Computing resource scheduling method and device, equipment and storage medium
By constructing a system state diagram and combining graph neural networks and large language models, the inefficiency caused by selecting equipment solely based on computing capacity in existing technologies is solved. This achieves joint optimization of computing resources and network status, thereby improving the overall task processing efficiency of the computing cluster.
Patent Information
- Application Number
- CN202611237187.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies select task execution devices based solely on the remaining computing capacity of the devices without considering network topology and link load. This can lead to tasks being assigned to devices with congested network paths but ample computing capacity, resulting in overall inefficiency.
By acquiring the location information of the output data of the predecessor task of the task to be scheduled, a system state diagram is constructed, and scheduling decisions are made using graph neural networks and large language models. The matching degree of computing devices and the smoothness of network paths are comprehensively considered to optimize task allocation.
This effectively avoids assigning tasks to devices with ample computing capacity but congested network paths, thus improving the overall task processing efficiency of the computing cluster.
Smart Images

Figure CN123068091A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of resource scheduling technology, and in particular to a method, apparatus, device and storage medium for computing resource scheduling. Background Technology
[0002] Current solutions select task execution devices based on the device's remaining computing capacity, without considering the device's location in the current network topology or the real-time load of the links that need to reach the device. This can lead to tasks being assigned to devices with the most abundant computing capacity but congested network paths, resulting in data transmission time far exceeding the computation itself. Consequently, the overall efficiency is actually lower than placing tasks on devices with unobstructed network paths but slightly lower computing capacity. Summary of the Invention
[0003] The main objective of this application is to provide a computing resource scheduling method, apparatus, device, and storage medium, which aims to solve the technical problem of low overall efficiency caused by selecting task execution devices based on the remaining computing capacity of the devices without considering the real-time load of the links.
[0004] To achieve the above objectives, this application proposes a computing resource scheduling method, the method comprising: Get the tasks to be scheduled in the ready queue. All predecessor tasks of the tasks to be scheduled have been completed. The attributes of the tasks to be scheduled include the location information of the output data of the predecessor tasks. Based on the current state of the tasks to be scheduled and the computing cluster, a system state diagram is constructed; The system state diagram is input into a graph neural network, which outputs a graph state vector. The graph state vector is projected into a prefix token sequence, and the prefix token sequence is input into a large language model to output a scheduling decision, so as to allocate the task to be scheduled to the corresponding target resource node.
[0005] Optionally, before obtaining the tasks to be scheduled from the ready queue, the method further includes: Obtain a directed acyclic graph workflow from user input, wherein the directed acyclic graph workflow includes multiple computation tasks and the data dependencies between the computation tasks; Traverse the directed acyclic graph workflow. When all predecessor tasks of the computation task have been executed, decouple the computation task from the directed acyclic workflow to obtain the task to be scheduled. The location information of the output data of the predecessor task is encapsulated as an attribute of the task to be scheduled, and the task to be scheduled is added to the ready queue.
[0006] Optionally, the system state diagram includes task nodes and resource nodes, and the construction of the system state diagram based on the current state of the task to be scheduled and the computing cluster includes: The task to be scheduled is taken as a task node, and the task attributes of the task to be scheduled and the location information of the output data of the predecessor task are encoded as the first attribute of the task node. Each computing device in the computing cluster is designated as a resource node, and the current state and device type of each computing device are encoded as the second attribute of the resource node. Based on the location information, a first type of edge is established between the task node and the resource node; Based on the physical connection relationship of the computing devices, a second type of edge is established between the resource nodes. The attributes of the second type of edge include the link bandwidth and real-time utilization of the corresponding link. A system state diagram is constructed based on the task nodes, resource nodes, the first type of edge, and the second type of edge.
[0007] Optionally, the step of inputting the system state graph into a graph neural network and outputting a graph state vector includes: The system state diagram is input into a graph neural network to perform multi-layer message passing, so that the task node and the resource node update information through edges; After message transmission is completed, the information of all nodes in the system state graph is aggregated through global pooling to generate graph feature vectors.
[0008] Optionally, the step of projecting the graph state vector into a prefix token sequence, inputting the prefix token sequence into a large language model, and outputting a scheduling decision to allocate the task to be scheduled to the corresponding target resource node includes: The graph state vector is converted into a prefix token sequence through a linear projection layer; Input the prefix token sequence into the large language model and output the hidden state representation; The hidden state representation is input into the first decision head, and the output is a first probability distribution representing the selection of each candidate resource node. The hidden state representation is input into the second decision head, and the output is a second probability distribution representing the selection of each candidate computational precision. A scheduling decision is determined based on the first probability distribution and the second probability distribution to allocate the task to be scheduled to the corresponding target resource node. The scheduling decision includes the target resource node and the target computational precision.
[0009] Optionally, after converting the graph state vector into a prefix token sequence through a linear projection layer, the method further includes: The graph state vector is read through a value assessment network; The graph state vector is processed by a shallow network in the value assessment network to output a state value estimate.
[0010] Optionally, after determining the scheduling decision based on the first probability distribution and the second probability distribution to allocate the task to be scheduled to the corresponding target resource node, the method further includes: The actual turnaround time of the task to be scheduled is obtained, and a reward signal is calculated based on the actual turnaround time. The reward signal is negatively correlated with the actual turnaround time. Based on the reward signal and the state value estimate output by the value assessment network, the generalized advantage estimate is determined. Based on the generalized dominance estimate, the parameters in the linear projection layer and the large language model are updated using the proximal policy optimization algorithm.
[0011] Furthermore, to achieve the above objectives, this application also proposes a computing resource scheduling device, which includes: The acquisition module is used to acquire the tasks to be scheduled in the ready queue. All predecessor tasks of the tasks to be scheduled have been completed. The attributes of the tasks to be scheduled include the location information of the output data of the predecessor tasks. The construction module is used to construct a system state diagram based on the current state of the task to be scheduled and the computing cluster; The output module is used to input the system state diagram into the graph neural network and output a graph state vector. The decision module is used to project the graph state vector into a prefix token sequence, input the prefix token sequence into the large language model, and output a scheduling decision to allocate the task to be scheduled to the corresponding target resource node.
[0012] In addition, to achieve the above objectives, this application also proposes a computing resource scheduling device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the computing resource scheduling method described above.
[0013] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the computing resource scheduling method described above.
[0014] This application achieves joint optimization of computing resource selection and network state awareness by acquiring the task to be scheduled, which includes the location information of the output data of the predecessor task, constructing a system state graph based on the task to be scheduled and the current state of the computing cluster, and then encoding it through a graph neural network and outputting the scheduling decision by a large language model. By incorporating network link status into the scheduling criteria, the scheduling decision can simultaneously consider the matching degree of computing devices and the smoothness of network paths, effectively avoiding the allocation of tasks to devices with sufficient computing capacity but congested network paths, thereby improving the overall task processing efficiency of the cluster. Attached Figure Description
[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a flowchart illustrating the first embodiment of the computational resource scheduling method of this application; Figure 2 This is a schematic diagram illustrating the workflow decoupling in this application; Figure 3 This is a schematic diagram of the Clos heterogeneous computing cluster structure in this application; Figure 4 This is a flowchart illustrating the second embodiment of the computational resource scheduling method of this application; Figure 5 This is a schematic diagram of the overall architecture of a GLS scheduling system; Figure 6 This is a schematic diagram of the module structure of the computing resource scheduling device according to an embodiment of this application; Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the computing resource scheduling method in the embodiments of this application.
[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.
[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.
[0021] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a computer, or an electronic device capable of performing the above functions. The following description uses a resource scheduling system as an example to illustrate this embodiment and the subsequent embodiments.
[0022] Based on this, embodiments of this application provide a computing resource scheduling method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the computational resource scheduling method of this application.
[0023] In this embodiment, the computing resource scheduling method includes: Step S10: Obtain the tasks to be scheduled in the ready queue. All predecessor tasks of the tasks to be scheduled have been completed. The attributes of the tasks to be scheduled include the location information of the output data of the predecessor tasks.
[0024] It should be noted that the ready queue is a global queue in the scheduling system that stores tasks in a schedulable state. A task to be scheduled is a task unit in the ready queue waiting to be allocated computing resources. A predecessor task is the direct upstream task of the task to be scheduled in the task dependency relationship. The location information of the predecessor task's output data is used to indicate the physical storage location of the input data required for the execution of the task to be scheduled, such as a resource node.
[0025] It is understood that the attributes of the task to be scheduled include computational workload, memory requirements, acceptable set of computational precision, and the location information of the output data of the predecessor task. Computational workload is used to characterize the scale of computing resources required by the task to be scheduled; memory requirements are used to characterize the storage capacity required during the execution of the task to be scheduled; and the acceptable set of computational precision is used to characterize the range of computational precision levels supported by the task to be scheduled.
[0026] In the specific implementation process, the originally complex global workflow optimization problem can be decomposed into a single-task scheduling problem that can be handled independently by decoupling. Based on the critical path principle, the overall completion time of the workflow is determined by the cumulative completion times of its constituent tasks. Optimizing the execution efficiency of each ready task can systematically shorten the overall completion time. At this point, before step S10, the following steps are also included: Obtain a directed acyclic graph workflow from user input, wherein the directed acyclic graph workflow includes multiple computation tasks and the data dependencies between the computation tasks; Traverse the directed acyclic graph workflow. When all predecessor tasks of the computation task have been executed, decouple the computation task from the directed acyclic workflow to obtain the task to be scheduled. The location information of the output data of the predecessor task is encapsulated as an attribute of the task to be scheduled, and the task to be scheduled is added to the ready queue.
[0027] It's important to note that a DAG (Directed Acyclic Graph) workflow consists of a set of computational tasks submitted by the user. Nodes represent computational tasks, and directed edges represent data dependencies between tasks. Data dependencies indicate that the output data of an upstream task serves as the input data for a downstream task, and a downstream task must wait for all its predecessor tasks to complete before it can begin execution.
[0028] It should be understood that computational tasks are the basic execution units in a directed acyclic graph workflow, and each computational task has independent computational and data requirements. Data dependencies between computational tasks are represented by directed edges, with the direction of the directed edges pointing from the upstream task to the downstream task, representing the direction of data flow.
[0029] Specifically, as an optional implementation, obtaining the directed acyclic graph workflow input by the user can involve a scheduling system receiving a workflow description file submitted by the user, parsing the identification information of each computational task, computational requirement parameters, and dependency descriptions between tasks in the workflow description file, and constructing the internal representation structure of the directed acyclic graph workflow based on the parsing results. The workflow description file adopts a standardized workflow definition format, such as JSON or YAML format.
[0030] Specifically, in this embodiment, the workflow preprocessing module is responsible for converting the complex DAG workflow into independent task units that the scheduler can process one by one. The first step is dependency detection: when all predecessor tasks of a task have been completed, the task is marked as ready and decoupled from the DAG structure to become an independent task to be scheduled. For example, the scheduling system maintains a list of predecessor tasks for each computation task. When a predecessor task is completed, it is removed from the list. When the list of predecessor tasks is empty, it is determined that all predecessor tasks of the computation task have been completed, triggering the decoupling operation. The second step is data source encapsulation: the location information of the output data generated by the predecessor tasks of the decoupled task is recorded as an attribute of the task. The scheduling system queries the execution records of each predecessor task, extracts the location information of the output data of the predecessor task from the execution records, and encodes the location information into the attribute fields of the task to be scheduled according to a preset format. The preset format includes a resource node identifier field and a storage path field. The third step is to enqueue the ready tasks: the encapsulated ready tasks are put into the global first-come-first-served ready queue, and the scheduler takes out the tasks from the queue one by one for processing.
[0031] In one example, reference Figure 2 , Figure 2 This is a schematic diagram illustrating the workflow decoupling in this application. 1. Input: Complex DAG workflow. The scheduling device obtains a complex DAG workflow input by the user as input. This DAG contains computational tasks in different execution states, where the completed task Task A is in the Done state, and the location of its output data is recorded as follows. Task B is in the Ready state because all its prerequisites have been met, while Task C is in the Wait state because its predecessor tasks have not all been completed. 2. Decoupling and Encapsulation. The scheduling device traverses and parses the DAG. When it detects that all predecessor tasks of Task B have been completed, it decouples Task B from the DAG structure. Simultaneously, it sets the output data position of Task B's predecessor task (i.e., Task A). Encapsulate the task attributes of Task B itself to form an atom. The atom The attributes include the computational workload. Memory requirements Acceptable set of computational precision and data source location 3. Global ready queue and future execution. After encapsulation, atomic... It is added to the Global Queue, along with other queued items. Other ready tasks wait to be retrieved sequentially by the GLS (Graph-LLMScheduler Scheduler, a scheduling system for graph neural networks and large language models). The GLS scheduler retrieves atoms from the global ready queue. Then, the scheduling decision is executed and the Allocate allocation operation is performed, ultimately transferring the atomic... Target resource nodes allocated to the computing cluster and with the selected precision This task enables decoupled scheduling of the entire process, from workflow parsing to resource allocation.
[0032] It should be understood that the technical reason for adopting this preprocessing method is that attempting to globally optimize the execution position and order of all tasks in a DAG is NP-hard in online scenarios where tasks arrive dynamically. According to the critical path principle, the overall completion time of a workflow is determined by the cumulative completion times of its constituent tasks; optimizing the execution efficiency of each task individually can systematically shorten the overall completion time. Simultaneously, encapsulating the data source location as a task attribute allows the scheduler to naturally handle dynamic workflow scenarios with conditional branches during subsequent decision-making.
[0033] Specifically, as an optional implementation, after obtaining the ready queue, retrieving the tasks to be scheduled from the ready queue can be done by the scheduling system traversing the global ready queue, retrieving the tasks to be scheduled from the head of the queue according to the first-come, first-served principle, and reading the attribute information of the tasks to be scheduled. The tasks in the global ready queue are arranged in the order of their entry into the queue, and the scheduling system processes each task in the queue sequentially.
[0034] In another alternative implementation, tasks in the ready queue are sorted by priority, which is determined based on the position of the task to be scheduled on the critical path in the directed acyclic graph workflow. The critical path is the longest path from the starting task to the ending task in the directed acyclic graph workflow, and tasks located on the critical path have a higher scheduling priority.
[0035] Optionally, when the ready queue is empty, the scheduling system enters a waiting state until a new task is marked as ready and added to the ready queue. When a task in the directed acyclic graph workflow is decoupled due to the completion of its predecessor task, the task is automatically added to the tail of the ready queue.
[0036] Step S20: Based on the current state of the task to be scheduled and the computing cluster, construct a system state diagram.
[0037] It should be noted that the system state diagram is a graph structure data dynamically constructed by the scheduling system for each task to be scheduled. It is used to represent the association between the task to be scheduled and the computing devices in the computing cluster, as well as the network connection relationships between the computing devices. The system state diagram includes two types of nodes: task nodes and resource nodes, as well as edges connecting the nodes.
[0038] It is understood that task nodes correspond to the tasks to be scheduled and are used to characterize the computational requirements and data source characteristics of the tasks to be scheduled. Resource nodes correspond to the computing devices in the computing cluster and are used to characterize the processing capacity and current load characteristics of each computing device. The edges are used to characterize the relationships between nodes, including the data dependencies between task nodes and resource nodes and the physical network connections between resource nodes. Of course, the edges can also reflect the connection relationships between L1 top-mounted switches and L2 subplane switches in the hierarchical network topology.
[0039] It should be understood that the computing cluster is a heterogeneous computing cluster, such as a Clos (Clos Network), which includes various types of computing devices. These computing devices include central processing units (CPUs), graphics processing units (GPUs), and dedicated accelerators. The computing devices are interconnected through a multi-layered switching network, forming a hierarchical network topology. The current state of the computing cluster includes the processing power, memory capacity, current load level, and device type of each computing device, as well as the link bandwidth and real-time utilization of each network link.
[0040] In one example, reference Figure 3 , Figure 3 This is a schematic diagram of the Clos heterogeneous computing cluster structure of this application. The computing cluster is the core node in the computing network that undertakes the actual computing tasks. Its topology, from bottom to top, includes multiple computing nodes and two levels of switching planes. Each computing node integrates various heterogeneous computing devices such as a central processing unit (CPU), a graphics processing unit (GPU), and a dedicated accelerator. Each node is connected to its corresponding subplane through a first-level switch (L1 switch). Multiple L1 switches within the same subplane converge and connect to a second-level switch (L2 switch), forming a hierarchical interconnection structure between the L2 subplane and the L1 subplane. Specifically, each L1 subplane covers a group of computing nodes. The CPU, GPU, and accelerator within a node communicate with each other and connect to the external network via the L1 switch of that subplane. Multiple L1 subplanes further exchange data across subplanes through L2 switches, thus forming the complete Clos heterogeneous computing cluster topology. In this topology, the links between L2 switches and L1 switches, and the links between L1 switches and each computing node, are all physical network connection edges. Their link bandwidth and real-time utilization are encoded as attributes of the second type of edge in the system state diagram, so that the scheduler can perceive the network path status when making decisions.
[0041] Understandably, when a cluster needs to process a large number of computing tasks simultaneously, the scheduling system needs to select a suitable computing device for each task, and also determine which network path the data required by the task should be transmitted along. These two decisions are interrelated: the same task executed on different devices will have different data transmission paths and latency. Therefore, considering both computing resource selection and network conditions during scheduling is key to achieving high-efficiency cluster operation.
[0042] In practical implementation, link bandwidth and real-time utilization can be encoded in the attributes of edges, so that network topology connectivity and link congestion status are explicitly incorporated into the system state representation, and the topology connectivity is preserved, enabling the scheduler to distinguish the differences in network paths traversed between the data source of a task and candidate devices. At this point, step S20 may include: The task to be scheduled is taken as a task node, and the task attributes of the task to be scheduled and the location information of the output data of the predecessor task are encoded as the first attribute of the task node. Each computing device in the computing cluster is designated as a resource node, and the current state and device type of each computing device are encoded as the second attribute of the resource node. Based on the location information, a first type of edge is established between the task node and the resource node; Based on the physical connection relationship of the computing devices, a second type of edge is established between the resource nodes. The attributes of the second type of edge include the link bandwidth and real-time utilization of the corresponding link. A system state diagram is constructed based on the task nodes, resource nodes, the first type of edge, and the second type of edge.
[0043] It should be noted that task nodes correspond to the tasks to be scheduled and are used to characterize the computational requirements and data source characteristics of the tasks to be scheduled. Resource nodes correspond to the computing devices in the computing cluster and are used to characterize the processing capacity and current load characteristics of each computing device. The edges are used to characterize the relationships between nodes, which may include the data dependencies between task nodes and resource nodes, as well as the physical network connections between resource nodes.
[0044] It is understood that the first attribute is the node feature vector of the task node, which is used to encode the multi-dimensional attribute information of the task to be scheduled.
[0045] Specifically, as an optional implementation, the encoding operation may include quantifying the computational workload, the memory requirements, the acceptable set of computational precision, and the location information, and concatenating them into a fixed-length feature vector according to a preset dimensional order, which serves as the first attribute of the task node.
[0046] It is understood that resource nodes are graph nodes representing the computing device in the system state graph. The second attribute is the node feature vector of the resource node, used to encode the multi-dimensional state information of the computing device. The current state includes processing power, memory capacity, and current load level.
[0047] Specifically, as an optional implementation, the encoding operation may further include quantifying the processing capability, memory capacity, current load level, and device type, and concatenating them into a fixed-length feature vector according to a preset dimensional order, which serves as the second attribute of the resource node. The device type is represented using one-hot encoding, and the processing capability, memory capacity, and current load level are represented using normalized numerical values.
[0048] It should be understood that the first type of edge is the connection edge between the task node and the resource node where the output data of the predecessor task is located. The first type of edge is used to characterize the data source location of the task to be scheduled, so that the system state diagram includes a direct connection relationship from the task to be scheduled to the computing device where its data source is located.
[0049] Specifically, the scheduling system reads the location information from the first attribute of the task node, parses the resource node identifier in the location information, searches for the resource node corresponding to the resource node identifier in the system state graph, and establishes a first-type edge connecting the task node and the resource node. This first-type edge is a directed edge, pointing from the resource node to the task node, representing the flow of data from the resource node to the task node.
[0050] It is understood that the second type of edge represents the connection edge between resource nodes reflecting the physical network connection. This second type of edge is used to characterize the network topology connection relationship between various computing devices in the computing cluster. The attributes of the second type of edge include the link bandwidth and real-time utilization rate of the corresponding link. The link bandwidth characterizes the maximum data transmission rate of the link, and the real-time utilization rate characterizes the proportion of bandwidth currently occupied by the link.
[0051] Specifically, the scheduling system queries the network topology configuration information of the computing cluster to obtain the physical connection relationships between each computing device. In the system state diagram, it establishes a second type of edge for resource node pairs with physical connections, and encodes the link bandwidth and real-time utilization corresponding to the connection relationship as edge attributes of the second type of edge. The second type of edge is an undirected edge, representing a bidirectional physical connection relationship.
[0052] Step S30: Input the system state diagram into the graph neural network and output the graph state vector.
[0053] It should be noted that the graph neural network is a graph neural network model based on a message-passing mechanism, used for feature extraction and representation learning of the system state graph. The graph state vector is a fixed-dimensional vector output by the graph neural network after encoding the system state graph, used to summarize all the key information in the system state graph, including the availability status of each computing device, the load status of each network link, and the data source location information of the task to be scheduled.
[0054] In the specific implementation process, multi-layer message passing can be performed on the system state graph, enabling each node to aggregate the state information and edge attribute information of its neighbors when updating its own state. This allows each node's encoding vector to simultaneously contain its own attribute information and neighbor information passed through the network topology. This ensures that the graph feature vector output after global pooling summarizes all the key information in the current system snapshot, namely, which devices are available, the load on each link in the network, and the location of the task data. This fixed-dimensional graph feature vector achieves structured encoding of the state of clusters of arbitrary size, enabling the subsequent inference and decision-making module to make scheduling decisions based on a complete state representation containing network topology information. At this point, step S30 may include: The system state diagram is input into a graph neural network to perform multi-layer message passing, so that the task node and the resource node update information through edges; After message transmission is completed, the information of all nodes in the system state graph is aggregated through global pooling to generate graph feature vectors.
[0055] It should be noted that multi-layer message passing includes multiple message passing layers, and each message passing layer performs a complete message passing operation on all nodes in the graph once. In each layer of message passing, each node first receives a message sent by its neighboring nodes, where the message is the current representation vector of the neighboring node. The node then aggregates the received message with its own representation vector to generate an updated representation vector.
[0056] Understandably, each node obtains information from both its own attributes and the states of its neighbors during message passing. After several layers of message passing, a global pooling operation aggregates all node information into a graph-level state vector with a fixed dimension. This vector summarizes all the key information in the current system snapshot: which devices are available, the load on each link in the network, and where the task data is located.
[0057] Specifically, during message transmission, the scheduling system reads the edge attributes of each edge and merges these attributes with the message content to generate an edge-enhanced message. This edge-enhanced message is then transmitted to the receiving node through the edges, incorporating the edge attribute information into the receiving node's representation vector. After message transmission is complete, global pooling is used to aggregate the information of all nodes in the system state graph. This global pooling employs summation pooling; the scheduling system sums the final representation vectors of all nodes in the system state graph element-wise to generate the graph feature vector. The specific implementation details of the above steps are relatively conventional and will not be described in this embodiment.
[0058] Understandingly, the heterogeneous devices and scheduled tasks in the computing cluster are abstracted into two types of graph nodes, and the physical network connections between devices are abstracted into graph edges. Link bandwidth and real-time utilization are encoded in the edge attributes. Message passing is based on this graph structure, ensuring that each node's encoded vector contains both its own attribute information and neighbor information transmitted through the network topology. When the scheduler selects a device for each task, its decision-making criteria naturally include network-side information such as which network paths must be traversed from the data source to the candidate device and the current status of these paths.
[0059] Step S40: Project the graph state vector into a prefix token sequence, input the prefix token sequence into the large language model, and output a scheduling decision to allocate the task to be scheduled to the corresponding target resource node.
[0060] It should be noted that the prefix token sequence is a fixed-length token sequence transformed from the graph state vector, used as the input prefix of the large language model. The large language model is a pre-trained language model with the ability to reason about complex constraints. The scheduling decision is a resource allocation scheme generated by the large language model based on the prefix token sequence, used to determine the execution position and computational precision of the task to be scheduled.
[0061] Understandably, the projection operation is implemented through a trainable linear projection layer, which can be a trained fully connected neural network layer, mapping the graph state vector from the output space of the graph neural network to the input embedding space of the large language model. The projection operation makes the graph encoding results compatible with the language model input format while maintaining the integrity of the graph structure information.
[0062] In this embodiment, by acquiring the task to be scheduled, which includes the location information of the output data of the predecessor task, and constructing a system state graph based on the task to be scheduled and the current state of the computing cluster, and then encoding it through a graph neural network and outputting the scheduling decision by a large language model, joint optimization of computing resource selection and network state awareness is achieved. Incorporating network link status into the scheduling criteria allows the scheduling decision to simultaneously consider the matching degree of computing devices and the smoothness of network paths, effectively avoiding the allocation of tasks to devices with sufficient computing capacity but congested network paths, thereby improving the overall task processing efficiency of the cluster.
[0063] Reference Figure 4 , Figure 4 This is a flowchart illustrating the second embodiment of the computing resource scheduling method of this application. Based on the first embodiment described above, a second embodiment of the computing resource scheduling method of this application is proposed.
[0064] In the second embodiment, step S40 includes: Step S401: Convert the graph state vector into a prefix token sequence through a linear projection layer.
[0065] Understandably, a trainable projection layer transforms the graph-level state vector output by the graph neural network into a fixed-length sequence of prefix tokens, which serves as the input prefix for the large language model. This mechanism separates the cluster size (up to hundreds of devices) from the model inference scale: when adjusting the cluster capacity, only the graph neural network capacity needs to be adjusted, while the large language model remains unchanged. Simultaneously, it complements the structured encoding capabilities of the graph neural network with the complex constraint inference capabilities of the large language model, enabling scheduling decisions to simultaneously consider the structural characteristics of the network topology and the implicit constraints between multiple factors.
[0066] It should be understood that the projection adaptation layer transforms the graph-level state vector output by the graph coding module into a set of prefix tokens through a trainable linear projection. The number of prefix tokens is fixed and has no direct correlation with the cluster size. Therefore, when the number of cluster nodes changes, only the capacity of the graph neural network needs to be expanded, and the inference core does not need to be changed.
[0067] In the specific implementation process, a value evaluation network that bypasses the inference model can be set up to directly assist in updating policy parameters during the training phase, forming an asymmetric policy network and value network architecture. The value network is used to assist in calculating the advantage function during the training phase and does not participate in the actual scheduling decision. This avoids the huge computational overhead caused by the deep propagation of gradients through a large language model during training. Furthermore, the value network is updated independently of the policy network and is not affected by the training instability that may be caused by deep gradient propagation in the inference model, making the training process more stable and efficient. At this point, after step S401, the following is also included: The graph state vector is read through a value assessment network; The graph state vector is processed by a shallow network in the value assessment network to output a state value estimate.
[0068] It should be noted that the value evaluation network is an auxiliary network independent of the large language model. It is used to evaluate the value of the current system state during the training phase and does not participate in the actual scheduling decision. The state value estimate is a scalar value output by the value evaluation network after processing the graph state vector, which is used to characterize the expected cumulative reward of the current system state.
[0069] Understandably, the value assessment network bypasses the large language model, directly reads the graph state vector, and processes it through a shallow network. This shallow network can be a neural network with fewer layers, including one or more fully connected layers. The design of the shallow network reduces training computational overhead without affecting the quality of the final decision, and makes the value assessment unaffected by the instability caused by deep gradient propagation in the large language model.
[0070] Optionally, the single-layer fully connected network has few parameters and low computational cost, making it suitable for training scenarios where value assessment is frequently updated. The value assessment network updates its parameters at a different frequency than the large language model. The large language model has a lower parameter update frequency, while the value assessment network has a higher parameter update frequency to adapt to the rapidly changing value function in reinforcement learning.
[0071] Step S402: Input the prefix token sequence into the large language model and output the hidden state representation.
[0072] It should be noted that the hidden state representation is the final layer representation output by the large language model after processing the prefix token sequence, which contains high-order features that integrate graph structure information and semantic reasoning information.
[0073] Specifically, as an optional implementation, the scheduling system inputs the prefix token sequence into the input embedding layer of the large language model. The input embedding layer fuses each token vector with the corresponding positional encoding information to generate an input representation with positional information. The input representation is then sequentially input into the multi-layer transformer of the large language model. Each transformer captures the dependencies between tokens through a self-attention mechanism, performs nonlinear transformations through a feedforward network, and extracts higher-order features layer by layer. The representation output by the final transformer layer is used as the hidden state representation.
[0074] It should be understood that the hidden state is represented as the representation vector output by the final layer transformer of the large language model, used as the input feature of the subsequent decision output head. The dimension of the hidden state representation is consistent with the dimension of the hidden layer of the large language model.
[0075] Step S403: Input the hidden state representation into the first decision head, and output the first probability distribution representing the selection of each candidate resource node.
[0076] Step S404: Input the hidden state representation into the second decision head, and output a second probability distribution representing the selection of each candidate computational precision.
[0077] It should be noted that the first decision head and the second decision head are two independent output modules downstream of the large language model, responsible for generating the resource selection probability distribution and the precision selection probability distribution, respectively. The first probability distribution is the set of probability values for each candidate resource node to be selected, and the second probability distribution is the set of probability values for each candidate computational precision to be selected.
[0078] Understandably, the first decision head primarily focuses on network topology, load balancing, and data locality factors to determine the execution device for the task to be scheduled. The second decision head primarily focuses on the trade-off between computational accuracy and execution efficiency to determine the computational precision of the task to be scheduled.
[0079] In one example, the inference core is based on a pre-trained large language model. Prefix token sequences enter the inference core, are processed by multiple Transformer layers, and then output as the final hidden state. The pre-trained weights of the inference core remain frozen during training; only a low-rank update matrix is introduced into the attention projection matrix for lightweight task adaptation, resulting in less than one percent of the total trainable parameters. This approach allows the same pre-trained model to serve multiple clusters with different configurations, each cluster only needing to maintain its own lightweight adapter.
[0080] The decision output head generates two probability distributions from the hidden states of the inference core through two independent linear projections. One is the probability distribution of the selection of candidate computing devices, primarily focusing on network topology, load balancing, and data locality. The other is the probability distribution of the selection of candidate computational precision, primarily focusing on the trade-off between computational accuracy and execution efficiency. The final scheduling action is obtained by jointly sampling from these two distributions, simultaneously determining on which device and with what precision the task will be executed.
[0081] Step S405: Determine a scheduling decision based on the first probability distribution and the second probability distribution to allocate the task to be scheduled to the corresponding target resource node. The scheduling decision includes the target resource node and the target computational precision.
[0082] It should be noted that the scheduling decision is a joint action generated based on the first and second probability distributions, used to simultaneously determine the execution location and computational precision of the task to be scheduled. The target resource node is the execution device selected from among the candidate resource nodes, and the target computational precision is the numerical precision level selected from among the candidate computational precision levels.
[0083] It is understandable that the process of determining the scheduling decision is a process of sampling or selecting the optimal combination from the first probability distribution and the second probability distribution.
[0084] In an optional implementation, sampling from the first probability distribution employs a top-k sampling strategy, whereby the scheduling system first selects the candidate resource node with the highest probability value from the first probability distribution. The sampling method from the second probability distribution is similar to that of the first probability distribution.
[0085] It should be understood that the scheduling system adds the tasks to be scheduled to the local task queue of the target resource node, and the target resource node executes each task in sequence according to the queue order. The local task queue adopts priority sorting or first-in-first-out sorting.
[0086] It is understandable that the target computational precision affects the execution efficiency and computational accuracy of the task to be scheduled on the target resource node. Higher computational precision provides more accurate computational results but consumes more computational resources and time, while lower computational precision improves execution efficiency but may reduce the accuracy of computational results.
[0087] In the specific implementation process, the reward signal can be calculated based on the actual turnaround time after task execution, and the generalized advantage estimate can be determined by combining the state value estimate output by the value network. The parameters in the linear projection layer and inference model are then updated using a proximal policy optimization algorithm. This allows the scheduling policy to autonomously explore better solutions through continuous interaction with the cluster environment, without being limited by the upper limit of traditional schedulers. Simultaneously, only the adaptation parameters in the linear projection layer and inference model are updated during training, while the pre-trained weights of the inference model remain frozen. This allows the same base model to serve multiple clusters with different configurations, with each cluster only needing to maintain its own lightweight adapter. At this point, after step S405, the following is also included: The actual turnaround time of the task to be scheduled is obtained, and a reward signal is calculated based on the actual turnaround time. The reward signal is negatively correlated with the actual turnaround time. Based on the reward signal and the state value estimate output by the value assessment network, the generalized advantage estimate is determined. Based on the generalized dominance estimate, the parameters in the linear projection layer and the large language model are updated using the proximal policy optimization algorithm.
[0088] It should be noted that the actual turnaround time is the time interval from when the task to be scheduled enters the ready queue to when it is completed, including waiting time, data transmission time, and computation execution time. The reward signal is an environmental feedback signal in reinforcement learning, used to evaluate the quality of the scheduling decision.
[0089] It should be understood that the reward signal is negatively correlated with the actual turnaround time; that is, the shorter the actual turnaround time, the larger the reward signal. This negative correlation is achieved through a reciprocal transformation or a negative linear transformation, ensuring that the reward signal maintains a moderate gradient across the entire value range and avoiding the gradient vanishing problem caused by large fluctuations in the actual turnaround time.
[0090] Specifically, as an optional implementation, the scheduling system records the timestamp of the task to be scheduled entering the ready queue and the timestamp of the task to be scheduled completing execution, and calculates the time difference between the two timestamps as the actual turnaround time; the actual turnaround time is normalized to generate a normalized turnaround time; and the reciprocal of the normalized turnaround time is calculated as the reward signal. The reward signal ideally approaches 1. When network congestion or excessive device load leads to an extended turnaround time, the reward signal decreases accordingly, thereby guiding the optimization direction of the scheduling strategy to align with the actual goal of shortening task turnaround time.
[0091] It should be noted that the generalized advantage estimate is a statistic used in reinforcement learning to evaluate the superiority or inferiority of the current scheduling decision relative to the average level. It is calculated based on the temporal difference error between the actual reward signal and the state value estimate predicted by the value assessment network. The value assessment network is a shallow neural network that bypasses the large language model mentioned above. Its input is a graph state vector, and its output is the state value estimate corresponding to that graph state vector. This estimate represents the expected cumulative reward that can be obtained under the current scheduling strategy in a certain system state.
[0092] Specifically, as an optional implementation, the scheduling device executes a scheduling decision and receives a reward signal. Then, the graph state vector of the current system state is... Input the value assessment network to obtain the value estimate of the current state. Simultaneously, the graph state vector of the next system state after execution will be generated. Input the value evaluation network to obtain the value estimate of the next state. Timing difference error Where γ is the discount factor. The generalized advantage estimate is obtained by exponentially weighting the time-series difference errors over multiple time steps, as follows:
[0093] in, These are GAE parameters used to control the trade-off between bias and variance. Generalized dominance estimate. The positive or negative sign of the value indicates the quality of the current scheduling decision relative to the average policy. A positive value indicates that the decision is better than the average level, and the policy parameters should be updated to reinforce that decision; a negative value indicates that the decision is worse than the average level, and the policy parameters should be adjusted in the opposite direction.
[0094] In this embodiment, the system employs an online reinforcement learning method for training, specifically using a near-end policy optimization algorithm. The scheduling policy continuously interacts with the cluster environment, receiving a reward in the form of the reciprocal of the normalized turnaround time after each scheduling decision. The technical consideration for using the reciprocal form of the reward is that turnaround time may fluctuate significantly in the early training phase (especially during network congestion), and the reciprocal form maintains a moderate gradient across the entire value range, avoiding the problem of the gradient tending to zero at large values in exponential rewards. During training, an advantage function is calculated based on the generalized advantage estimation method to evaluate the merits of each decision relative to the average level. The policy parameters are then updated by pruning the target based on the advantage function, while a policy entropy regularization term is added to encourage exploration. Each update only involves the parameters of the projection adaptation layer and the low-rank adaptation matrix parameters in the inference core; the pre-trained weights of the inference core remain frozen.
[0095] In this embodiment, the graph state vector is converted into a prefix token sequence through a linear projection layer, achieving an effective connection between the structured encoding output of the graph neural network and the large language model. After reasoning based on the prefix token sequence, the large language model outputs the resource node selection probability distribution and the computational precision selection probability distribution by the first and second decision heads, respectively. By complementing the structured encoding capability of the graph neural network with the complex constraint reasoning capability of the inference model, the scheduling decision can simultaneously consider the structural characteristics of the network topology and the implicit constraints between multiple factors such as computational workload, data location, and precision requirements.
[0096] For example, to help understand the implementation flow of the computing resource scheduling method obtained by combining this embodiment with the above embodiment one, please refer to... Figure 5 , Figure 5This is a schematic diagram of the overall architecture of a GLS scheduling system. Specifically, the system input side includes an HPC / AI workflow pool, which stores various high-performance computing or artificial intelligence workflows submitted by users. These workflows are decoded and processed by an encoder, parsed into atomic tasks, and stored in a task pool. The system simultaneously reads the current system state graph from the cluster environment, which includes the resource status of computing devices and real-time network link status information. The encoder fuses the attributes of the tasks to be scheduled in the task pool with the system state graph, generating a graph state vector, which is then input into an embedding encoding module. This vector is converted into a prefix token sequence, and then the LoRa adaptation layer performs task-specific parameter adaptation on the large language model. The adapted inference results are sent to two parallel branches—the resource head and the precision head—via the output head. The LoRa adaptation layer includes a low-rank matrix and a large language model (Qwen 3 8B). The resource head branch outputs the probability distribution of each candidate resource node being selected. After filtering out unavailable or non-compliant candidate resource nodes, it is re-normalized through the SoftMax layer and finally outputs the resource-dimensional action after sampling. The precision head branch similarly outputs the probability distribution of each candidate's computational precision and outputs the precision-dimensional action after sampling. The two parts of the action are combined to form the final scheduling action and are sent to the cluster environment for execution. After the task is completed, the environment calculates a reward signal based on the actual turnaround time. This reward signal is fed back to the critic network to calculate the advantage function and update the trainable parameters in the LoRa adaptation layer and the linear projection layer, forming a closed-loop learning architecture of "perception-decision-execution-feedback-evolution".
[0097] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the computing resource scheduling method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.
[0098] This application also provides a computing resource scheduling device; please refer to... Figure 6 The computing resource scheduling device includes: The acquisition module 10 is used to acquire the tasks to be scheduled in the ready queue. All the predecessor tasks of the tasks to be scheduled have been completed. The attributes of the tasks to be scheduled include the location information of the output data of the predecessor tasks. The construction module 20 is used to construct a system state diagram based on the current state of the task to be scheduled and the computing cluster; Output module 30 is used to input the system state diagram into a graph neural network and output a graph state vector; The decision module 40 is used to project the graph state vector into a prefix token sequence, input the prefix token sequence into the large language model, and output a scheduling decision to allocate the task to be scheduled to the corresponding target resource node.
[0099] The computing resource scheduling device provided in this application, employing the computing resource scheduling method in the above embodiments, can solve the technical problem of low overall efficiency caused by selecting task execution devices based on the remaining computing capacity of the devices without considering the real-time load of the links. Compared with the prior art, the beneficial effects of the computing resource scheduling device provided in this application are the same as those of the computing resource scheduling method provided in the above embodiments, and other technical features in the computing resource scheduling device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.
[0100] This application provides a computing resource scheduling device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the computing resource scheduling method in the first embodiment described above.
[0101] The following is for reference. Figure 7 This document illustrates a structural diagram of a computing resource scheduling device suitable for implementing embodiments of this application. The computing resource scheduling device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The computing resource scheduling device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.
[0102] like Figure 7As shown, the computing resource scheduling device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the computing resource scheduling device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the computing resource scheduling device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows computing resource scheduling devices with various systems, it should be understood that it is not required to implement or possess all of the systems shown. More or fewer systems may be implemented alternatively.
[0103] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.
[0104] The computing resource scheduling device provided in this application, employing the computing resource scheduling method in the above embodiments, can solve the technical problem of low overall efficiency caused by existing methods that select task execution devices based on the remaining computing capacity of the devices without considering the real-time load of the links. Compared with the prior art, the beneficial effects of the computing resource scheduling device provided in this application are the same as those of the computing resource scheduling method provided in the above embodiments, and other technical features in this computing resource scheduling device are the same as those disclosed in the previous embodiment method, and will not be repeated here.
[0105] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.
[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0107] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the computing resource scheduling method in the above embodiments.
[0108] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.
[0109] The aforementioned computer-readable storage medium may be included in a computing resource scheduling device; or it may exist independently and not be assembled into a computing resource scheduling device.
[0110] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by a computing resource scheduling device, cause the computing resource scheduling device to perform the computing resource scheduling method described above.
[0111] Computer program code for performing the operations of this application can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, and C++, and conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0112] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0113] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.
[0114] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described computing resource scheduling method. This solves the technical problem of low overall efficiency caused by prior art where task execution devices are selected based on remaining computing capacity without considering real-time link load. Compared to the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the computing resource scheduling method provided in the above embodiments, and will not be repeated here.
[0115] The above description is only a part of the embodiments of this application and does not limit the scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included within the protection scope of this application.
[0116] It should be noted that the data collection, tag management, rule setting, and push decision-making processes involved in this application are designed to work with other technical features to solve technical problems. They do not involve or support any illegal activities. Any data processing that may violate laws and regulations (such as unauthorized collection of privacy data, generation of discriminatory tags, setting unfair rules, or pushing illegal information) is not within the scope of protection of this application's technical solution. Of course, the user data in this application will be encrypted, anonymized, or de-identified before storage to ensure user data security.
Claims
1. A method for scheduling computing resources, characterized in that, The computing resource scheduling method includes: Get the tasks to be scheduled in the ready queue. All predecessor tasks of the tasks to be scheduled have been completed. The attributes of the tasks to be scheduled include the location information of the output data of the predecessor tasks. Based on the current state of the tasks to be scheduled and the computing cluster, a system state diagram is constructed; The system state diagram is input into a graph neural network, which outputs a graph state vector. The graph state vector is projected into a prefix token sequence, and the prefix token sequence is input into a large language model to output a scheduling decision, so as to allocate the task to be scheduled to the corresponding target resource node.
2. The computing resource scheduling method as described in claim 1, characterized in that, Before obtaining the tasks to be scheduled from the ready queue, the process also includes: Obtain a directed acyclic graph workflow from user input, wherein the directed acyclic graph workflow includes multiple computation tasks and the data dependencies between the computation tasks; Traverse the directed acyclic graph workflow. When all predecessor tasks of the computation task have been executed, decouple the computation task from the directed acyclic workflow to obtain the task to be scheduled. The location information of the output data of the predecessor task is encapsulated as an attribute of the task to be scheduled, and the task to be scheduled is added to the ready queue.
3. The computing resource scheduling method as described in claim 1, characterized in that, The system state diagram includes task nodes and resource nodes. The construction of the system state diagram based on the current state of the tasks to be scheduled and the computing cluster includes: The task to be scheduled is taken as a task node, and the task attributes of the task to be scheduled and the location information of the output data of the predecessor task are encoded as the first attribute of the task node. Each computing device in the computing cluster is designated as a resource node, and the current state and device type of each computing device are encoded as the second attribute of the resource node. Based on the location information, a first type of edge is established between the task node and the resource node; Based on the physical connection relationship of the computing devices, a second type of edge is established between the resource nodes. The attributes of the second type of edge include the link bandwidth and real-time utilization of the corresponding link. A system state diagram is constructed based on the task nodes, resource nodes, the first type of edge, and the second type of edge.
4. The computing resource scheduling method as described in claim 1, characterized in that, The step of inputting the system state diagram into a graph neural network and outputting a graph state vector includes: The system state diagram is input into a graph neural network to perform multi-layer message passing, so that the task node and the resource node update information through edges; After message transmission is completed, the information of all nodes in the system state graph is aggregated through global pooling to generate graph feature vectors.
5. The computing resource scheduling method as described in claim 1, characterized in that, The step of projecting the graph state vector into a prefix token sequence, inputting the prefix token sequence into a large language model, and outputting a scheduling decision to allocate the task to be scheduled to the corresponding target resource node includes: The graph state vector is converted into a prefix token sequence through a linear projection layer; Input the prefix token sequence into the large language model and output the hidden state representation; The hidden state representation is input into the first decision head, and the output is a first probability distribution representing the selection of each candidate resource node. The hidden state representation is input into the second decision head, and the output is a second probability distribution representing the selection of each candidate computational precision. A scheduling decision is determined based on the first probability distribution and the second probability distribution to allocate the task to be scheduled to the corresponding target resource node. The scheduling decision includes the target resource node and the target computational precision.
6. The computing resource scheduling method as described in claim 5, characterized in that, After converting the graph state vector into a prefix token sequence through a linear projection layer, the method further includes: The graph state vector is read through a value assessment network; The graph state vector is processed by a shallow network in the value assessment network to output a state value estimate.
7. The computing resource scheduling method as described in claim 5, characterized in that, After determining the scheduling decision based on the first probability distribution and the second probability distribution to allocate the task to be scheduled to the corresponding target resource node, the method further includes: The actual turnaround time of the task to be scheduled is obtained, and a reward signal is calculated based on the actual turnaround time. The reward signal is negatively correlated with the actual turnaround time. Based on the reward signal and the state value estimate output by the value assessment network, the generalized advantage estimate is determined. Based on the generalized dominance estimate, the parameters in the linear projection layer and the large language model are updated using the proximal policy optimization algorithm.
8. A computing resource scheduling device, characterized in that, The device includes: The acquisition module is used to acquire the tasks to be scheduled in the ready queue. All predecessor tasks of the tasks to be scheduled have been completed. The attributes of the tasks to be scheduled include the location information of the output data of the predecessor tasks. The construction module is used to construct a system state diagram based on the current state of the task to be scheduled and the computing cluster; The output module is used to input the system state diagram into the graph neural network and output a graph state vector. The decision module is used to project the graph state vector into a prefix token sequence, input the prefix token sequence into the large language model, and output a scheduling decision to allocate the task to be scheduled to the corresponding target resource node.
9. A computing resource scheduling device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the computing resource scheduling method as described in any one of claims 1 to 7.
10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the computing resource scheduling method as described in any one of claims 1 to 7.
