Task scheduling method and system of computing power network

Abstract

The application provides a kind of computing power network task scheduling method and system, the method comprises: obtaining the attribute of updated computing power network resource state and real-time computing task, determine the resource balance value of each routing pair of all scheduling schemes based on the updated resource state, select the scheduling scheme with the minimum resource balance value as the candidate scheduling scheme of each routing pair;Based on Markov decision process, the continuous scheduling problem of real-time computing task is characterized, the attribute of computing task, updated resource state, action and first reward are determined;Based on the attribute of computing task, updated resource state and load balancing experience, the candidate computing node with the maximum first reward in the candidate scheduling scheme of each routing pair in each time slot and the corresponding candidate segmented routing path are determined as the target computing node and the corresponding target segmented routing path of each computing task received by the source node in each routing pair, and the scheduling of the corresponding task is executed.The application improves the task success rate and load balancing degree of computing power network.

Description

Technical Field

[0001] This invention relates to the field of computing technology for computing networks, and in particular to a task scheduling method and system for large-scale computing networks. Background Technology

[0002] In recent years, with the continuous innovation and development of information technology, cutting-edge technologies such as computer vision and AIGC have emerged. These technologies require powerful computing power to run complex algorithms to achieve rapid and accurate analysis of massive amounts of data. To meet the stringent quality-of-service requirements of latency-sensitive and computationally intensive businesses in these applications, edge computing and fog computing are considered promising computing paradigms. They push computing power down to the user edge and have gradually evolved into various efficient collaborative computing paradigms such as cloud-edge-device vertical collaborative computing and small-scale edge horizontal collaborative computing. However, with the explosive growth of data and the rapid evolution of algorithms, computing nodes have gradually exhibited ubiquitous heterogeneous characteristics in terms of location, vendor, form, and capabilities. This has led to a contradiction between the limited resources of a single computing node and the uneven resource utilization between heterogeneous computing power, making it difficult to meet the exponentially growing computing power demand.

[0003] Fortunately, Computing Power Networks (CPNs) have emerged as a promising new paradigm, aiming to integrate ubiquitous heterogeneous resources based on ubiquitous network connectivity. Building upon interconnectivity, it draws inspiration from the "crowdsourcing" sharing economy concept found in services like Uber and Grab, encouraging both computing power providers and consumers to join the network for transactions, thus achieving resource interconnectivity. This computing paradigm stems from three key facts. First, for resource consumers, different computing tasks have varying requirements for service quality such as latency, accuracy, and reliability. Furthermore, large-scale tasks can typically be broken down into multiple subtasks requesting different computing services for parallel computation, a requirement that a single computing node cannot easily meet. Second, for resource providers, the ubiquitous heterogeneous computing resources and the massive application services deployed on them differ, hindering horizontal resource collaboration and ultimately leading to uneven resource utilization among multiple computing nodes. Finally, with breakthroughs in 6G technology, networks will gradually acquire the capabilities of inherent intelligence, ubiquitous interconnectivity, and multi-dimensional integration, effectively supporting deep perception and collaborative orchestration of heterogeneous ubiquitous resources. Therefore, the computing power network aims to promote the expansion of the existing network from a "ubiquitous connectivity platform" to a "heterogeneous resource collaboration platform" that integrates communication, sensing, computing and storage, and build a more collaborative and efficient green computing ecosystem.

[0004] Unlike traditional task scheduling, in computing power networks, scheduling is not limited to vertical offloading decisions or small-scale horizontal node collaborative decisions. Instead, it is based on a deep understanding of heterogeneous computing resources and combined with network status to perform collaborative orchestration and scheduling of computing tasks across the entire network. However, to achieve efficient collaborative scheduling in large-scale computing power networks, the following challenges still need to be overcome: 1) Resources: The resource requirements of computing tasks have spatiotemporal variations, and the computing network resources themselves are ubiquitous and heterogeneous, which can easily lead to computing power overload and network congestion due to uneven resource utilization. 2) Tasks: There is resource competition among massive computing tasks, and their service quality requirements vary, resulting in coupling and complexity in network-wide scheduling. 3) Network: Collaborative scheduling involves network transmission issues. Routing requires simultaneous consideration of computing power status and network status, which can easily lead to information redundancy and solution complexity in large-scale networks. 4) Algorithm: Although machine learning algorithms are often used to solve continuous scheduling problems, the high complexity of large-scale computing networks can easily lead to the curse of dimensionality, making it difficult for the algorithm to converge. Furthermore, the algorithm has poor robustness and is heavily dependent on the network state. When the network undergoes even a small change, the algorithm needs to be adjusted. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a task scheduling method and system for large-scale computing networks to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides a task scheduling method for a computing power network, the method comprising the following steps:

[0007] Obtain updated computing network resource status and real-time received computing task attributes, including resource requirements, source node, and destination node;

[0008] Based on the updated computing power network resource status, the resource balance value corresponding to each scheduling scheme of each routing pair is determined. The scheduling scheme with the smallest resource balance value is selected as the candidate scheduling scheme for each routing pair. Each routing pair includes a source node and a corresponding destination node. The candidate scheduling scheme includes at least one candidate computing node and at least one corresponding candidate segmented routing path.

[0009] Based on the Markov decision process to characterize the continuous scheduling problem of real-time computing tasks, the attributes of computing tasks, the updated computing network resource status, actions, and the first reward are determined. The actions include candidate scheduling schemes for each routing pair. The first reward includes second rewards for the success rate of computing tasks, resource balance, maximum computing node resource utilization, and maximum link utilization, as well as instantaneous rewards for the Lyapunov drift factor in each time slot.

[0010] Based on the attributes of the computing tasks, the updated computing power network resource status, and load balancing experience, the candidate computing node and the corresponding candidate segmented routing path with the maximum first reward in the candidate scheduling scheme of each routing pair in each time slot are determined as the target computing node and the corresponding target segmented routing path of each computing task received by the source node in each routing pair, and the scheduling of the corresponding computing tasks is executed according to the target computing node and the corresponding target segmented routing path.

[0011] In some embodiments of the present invention, the computing power network resources include computing power resources and network resources. Based on the updated computing power network resource status, the resource balance value corresponding to each scheduling scheme of each routing pair is determined by sensing the computing power resources and the segmented multi-factor weighted link weights in the network resources.

[0012] In some embodiments of the present invention, based on the updated computing power network resource status, the resource equilibrium value corresponding to each scheduling scheme of each routing pair is determined by sensing the computing power resources and the segmented multi-factor weighted link weights in the network resources, including:

[0013] The computing node resources and link weights in all scheduling schemes of each routing pair are respectively determined as the computing power resource value and network resource value of each routing pair. The link weights and the product of link bandwidth and capacity in the first half of the segmented routing path are inversely correlated, and the link weights and link reliability in the second half of the segmented routing path are inversely correlated.

[0014] The standard deviations of the computing resource values ​​and network resource values ​​of all routing pairs are respectively determined as the computing resource equilibrium value and the network resource equilibrium value;

[0015] The weighted sum of the computing power resource balance value and the network resource balance value is used as the resource balance value corresponding to each scheduling scheme of each routing pair.

[0016] In some embodiments of the present invention, in the step of selecting the scheduling scheme with the minimum resource balance value as the candidate scheduling scheme for each routing pair, if the computing node and segmented routing path in the selected scheduling scheme exceed the routing range of the computing power network and are limited by the routing boundary value, the computing node and segmented routing path exceeding the routing boundary value are pulled back into the routing range based on the boundary reflection method.

[0017] In some embodiments of the present invention, the segmented routing path includes a routing path from the source node to the computing node and a routing path from the computing node to the destination node.

[0018] In some embodiments of the present invention, the source node and the destination node corresponding to the computing task are the same node or different nodes.

[0019] In some embodiments of the present invention, the computing network resource status includes the computing power, storage capacity and connected load status of each node, as well as the security, reliability and connection status of each link.

[0020] The resource requirements of the computing task include the type of computing task, the amount of source data, the amount of computation, the amount of computing result data, latency requirements, and security requirements.

[0021] In some embodiments of the present invention, after obtaining the updated computing network resource status and the attributes of the computing tasks received in real time, the method further includes:

[0022] Based on the attributes of the computing task, nodes and links matching the updated computing network resource status are selected; and / or,

[0023] The computational tasks received by the source node in the same routing pair are categorized.

[0024] Another aspect of the present invention provides a task scheduling system for a computing power network. The system includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the aforementioned task scheduling method for a computing power network.

[0025] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned task scheduling method for a computing power network.

[0026] Another aspect of the present invention provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the aforementioned task scheduling method for a computing power network.

[0027] The task scheduling method and system for computing power networks of the present invention can solve the problems of solution complexity, complex coupling, dimensional curse and insufficient robustness brought about by large-scale computing power networks, improve the task success rate and load balancing of the entire computing power network, and avoid computing power overload and network congestion.

[0028] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0029] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0030] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0031] Figure 1 This is a schematic diagram of the computing network architecture in one embodiment of the present invention;

[0032] Figure 2 This is a flowchart illustrating a task scheduling method for a computing network according to an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of the technical route of the collaborative scheduling scheme for computing tasks in a large-scale computing network according to one embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of a network scheduling topology consisting of 16 CPN routers in one embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0036] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0037] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0038] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0039] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0040] Figure 1 This is a schematic diagram of the computing network architecture in one embodiment of the present invention. Figure 1 As shown, the architecture or framework of a computing power network, consisting of a user layer, a network layer, and a control layer, is used for scheduling computing tasks within the network. For the user layer, tasks where users generate requests through numerous applications to obtain computing resources (computing power network resources) are collectively referred to as computing tasks. In the trend of the Internet of Things and distributed computing, the generation and destination of computing tasks may be located in the same place or in different locations. For example, this could be due to the user's own movement (e.g., a user in a car submits a traffic flow prediction task at intersection A from point B, and the prediction result needs to be applied when the user moves to point A); it could be a case of users with different source and destination locations (e.g., a user collects a large number of images at point A, and the image calculation results need to be transmitted to point B for analysis); or it could be a case of distributed collaborative computing tasks (e.g., a user submits a large model pre-training computing task at point A, and the model training results need to be transmitted to point B for further fine-tuning). Based on this, this invention differs from the traditional assumption that only the source and destination are the same; it also considers computing tasks with different source and destination locations, and assumes that each task is an indivisible minimum task.

[0041] For the network layer, it consists of basic computing resources and network resources. The ubiquitous heterogeneous computing resources are CPN nodes (computing resources), and the routing and switching devices deployed at each node are CPN routers. Network resources are the links (resources) formed between CPN nodes. Unlike the vertical hierarchy of cloud, edge, and endpoint, CPN nodes in the computing power network are treated as distributed nodes with equal status, and each node is defined with a Computing Power Identifier (CPI): {IP address, security level, computing power} to distinguish different resources. In addition to traditional routing functions, CPN routers also act as CPN gateways, serving as the access point for CPN nodes and user-submitted computing tasks within the computing power network. Therefore, each CPN router maintains a virtual task arrival queue to store computing tasks accessing the computing power network (computing tasks are transmitted to the task arrival queue via wireless links).

[0042] For the control layer, the computing network adopts a centralized architecture in Software-Defined Networking (SDN) that separates data forwarding from control to orchestrate computing network resources. At the core of this layer is a programmable computing network controller. It exchanges information with the data plane at the network layer via the OpenFlow protocol through the SDN Control-Data Plane Interface (CDPI), and communicates with the computing network application services at the control layer through the SDN Northbound Interface (NBI) to invoke program functions. The computing network application service includes a computing network monitoring module (CMM), a collaborative scheduling module (CSM), and a policy conversion module (PDM). CMM uses a link layer discovery protocol to obtain the computing network topology and employs active or passive measurement methods such as Telemetry and Ping to periodically obtain the resource status of the computing network. CSM combines the computing network resource status information with an integrated algorithm to generate collaborative scheduling decisions based on computing power-aware routing for computing tasks. PDM converts the routing and scheduling decisions into Open Flow messages sent to the network layer to update the routing table entries of the CPN routers and guide the scheduling process of computing tasks.

[0043] In the aforementioned computing network framework, the scheduling of various computing tasks accessing each source node within each time slot involves the coordination of multiple routes. Specifically, this includes the CPN router r at the source node where the computing task accesses, and the CPN router r at the computing CPN node n selected by the computing task. * The coordination between the computation results and the CPN router r′ at the destination node is also crucial. In a large computing network, the decision-making, announcement, and updates of routes, as well as the selection and allocation of computing network resources, are extremely complex.

[0044] To address these challenges and to overcome the numerous difficulties in achieving efficient collaborative scheduling in large-scale computing networks, such as the spatiotemporal variability of resource demands from computing tasks and the ubiquitous heterogeneity of computing network resources leading to uneven utilization of ubiquitous resources resulting in computing overload and network congestion; resource competition among massive computing tasks and the service quality requirements of different tasks significantly increasing the coupling and complexity of network-wide scheduling; high information redundancy and solution complexity in large-scale computing networks; and the dimensionality curse and difficulties in convergence of machine learning algorithms caused by the high complexity in large-scale computing networks, this invention provides a task scheduling method and system for large-scale computing networks. This method and system are based on a collaborative scheduling framework of computing power-aware routing, aiming to solve the solution complexity brought about by large-scale computing networks and improve the overall task success rate and load balancing (the utilization balance of computing network resources).

[0045] Figure 2 This is a flowchart illustrating a task scheduling method for a computing network according to an embodiment of the present invention. Figure 2 As shown, the method includes the following steps:

[0046] Step S210: Obtain the updated computing network resource status and the attributes of the computing tasks received in real time, including resource requirements, source node, and destination node.

[0047] Specifically, in the user layer, user-submitted computing tasks are transmitted wirelessly to the task arrival queue in the CPN router at the source CPN node. Through deep perception and intent translation of the computing tasks, the resource requirements of the computing tasks and the computing power network requirements of the source CPN node and destination CPN node can be estimated. The resource requirements of the computing tasks include the type of computing task (e.g., video rendering task, model training task), source data volume, computational load, computation result data volume, latency requirements, and security requirements. In the computing power network, due to the requirements of the computing tasks, the source CPN node and destination CPN node corresponding to the computing task can be the same node or different nodes. The computing power network resource status includes the computing power, storage capacity, and connected load status of each CPN node, as well as the security, reliability, and connection status of each link.

[0048] Step S220: Based on the updated computing power network resource status, determine the resource balance value corresponding to each scheduling scheme of each routing pair, and select the scheduling scheme with the smallest resource balance value as the candidate scheduling scheme for each routing pair. Each routing pair includes a source node and a corresponding destination node. The candidate scheduling scheme includes at least one candidate computing node and at least one corresponding candidate segmented routing path.

[0049] The collaborative scheduling problem based on computational power-aware routing, which aims to reduce the solution complexity of large-scale computing power networks and improve the task success rate and resource utilization balance of the entire computing power network, is constructed as a revised Traveling Salesman Model (TSM). The revised TSM is then decomposed into two problems: the selection of candidate computing power network resources and collaborative task scheduling, to simplify the solution. Finally, it is extended to a long-term continuous scheduling process to form a Markov Decision Process (MDP) with reliability constraints, which effectively reduces the solution complexity of collaborative scheduling problems involving multiple source nodes, multiple destination nodes, and multiple computing tasks in large-scale computing power networks.

[0050] The Traveling Salesman Problem (TSP) is a classic combinatorial optimization problem aiming to find the shortest path through a given set of cities and back to the origin. In networking, the TSP is often used for path selection between source and destination nodes. This method's cooperative scheduling problem requires providing the cooperative scheduling decision defined by the Computational Scheduling Model (CSM), which involves selecting an intermediate computing node for each computational task and determining the complete path from the source node to the destination node, including the selected computing nodes along the way. This is the segmented routing path, which includes the routing path from the source node to the computing node and the routing path from the computing node to the destination node. It should be noted that the range of selectable computing nodes includes all nodes in the computing network, including both the source and destination nodes. Therefore, this invention improves upon the TSP by constructing a modified colored multi-warehouse multi-TSP open path model. Here, "multi-warehouse" represents multiple CPN routers, "multi-TSP" represents multiple computational tasks, "city" represents CPN nodes, "colored" indicates the security attributes of CPN nodes, and "open path" indicates that the source and destination nodes are not the same. For each traveling salesman, one and only one city from a given set of cities must be visited. That is, for each computational task, only one node is selected from all nodes between the source and destination nodes as the computational node, and the destination node is reached. Therefore, the problem solved by this invention is how to select the visited city and route, i.e., selecting the computational node and the corresponding complete path (segmented routing path), to provide a high task success rate and balanced resource utilization while satisfying the security requirements, link capacity constraints, and system reliability guarantees of the computational task. The TSP problem itself is an NP-hard problem, therefore... Figure 3As shown in the technical approach to solving the problem, the complex problem described above is decomposed into two sub-problems: candidate computing power network resources (candidate scheduling schemes) selection based on computing power-aware routing, and task collaborative scheduling based on traffic engineering. The former requires determining all candidate computing nodes and all candidate routing paths between each pair of source and destination CPN routers. The latter, based on the former, combines the attributes of the computing task to determine the target computing node and routing path from all candidate computing nodes and routing paths to complete task scheduling. In this method, computing power network resources are simply referred to as resources, including computing power resources and network resources. Computing power resources are computing node resources, and network resources are link resources in the routing path.

[0051] For candidate resource selection based on computational power-aware routing, it is necessary to determine candidate computing nodes and candidate segmented routing paths consisting of the corresponding preceding and following paths for the routing pair (r, r′). For example, CPN node n is selected as a candidate computing node for computation tasks, and one or more routing paths from the source node r to the computing node n need to be determined. and one or more routing paths from computing node n to destination node r′ This is used to transmit a single computing task connected to the source node r in the route pair, or to distribute the transmission of multiple computing tasks, in order to avoid network congestion. When determining candidate computing nodes and segmented routing paths, it is necessary to consider the balance of candidate computing power network resources between the source and destination nodes in the route pair, so as to ensure the fairness of the allocation of candidate computing power network resources as much as possible. Therefore, in step S220, based on the updated computing power network resource status, the resource balance value corresponding to each scheduling scheme of each route pair is determined by sensing the computing power resources and the segmented multi-factor weighted link weights in the network resources. Among them, the scheduling scheme includes at least one computing node and at least one corresponding segmented routing path.

[0052] Specifically, based on the updated computing power network resource status, the resource balance value corresponding to each scheduling scheme of each routing pair is determined by sensing the computing power resources and the segmented multi-factor weighted link weights in the network resources, including the following steps:

[0053] Step S221: Determine the computing power resource value and network resource value of each routing pair as the computing node resource value and network resource value of each routing pair respectively in all scheduling schemes of each routing pair and all link weights in all segmented routing paths. The link weight and the product of link bandwidth and capacity in the first half of the segmented routing path are inversely correlated, and the link weight and link reliability in the second half of the segmented routing path are inversely correlated.

[0054] Specifically, the link weights in the segmented routing path are determined using a segmented multi-factor weighting method. The link weight in the first half of the segmented routing path is the reciprocal of the product of the link bandwidth and capacity. This means that the bandwidth and capacity of the link are considered in the first half, which can avoid the high latency and overload risks caused by long-distance transmission of the original data. The link weight in the second half of the segmented routing path is the reciprocal of the link reliability. This means that the reliable transmission of the calculation results is considered more in the second half, which can avoid the loss of calculation results.

[0055] Step S222: Determine the standard deviations of the computing resource values ​​and network resource values ​​of all routing pairs in the computing power network as the computing resource equilibrium value and the network resource equilibrium value, respectively.

[0056] Step S223: The weighted sum of the computing power resource balancing value and the network resource balancing value is used as the resource balancing value for each routing pair and for all scheduling schemes. This resource balancing value can be used to measure whether the computing power resources (computing nodes) and network resources (links in the segmented routing path) allocated to the computing tasks in the scheduling scheme are balanced. The smaller the resource balancing value, the higher the load balancing degree.

[0057] In some embodiments, in step S220, when selecting the scheduling scheme with the smallest resource balance value as the candidate scheduling scheme for each routing pair, if the computing nodes and segmented routing paths in the selected scheduling scheme exceed the routing range of the computing power network and are limited by the routing boundary value, the computing nodes and segmented routing paths exceeding the routing boundary value are pulled back into the routing range based on the boundary reflection method.

[0058] In the process of selecting candidate computing nodes and candidate segmented routing paths, i.e., candidate scheduling schemes, finding the optimal candidate resources among all routing pairs becomes difficult as the network size increases. Therefore, this method considers using an improved metaheuristic algorithm, namely an improved atomic orbital search algorithm, to obtain near-optimal feasible solutions as candidate ranges with relatively low time and space costs, i.e., to determine candidate scheduling schemes or candidate resources, effectively reducing network computational complexity and information redundancy. Inspired by the fundamental principles of quantum mechanics, this algorithm constructs a quantum-based atomic model and disperses electrons around the atomic nucleus in Q virtual layers according to electron density configuration. Based on the principle of atomic energy absorption or emission, it simulates the movement of the electron cloud around the atomic nucleus at different locations, i.e., simulating the iterative process of generating candidate scheduling schemes. To this end, each selectable scheduling scheme (including computing nodes and corresponding segmented routing paths) for each routing pair is defined as an electron, and the resource equilibrium value corresponding to the scheduling scheme is the energy of that electron. The optimization objective of the iterative process is to find low-energy electrons, i.e., to find the scheduling scheme with the minimum resource equilibrium value: computing nodes and corresponding segmented routing paths. In this method, due to the large number of elements in the solution (selectable nodes and paths for each routing pair), the probability of the entire solution being limited to boundary values ​​because elements exceed the routing range of the computing power network increases significantly. This makes the original atomic orbital search algorithm prone to getting stuck at boundary values, leading to a reduction in the number of effective solutions. Therefore, this method improves the original atomic orbital search algorithm, forming an improved atomic orbital search algorithm. This algorithm uses boundary reflection to pull the solution back into the range for elements exceeding the boundary in each electron. In other words, it pulls the computing nodes and segmented routing paths that exceed the routing boundary value in each scheduling scheme back into the routing range. For example, if there are 10 CPN routers (CPN nodes) deployed in the computing power network, numbered 1-10, then the routing range is 1-10, and the boundary value is 10. For the scheme that selects router number 11, since router 11 does not exist, the solution needs to be reflected back into the range of 1-10. Without the boundary reflection method, the solution selected by number 11 would automatically become 10, but with reflection, 11 can become any other number within the range of 1-10.

[0059] For task collaborative scheduling based on traffic engineering, for a newly arriving computation task in the task arrival queue, the CPN router where the source node of the computation task is located and the CPN router where the destination node is located are used to determine the CPN node for computation and the segmented routing path for transmission from the candidate computing power network resources (candidate scheduling schemes). These are then used as the target computation node and target segmented routing path for the corresponding computation task, i.e., the target scheduling scheme. For traditional methods that do not employ a computation power-aware routing-based collaborative scheduling mechanism, the solution space is enormous, encompassing all CPN nodes and any feasible path in the computing power network. However, this method, based on computation power-aware routing, reduces the cost of route announcement and computing power addressing at the control plane, narrows the solution space of scheduling decisions to the candidate resource range, and considers the balance and fairness of candidate resource selection in the computation power-aware candidate resource selection. Based on this, when facing multi-computing-power network resources, the first step is to select CPN computing nodes, and then segment routing traffic engineering (SR-TE) technology is used to select the upstream and downstream transmission paths, thereby completing the coordinated scheduling of computing tasks across multiple computing nodes and multiple network links. Therefore, the scheduling decision for all global computing tasks can be defined as the traffic engineering scheduling decision between source and destination nodes in all routing pairs.

[0060] In traditional networks, the typical goal of traffic engineering is to reduce maximum link utilization. However, in computing power networks, it's also necessary to reduce maximum computing power (computing node) resource utilization and maximize the balance of computing power network resource utilization. Simultaneously, considering the coupling and competition among users, the completion rate of computing tasks should be maximized. Furthermore, it's necessary to ensure reliable allocation of users across multiple paths to meet their specific needs. Therefore, considering the complex coupling characteristics of the entire computing power network, the multi-optimization objectives of collaborative task scheduling decisions are: while satisfying security, link capacity, and reliability constraints, to increase the number of successful tasks and the balance of computing power network resource utilization, and to reduce maximum computing power resource utilization and maximum link utilization. In other words, to achieve a high success rate for computing tasks, achieve load balancing, and avoid network congestion.

[0061] Step S230: Based on the Markov decision process characterizing the continuous scheduling problem of real-time computing tasks, determine the attributes of the computing tasks, the updated computing network resource status, actions, and the first reward. The actions include candidate scheduling schemes for each routing pair. The first reward includes second rewards for the success rate of the computing tasks, resource balance, maximum computing node resource utilization, and maximum link utilization, as well as instantaneous rewards for the Lyapunov drift factor in each time slot.

[0062] Since the resource status of a computing network is constantly changing and updating, it is necessary to consider long-term dynamic scheduling decisions (scheduling schemes) to achieve long-term continuous collaborative scheduling. The resource status of the computing network at the next moment is only related to the resource status at the current moment and is independent of the resource status at previous moments. That is, the long-term collaborative scheduling process in the computing network has the Markov property, so the long-term continuous scheduling process of the computing network can be described as a Markov decision process. This process is specifically an interaction process between an agent and the environment, that is, the agent makes decisions or actions by observing the environment, thereby causing changes in the environment and obtaining certain rewards. Among them, 1) For the state, the agent needs to observe the attributes of newly arriving computing tasks and the resource status of the computing network to make collaborative scheduling decisions. Therefore, the state is defined as a combination of the attributes of the computing task and the resource status of the computing network, such as the computational load of the computing task, the source CPN router, the destination CPN router, the latency requirements, and the connection status of the links and the load status of the CPN nodes. 2) For actions, after observing the environment, the agent needs to formulate a collaborative scheduling decision based on traffic engineering for all computing tasks accessing the computing network. 3) Regarding rewards, to guide the agent's correct exploration, the agent will receive certain rewards after making decisions and interacting with the environment. The higher the success rate of computational tasks, the more balanced the utilization of computing network resources, and the smaller the maximum utilization rate of computing network resources (maximum computing node resource utilization and maximum link utilization), the higher the reward value. 4) In addition, considering the reliability guarantees provided by the system, the scheduling decisions should also meet reliability constraints. Based on this, the long-term collaborative scheduling process is constructed as an MDP process with reliability constraints. Its purpose is to comprehensively consider the computational tasks and computing network resources of the entire network during long-term continuous scheduling, improve the success rate of computational tasks and the balance of computing network resource utilization, reduce the maximum utilization rate of computing network resources, and meet the link reliability guarantee of the system.

[0063] In Markov decision processes with reliability constraints, the introduction of long-term constraints on reliability satisfaction increases the complexity of the problem and can lead to ineffective solutions. Lyapunov stability, a theory commonly used in dynamical control to describe the asymptotic stability of nonlinear systems, has also been frequently applied in recent years to address network service quality constraints. Specifically, defining a positive definite Lyapunov function and proving that the system tends towards an asymptotically stable point over time proves that the system satisfies the long-term constraints. Therefore, to observe the long-term constraint satisfaction status of user reliability satisfaction, this method first constructs a long-term reliability satisfaction deviation queue Δ. t+1This is used to represent the accumulated satisfaction difference up to time slot t+1 (referring to the difference between the system's guaranteed reliability and the actual reliability; if the actual reliability is greater than the system's guaranteed reliability, then the satisfaction difference is 0), as shown in the following formula:

[0064] Δ t+1 =P{Ω-Y(t+1)+Δ t}, Δ0=0

[0065] Where Ω represents the system's guaranteed or provided reliability, Y(t+1) represents the average reliability of the computation task traversing the link, and Ρ{x} is a positive function, meaning that Ρ{x} = 0 when x is negative. Then, the classic Lyapunov function is used to observe the convergence of this deviation queue, denoted as... Next, in order to narrow down the long-term convergence process of the satisfaction difference (the change of the satisfaction difference with time t) to the observation in each scheduling decision, a Lyapunov drift factor is introduced to capture the change of the Lyapunov value between every two consecutive time slots, denoted as σ. t+1 =L(Δ t+1 )-L(Δ t That is, the convergence of the deviation queue is mapped to the magnitude of the drift value in each time slot through the Lyapu transformation. When the drift value σ t+1 When the value approaches zero, the system satisfies the long-term reliability constraint, thus enabling long-term continuous cooperative scheduling. Therefore, this long-term reliability constraint can be transformed into an instantaneous reward for the drift factor within each time slot. Simultaneously, to guide the agent in making scheduling decisions by observing the satisfaction state of the deviation queue, the reliability satisfaction deviation queue is added as a new state, thereby transforming the MDP with reliability constraints into an unconstrained MDP.

[0066] Step S240: Based on the attributes of the computing task, the updated computing power network resource status, and load balancing experience, determine the candidate computing node and the corresponding candidate segmented routing path with the maximum first reward in the candidate scheduling scheme of each routing pair in each time slot as the target computing node and the corresponding target segmented routing path of each computing task received by the source node in each routing pair, and execute the scheduling of the corresponding computing task according to the target computing node and the corresponding target segmented routing path.

[0067] Based on the aforementioned selection and constraint transformation process of candidate resources or candidate scheduling schemes, Deep Reinforcement Learning (DRL) algorithm is used to solve the unconstrained Multivariate Dependency Programming (MDP) corresponding to long-term continuous collaborative scheduling. DRL algorithms are commonly used for solving MDPs, training neural networks to learn the interaction process between agents and the environment to obtain long-term decisions. However, the large-scale state and action spaces in computing power networks are susceptible to the curse of dimensionality, leading to learning difficulties. Therefore, this method improves the traditional DRL algorithm by proposing a deep reinforcement learning algorithm based on computing power awareness and experience-driven approaches. It reduces the solution space through computing power awareness and utilizes load balancing experience for path planning, effectively addressing the problems of dimensionality curse and insufficient robustness in large-scale computing power networks. Specifically, on the one hand, by using the determined candidate resources or candidate scheduling schemes based on computing power awareness as input to the improved DRL algorithm, a subset can be selected from the original solution space of traditional methods, ensuring that each routing pair in the large-scale computing power network has a balanced and fair pool of candidate computing power network resources. This reduces the spatial dimensionality of the entire network's collaborative scheduling, thereby alleviating the training complexity of the model in the algorithm. On the other hand, a load balancing strategy is adopted as an experience for collaborative scheduling based on traffic engineering, and a feasible optimization solution is formed based on candidate resources. This further guides the agent to learn in the right direction, increases the robustness of the algorithm to adapt to different environments, reduces the dependence on network state, and thus eliminates the need for frequent adjustments when the network undergoes minor changes.

[0068] In some embodiments, after obtaining the updated computing network resource status and the attributes of the real-time received computing tasks in step S210, the method further includes the following steps:

[0069] Step S212: Based on the attributes of the computing task, filter out nodes and links matching the computing task from the updated computing network resource status; and / or,

[0070] Step S213: Classify the computation tasks received by the source node in the same routing pair.

[0071] The above steps first cleanse the acquired computing network resource status, filtering out effective available resource information from the high-dimensional resource status information of the computing network. Then, from this effective available resource information, selectable computing nodes and links are determined, thereby identifying candidate computing nodes and candidate paths, which can improve accuracy and computational efficiency. Categorizing computational tasks for the same routing pair facilitates a more balanced and fair allocation of computing and network resources to the computational tasks of various routing pairs in the computing network, resulting in greater efficiency.

[0072] Specifically, such as Figure 1As shown, the overall process of collaborative scheduling based on computing power-aware routing includes the following steps:

[0073] ① Computing task access: The computing tasks submitted by users are transmitted via wireless link to the arrival queue in the CPN router at the source node where the user submitted the task. Through deep perception and intent translation of the computing tasks, the type of computing task, the amount of computing, the amount of source data, the amount of computing result data can be estimated, and the latency requirements, security requirements and destination node requirements of the computing task can be known.

[0074] ② CPN Status Information Acquisition: The CMM module acquires the updated CPN status, including CPN resource status and computing task attributes;

[0075] ③ State information cleaning: The CMM module cleans the collected CPN states and filters out valid information from the high-dimensional states; that is, steps S210, S212 and S213 are executed by the CMM module.

[0076] ④ Cooperative scheduling decision: The CSM module adopts a cooperative scheduling method based on computing power-aware routing for dynamic continuous scheduling. When it is necessary to update the route, it first dynamically selects candidate resources. Otherwise, it only needs to generate scheduling decisions based on candidate resources using an integrated algorithm (including the target computing nodes and corresponding target segment routing paths for each computing task in each time slot). That is, the scheduling decision generation steps in steps S220, S230 and S240 are executed by the CSM module.

[0077] Each CPN node only advertises its status to the selected route and updates it periodically when significant changes occur. This reduces routing and scheduling costs in large-scale computing networks and significantly improves information propagation efficiency. To achieve the offloading of massive computing tasks in the computing network to avoid congestion, the CSM module uses a multi-route algorithm to determine candidate SR paths for the source data of computing tasks from the CPN router at the source node to the CPN router at the candidate computing node, and candidate SR paths for the computing results from the CPN router at the candidate computing node to the CPN router at the destination node. The CSM module selects the target computing node and target segmented route path corresponding to the execution of computing task scheduling from the candidate resources. In Overlay mode, SR-TE technology is used to insert the selected route segment into the source data packet to guide its routing forwarding, and the segment combination can be dynamically adjusted to adapt to real-time changes in the computing network status.

[0078] ⑤ Execution of computation tasks: The PDM module converts the scheduling decision into an OpenFlow message and sends it to the CPN router at the source node of the corresponding computation task to guide the scheduling of the computation task to the corresponding CPN target computation node. Based on the latency requirements and computation volume of the computation task, the PDM module determines the corresponding priority and adds the task to the computation queue of the corresponding CPN target computation node in priority order. That is, the scheduling execution step in step S240 is executed by the PDM module. Generally speaking, the higher the latency requirement and the smaller the computation volume, the higher the priority of the computation task.

[0079] ⑥ Transmission of computation results: The computation results obtained by the target computing node after performing computation on the computation task are transmitted to the CPN router (destination node) where the destination is located through the second half of the target segmented routing path in the corresponding scheduling decision, and finally transmitted to the destination through the wireless link.

[0080] Through the above process, the collaborative scheduling architecture based on computing power-aware routing can effectively manage and optimize the scheduling and transmission of all computing tasks on all nodes in the computing power network, ensuring that computing tasks are completed efficiently and securely.

[0081] Figure 4 Here is an example of a network scheduling topology consisting of 16 CPN routers, such as... Figure 4As shown, each CPN node is deployed with a corresponding CPN router. In this network scheduling topology, for scheduling four computing tasks that access the CPN router (source node) numbered 6 and transmit the computing results of all tasks to the CPN router (destination node) numbered 11, this method uses a computing power-aware candidate resource selection process to determine the candidate computing nodes in the candidate scheduling scheme of the route pair (6,11) as all nodes corresponding to CPN routers (2,6,7,9,16), that is, the schedulable range of the four computing tasks. Ultimately, the scheduling decisions for the four computational tasks (red, blue, black, and green tasks) are shown by the arrows in the figure. The target computation node for the red task is the CPN node at CPN router 6, and the target segmented routing path is the red routing path. This means the computation is performed directly at the source node, and the result is transmitted along the red routing path to the destination node at CPN router 11. The target computation nodes for both the blue and black tasks are the CPN nodes at CPN router 2, and the target segmented routing paths are the blue and black segmented routing paths, respectively. This means the task first follows the first half of the blue and black segmented routing paths, respectively. The calculations for tasks originating from the source node are transmitted to the CPN node at CPN router 2. After the calculations are completed, the results are transmitted along the latter half of the blue and black segmented routing paths to the destination node at CPN router 11, respectively. The target calculation node for the green line task is the CPN node at CPN router 16, and the target segmented routing path is the green segmented routing path. Specifically, the calculations are first transmitted from the source node to the CPN node at CPN router 16 along the first half of the green segmented routing path. After the calculations are completed, the results are transmitted along the latter half of the green segmented routing path to the destination node at CPN router 11. This method also provides corresponding scheduling decisions for the calculation tasks of other routing pairs in the network.

[0082] Corresponding to the above method, the present invention also provides a task scheduling system for a computing power network. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the aforementioned task scheduling method for a computing power network.

[0083] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned task scheduling method for a computing network. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0084] This invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the aforementioned task scheduling method for a computing power network.

[0085] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0086] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0087] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A task scheduling method for a computing power network, characterized in that, The method includes: The system obtains updated computing network resource status and real-time received computing task attributes, including resource requirements, source node, and destination node. The computing network resources include computing resources and network resources. Based on the updated computing power network resource status, the resource equilibrium value corresponding to each scheduling scheme of each routing pair is determined by sensing computing power resources and using segmented multi-factor weighted link weights in the network resources. The scheduling scheme with the smallest resource equilibrium value is selected as the candidate scheduling scheme for each routing pair. Each routing pair includes a source node and a corresponding destination node. The candidate scheduling scheme includes at least one candidate computing node and at least one corresponding candidate segmented routing path. Specifically, the resource equilibrium value corresponding to each scheduling scheme of each routing pair is determined by sensing computing power resources and using segmented multi-factor weighted link weights in the network resources. The equilibrium value includes: determining the computing resource value and network resource value of each routing pair as the computing node resources and the link weights of each segmented routing path in all scheduling schemes of each routing pair, respectively; wherein the link weights and the product of link bandwidth and capacity in the first half of the segmented routing path are inversely correlated, and the link weights and link reliability in the second half of the segmented routing path are inversely correlated; determining the standard deviations of the computing resource values ​​and network resource values ​​of all routing pairs as the computing resource equilibrium value and the network resource equilibrium value, respectively; and using the weighted sum of the computing resource equilibrium value and the network resource equilibrium value as the resource equilibrium value corresponding to each scheduling scheme of each routing pair. Based on the Markov decision process to characterize the continuous scheduling problem of real-time computing tasks, the attributes of computing tasks, the updated computing network resource status, actions, and the first reward are determined. The actions include candidate scheduling schemes for each routing pair. The first reward includes second rewards for the success rate of computing tasks, resource balance, maximum computing node resource utilization, and maximum link utilization, as well as instantaneous rewards for the Lyapunov drift factor in each time slot. Based on the attributes of the computing tasks, the updated computing power network resource status, and load balancing experience, the candidate computing node and the corresponding candidate segmented routing path with the maximum first reward in the candidate scheduling scheme of each routing pair in each time slot are determined as the target computing node and the corresponding target segmented routing path of each computing task received by the source node in each routing pair, and the scheduling of the corresponding computing tasks is executed according to the target computing node and the corresponding target segmented routing path.

2. The method according to claim 1, characterized in that, In the step of selecting the scheduling scheme with the minimum resource balance value as the candidate scheduling scheme for each routing pair, if the computing nodes and segmented routing paths in the selected scheduling scheme are beyond the routing range of the computing power network and are limited by the routing boundary value, the computing nodes and segmented routing paths that exceed the routing boundary value are pulled back into the routing range based on the boundary reflection method.

3. The method according to any one of claims 1 to 2, characterized in that, The segmented routing path includes the routing path from the source node to the compute node and the routing path from the compute node to the destination node; The source node and destination node corresponding to the computing task can be the same node or different nodes.

4. The method according to any one of claims 1 to 2, characterized in that, The computing network resource status includes the computing power, storage capacity, and connected load status of each node, as well as the security, reliability, and connection status of each link. The resource requirements of the computing task include the type of computing task, the amount of source data, the amount of computation, the amount of computing result data, latency requirements, and security requirements.

5. The method according to any one of claims 1 to 2, characterized in that, After obtaining the updated computing network resource status and the attributes of the computing tasks received in real time, the method further includes: Based on the attributes of the computing task, nodes and links matching the updated computing network resource status are selected; and / or, The computational tasks received by the source node in the same routing pair are categorized.

6. A task scheduling system for a computing network, comprising a processor, a memory, and computer instructions stored in the memory, characterized in that, The processor is configured to execute the computer instructions, and when the computer instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

8. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the method according to any one of claims 1 to 5.

Images