents the action performed by the agent at step t; The state space consists of three parts: the remaining capacity of the link in different time slots, the remaining computing power of the computing node in different time slots, and the computing tasks that need to be scheduled at step t; in, represents the state space at step t; represents the remaining capacity of all links in different time slots at step t; Represents the remaining computing power of all computing nodes in different time slots at step t; Represents the computing task that needs to be scheduled at step t; represents step t, time slot Downlink The remaining capacity; represents step t, time slot Downstream computing nodes chain The remaining computing power; The reward function is given by the following formula: The final solution obtained by the D3QN-based deterministic transmission scheduling algorithm for computing-network integration includes the time slot when the computing task is sent into the network, the transmission path from the source node to the destination node, and the computing power node responsible for the calculation in the transmission path.

2. The system according to claim 1, wherein: The computing network perception module in the deterministic transmission scheduling system for computing network integration is used to send perception commands to each node, instructing each node to send perception information to it. The perception information includes: the IP address of this node, the remaining computing power of this node, the list of IP addresses of adjacent nodes, the list of delays between adjacent nodes, and the list of remaining link capacities between adjacent nodes. After receiving the perception information of all nodes, the complete computing network status information is integrated and uploaded to the system coordination module.

3. The system according to claim 1, wherein: Deterministic transmission mechanisms represented by Round Robin Specific Queuing Forwarding (CSQF) and Round Robin Queuing Forwarding (CQF) are deployed in the switches and routers.

4. The system according to any one of claims 1 to 3, characterized in that A network is set up to include an end layer, an edge layer and a cloud layer, wherein the end layer includes a user terminal that initiates a computing task, and the edge layer includes heterogeneous edge computing nodes, transmission nodes and heterogeneous networks. The deterministic transmission scheduling system for computing-network integration is deployed in the cloud layer, and the service parsing module in the deterministic transmission scheduling system for computing-network integration is used as a task perception server, the computing-network perception module is used as a computing-network perception server, the transmission scheduling algorithm module is used as an algorithm training server, the algorithm deployment module is used as an algorithm deployment server, and the system coordination module is used as a cloud computing server. The cloud computing server processes computing tasks, the task perception server collects computing task information from the TSN switch of the end layer, the computing-network perception server collects computing power resources and network resources, the algorithm training server collects information from the task perception server and the computing-network perception server as status, uses the scheduling plan as an action, executes the D3QN algorithm, performs neural network training, and obtains a transmission scheduling plan. The algorithm deployment server sends the transmission scheduling plan to the switches and routers in the network.

Images