Heterogeneous computing resource scheduling system of large model

Abstract

The invention provides a large-model heterogeneous computing resource scheduling system, which relates to the field of electric digital data processing, and comprises a task decomposition adaptation module, a heterogeneous resource management module, a task scheduling optimization module and a data communication coordination module, the task decomposition adaptation module is used for decomposing tasks of a large model and adapting the tasks to corresponding computing resources, the heterogeneous resource management module is used for managing and monitoring heterogeneous computing resources, and the task scheduling optimization module is used for providing an efficient task scheduling strategy. The data communication coordination module is used for carrying out data communication among different resource nodes; according to the system, the scheduling strategy can be adjusted when the specific type of resources are insufficient, and the task running fluency in a large model is improved.

Description

Technical Field

[0001] The present invention relates to the field of electronic digital data processing, and in particular to a large-model heterogeneous computing resource scheduling system. Background Art

[0002] With the development of artificial intelligence, big data, and high-performance computing, more and more large-scale computing tasks are placing increasing demands on computing resources, especially in scenarios such as deep learning model training, scientific computing, and video rendering. These tasks usually have high computational complexity and diverse resource requirements, and require full use of heterogeneous computing resources to achieve efficient processing. However, traditional resource scheduling systems mainly optimize for a single type of resource and are difficult to flexibly allocate between multiple resource types, resulting in low resource utilization and difficulty in fully utilizing computing performance.

[0003] Many resource scheduling systems have been developed. After a lot of searching and reference, it is found that the existing resource scheduling systems are such as the system disclosed in the publication number CN118193175B. These system methods generally include: layering heterogeneous tasks with time dependencies and calculating the deadlines of each layer, fine-grained calculation of the deadlines of each heterogeneous task, estimating the cost and reliability of the task using various resources, and allocating different resources to multiple heterogeneous tasks with the goal of balancing processing performance, cost overhead and reliability. However, when allocating resources, the system still allocates the resource type adapted to the task, and cannot optimize the processing when the adapted resources are insufficient to improve the smoothness of task execution. Summary of the invention

[0004] The purpose of the present invention is to propose a large-model heterogeneous computing resource scheduling system to address the existing deficiencies.

[0005] The present invention adopts the following technical solution: A large-model heterogeneous computing resource scheduling system includes a task decomposition adaptation module, a heterogeneous resource management module, a task scheduling optimization module and a data communication coordination module; The task decomposition and adaptation module is used to decompose the tasks of the large model and adapt the corresponding computing resources, the heterogeneous resource management module is used to manage and monitor heterogeneous computing resources, the task scheduling optimization module is used to provide an efficient task scheduling strategy, and the data communication coordination module is used to perform data communication between different resource nodes; The task decomposition and allocation module includes a task analysis unit, a task disassembly unit and a resource adaptation unit. The task analysis unit is used to analyze the demand characteristics of the large model task, the task disassembly unit is used to disassemble the task into subtasks, and the resource adaptation unit is used to match the subtask with a suitable computing resource type; The heterogeneous resource management module includes a resource monitoring unit, a resource registration unit and a resource scheduling unit, wherein the resource monitoring unit is used to obtain the usage status of computing resources in real time, the resource registration unit is used to register the registration information of resources, and the resource scheduling unit is used to send subtasks to corresponding computing resources; The task scheduling optimization module includes a basic strategy generation unit, a type optimization adjustment unit and a historical data analysis unit. The basic strategy generation unit generates a scheduling plan based on the adaptation information of the task. The type optimization adjustment unit is used to convert the adaptation resource type of the subtask to generate a scheduling plan. The historical data analysis unit is used to collect optimization effect information for analysis and processing; The data communication coordination module includes a task association management unit, a data identification processing unit and a directional transmission processing unit. The task association management unit is used to register and manage the association between the data generated by the tasks, the data identification processing unit is used to identify the data to be transmitted, and the directional transmission processing unit transmits the data to the corresponding resource node based on the identification information and the association information.

[0006] Furthermore, the basic strategy generation unit includes a policy judgment processor, a node selection processor and a solution output processor. The policy judgment processor is used to judge and select the basic strategy or optimization strategy to process the subtask. The node selection processor selects the resource node for executing the subtask based on the basic strategy. The solution output processor is used to output the scheduling solution to the resource scheduling unit.

[0007] Furthermore, the node selection processor calculates the matching index P of the resource node according to the following formula: ; in, is the utilization rate of the corresponding type of resources of the resource node, Ve is the number of idle resources of the corresponding type of resources of the resource node, and Vd is the number of resources required by the subtask for the corresponding type; The node selection processor selects a resource node with a maximum matching index as an execution node of the subtask.

[0008] Furthermore, the type optimization and adjustment unit includes a type conversion evaluation processor, a waiting comparison processor and an adjustment output processor. The type conversion evaluation processor is used to evaluate the execution performance of the subtask after the resource type conversion, the waiting comparison processor is used to compare the efficiency of waiting for the original resource type to be executed and the efficiency after the resource type is converted, and the adjustment output processor is used to output the scheduling plan after the type conversion to the resource scheduling unit.

[0009] Furthermore, the evaluation process of the type conversion evaluation processor includes the following steps: S1, select the resource type after conversion, called the target type; S2, select the resource node with the largest amount of target type resources, called the target node; S3. Calculate the performance index Q of the target node according to the following formula: ; in, To preserve the proportion, is the adaptation coefficient of the subtask to the target type resource, is the idle amount of target type resources in the target node, V 0 is the resource base, e is a natural constant; S4, repeating steps S1 to S3 until the performance index of the target node under all resource types is obtained; The waiting comparison processor calculates the comparison judgment value according to the following formula: : ; Among them, Tw is the time that the subtask waits for the adaptation resource node, and Tr is the running time of the subtask under the adaptation resource processing; When the comparison judgment value is greater than 0, the target node with the largest performance index is used as the node for executing the subtask, and the corresponding target type resource is used as the resource for processing the subtask; When the comparison judgment value is not greater than 0, the subtask is returned to the basic strategy generation unit and waits for the emergence of a node that can meet the adaptation resources of the subtask.

[0010] The beneficial effects achieved by the present invention are: This system adapts resources for subtasks. Under normal circumstances, it selects appropriate nodes to execute subtasks based on adapted resources. When adapted resources are insufficient, it analyzes the real-time resource status of each node, adjusts resource types, and makes full use of idle resources to run subtasks, reducing the occurrence of subtask blocking and improving the running smoothness of the entire model task.

[0011] To further understand the features and technical contents of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are only for reference and description and are not intended to limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the overall structural framework of the present invention; Figure 2 This is a schematic diagram of the task decomposition and allocation module of the present invention; Figure 3 This is a schematic diagram of the heterogeneous resource management module of the present invention; Figure 4 This is a schematic diagram of the task scheduling optimization module of the present invention; Figure 5 This is a schematic diagram of the data communication coordination module of the present invention; Figure 6 This is a comparison chart of the actual test results of the present invention. DETAILED DESCRIPTION

[0013] The following is an explanation of the embodiments of the present invention through specific embodiments. Those skilled in the art can understand the advantages and effects of the present invention from the contents disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depicted according to actual sizes. It is stated in advance. The following embodiments will further explain the relevant technical contents of the present invention in detail, but the disclosed contents are not intended to limit the scope of protection of the present invention.

[0014] Embodiment 1: This embodiment provides a large-scale heterogeneous computing resource scheduling system, combined with Figure 1 , including task decomposition adaptation module, heterogeneous resource management module, task scheduling optimization module and data communication coordination module; The task decomposition and adaptation module is used to decompose the tasks of the large model and adapt the corresponding computing resources, the heterogeneous resource management module is used to manage and monitor heterogeneous computing resources, the task scheduling optimization module is used to provide an efficient task scheduling strategy, and the data communication coordination module is used to perform data communication between different resource nodes; The task decomposition and allocation module includes a task analysis unit, a task disassembly unit and a resource adaptation unit. The task analysis unit is used to analyze the demand characteristics of the large model task, the task disassembly unit is used to disassemble the task into subtasks, and the resource adaptation unit is used to match the subtask with a suitable computing resource type; The heterogeneous resource management module includes a resource monitoring unit, a resource registration unit and a resource scheduling unit, wherein the resource monitoring unit is used to obtain the usage status of computing resources in real time, the resource registration unit is used to register the registration information of resources, and the resource scheduling unit is used to send subtasks to corresponding computing resources; The task scheduling optimization module includes a basic strategy generation unit, a type optimization adjustment unit and a historical data analysis unit. The basic strategy generation unit generates a scheduling plan based on the adaptation information of the task. The type optimization adjustment unit is used to convert the adaptation resource type of the subtask to generate a scheduling plan. The historical data analysis unit is used to collect optimization effect information for analysis and processing; The data communication coordination module includes a task association management unit, a data identification processing unit and a directional transmission processing unit. The task association management unit is used to register and manage the association between the data generated by the tasks, the data identification processing unit is used to identify the data to be transmitted, and the directional transmission processing unit transmits the data to the corresponding resource node based on the identification information and the association information.

[0015] The basic strategy generation unit includes a policy judgment processor, a node selection processor and a solution output processor. The policy judgment processor is used to judge and select a basic strategy or an optimization strategy to process a subtask. The node selection processor selects a resource node to execute the subtask based on the basic strategy. The solution output processor is used to output the scheduling solution to the resource scheduling unit.

[0016] The node selection processor calculates the matching index P of the resource node according to the following formula: ; in, is the utilization rate of the corresponding type of resources of the resource node, Ve is the number of idle resources of the corresponding type of resources of the resource node, and Vd is the number of resources required by the subtask for the corresponding type; The node selection processor selects a resource node with a maximum matching index as an execution node of the subtask.

[0017] The type optimization and adjustment unit includes a type conversion evaluation processor, a waiting comparison processor and an adjustment output processor. The type conversion evaluation processor is used to evaluate the execution performance of the subtask after the resource type conversion, the waiting comparison processor is used to compare the efficiency of waiting for the original resource type to be executed and the efficiency after the resource type is converted, and the adjustment output processor is used to output the scheduling plan after the type conversion to the resource scheduling unit.

[0018] The evaluation process of the type conversion evaluation processor includes the following steps: S1, select the resource type after conversion, called the target type; S2, select the resource node with the largest amount of target type resources, called the target node; S3. Calculate the performance index Q of the target node according to the following formula: ; in, To preserve the proportion, is the adaptation coefficient of the subtask to the target type resource, is the idle amount of target type resources in the target node, V 0 is the resource base, e is a natural constant; S4, repeating steps S1 to S3 until the performance index of the target node under all resource types is obtained; The waiting comparison processor calculates the comparison judgment value according to the following formula: : ; Among them, Tw is the time that the subtask waits for the adaptation resource node, and Tr is the running time of the subtask under the adaptation resource processing; When the comparison judgment value is greater than 0, the target node with the largest performance index is used as the node for executing the subtask, and the corresponding target type resource is used as the resource for processing the subtask; When the comparison judgment value is not greater than 0, the subtask is returned to the basic strategy generation unit and waits for the emergence of a node that can meet the adaptation resources of the subtask.

[0019] Embodiment 2: This embodiment includes all the contents of Embodiment 1, and provides a large-model heterogeneous computing resource scheduling system, including a task decomposition adaptation module, a heterogeneous resource management module, a task scheduling optimization module and a data communication coordination module; The task decomposition and adaptation module is used to decompose the tasks of the large model and adapt the corresponding computing resources, the heterogeneous resource management module is used to manage and monitor heterogeneous computing resources, the task scheduling optimization module is used to provide an efficient task scheduling strategy, and the data communication coordination module is used to perform data communication between different resource nodes; Combination Figure 2 The task decomposition and allocation module includes a task analysis unit, a task disassembly unit and a resource adaptation unit. The task analysis unit is used to analyze the demand characteristics of the large model task, the task disassembly unit is used to disassemble the task into subtasks, and the resource adaptation unit is used to match the subtask with a suitable computing resource type; Combination Figure 3 , the heterogeneous resource management module includes a resource monitoring unit, a resource registration unit and a resource scheduling unit, the resource monitoring unit is used to obtain the usage status of computing resources in real time, the resource registration unit is used to register the registration information of resources, and the resource scheduling unit is used to send subtasks to corresponding computing resources; Combination Figure 4 The task scheduling optimization module includes a basic strategy generation unit, a type optimization adjustment unit and a historical data analysis unit. The basic strategy generation unit generates a scheduling plan based on the adaptation information of the task. The type optimization adjustment unit is used to convert the adaptation resource type of the subtask to generate a scheduling plan. The historical data analysis unit is used to collect optimization effect information for analysis and processing; Combination Figure 5The data communication coordination module includes a task association management unit, a data identification processing unit and a directional transmission processing unit. The task association management unit is used to register and manage the association between the data generated by the task, the data identification processing unit is used to identify the data to be transmitted, and the directional transmission processing unit transmits the data to the corresponding resource node based on the identification information and the association information; The task analysis unit includes a description analysis processor, a characteristic analysis processor and a condition analysis processor, wherein the description analysis processor is used to parse the description information of the task, the characteristic analysis processor is used to parse the internal characteristics of the task, and the condition analysis processor is used to parse the dependency conditions of the task; The task disassembly unit includes a subtask generation processor, a task encoding processor and a task packaging processor. The subtask generation processor disassembles the task based on the parsing information to obtain the execution data of the subtask. The task encoding processor encodes the subtask based on the dependency relationship. The task packaging processor is used to package the execution data and the demand characteristics of the subtask to obtain the subtask data packet. The subtask data packet uses the subtask code as an identification mark. The subtask data packet includes execution data and annotation data. The execution data is used to execute the subtask, and the annotation data is used to describe the relevant information of the subtask. The resource adaptation unit includes a data packet management processor, a type matching processor and a demand assessment processor, wherein the data packet management processor is used to receive and manage subtask data packets, the type matching processor obtains the resource type of the subtask based on annotation information matching, and the demand assessment processor is used to assess the demand of the subtask for the corresponding resource type; The resource monitoring unit includes a status acquisition processor, a data aggregation processor and a query interface processor. The status acquisition processor is used to collect the usage status of all types of computing resources in the resource node, the data aggregation processor is used to summarize the collected data and store it, and the query interface processor is used to provide an interface for querying the computing resource status; The resource registration unit includes a resource registration processor, a resource attribute manager and a resource topology manager. The resource registration processor is used to receive and register a new node and record the communication address. The resource attribute manager is used to manage the resource type and resource quantity contained in each resource node. The resource topology manager is used to manage the network communication parameters between nodes. The resource scheduling unit includes a scheme receiving processor, a data acquisition processor and a directional scheduling processor, wherein the scheme receiving processor is used to receive the scheduling scheme of the subtask, the data acquisition processor is used to obtain the execution data of the subtask, and the directional scheduling processor is used to obtain the communication address of the resource node in the scheduling scheme and send the execution data to the resource node; The basic strategy generation unit includes a strategy judgment processor, a node selection processor and a solution output processor, wherein the strategy judgment processor is used to judge and select a basic strategy or an optimization strategy to process a subtask, the node selection processor selects a resource node for processing a subtask based on the basic strategy, and the solution output processor is used to output a scheduling solution to the resource scheduling unit; When the type and quantity of idle resources of the resource node meet the requirements of the subtask, the strategy judgment processor selects the basic strategy, otherwise selects the optimization strategy; The node selection processor calculates the matching index P of the resource node according to the following formula: ; in, is the utilization rate of the corresponding type of resources of the resource node, Ve is the number of idle resources of the corresponding type of resources of the resource node, and Vd is the number of resources required by the subtask for the corresponding type; The node selection processor selects the resource node with the largest matching index as the execution node of the subtask; The type optimization and adjustment unit includes a type conversion evaluation processor, a waiting comparison processor and an adjustment output processor, wherein the type conversion evaluation processor is used to evaluate the execution performance of the subtask after the resource type conversion, the waiting comparison processor is used to compare the efficiency of waiting for the original resource type to be executed and the efficiency after the resource type is converted, and the adjustment output processor is used to output the scheduling scheme after the type conversion to the resource scheduling unit; The evaluation process of the type conversion evaluation processor includes the following steps: S1, select the resource type after conversion, called the target type; S2, select the resource node with the largest amount of target type resources, called the target node; S3. Calculate the performance index Q of the target node according to the following formula: ; in, To preserve the proportion, is the adaptation coefficient of the subtask to the target type resource, is the idle amount of target type resources in the target node, V 0 is the resource base, e is a natural constant; S4, repeating steps S1 to S3 until the performance index of the target node under all resource types is obtained; The waiting comparison processor calculates the comparison judgment value according to the following formula: : ; Among them, Tw is the time that the subtask waits for the adaptation resource node, and Tr is the running time of the subtask under the adaptation resource processing; When the comparison judgment value is greater than 0, the target node with the largest performance index is used as the node for executing the subtask, and the corresponding target type resource is used as the resource for processing the subtask; When the comparison judgment value is not greater than 0, the subtask is returned to the basic strategy generation unit and waits for the emergence of a node that can meet the adaptation resources of the subtask; The historical data analysis unit continuously adjusts the adaptation coefficient by obtaining the actual case data of the subtasks running the adaptation resources. Carry out correction processing; The task association management unit includes a running task register, an association relationship register and an information editing processor, wherein the running task register is used to store the running subtask code, the association relationship register is used to store the data relationship between the running subtasks, and the information editing processor is used to receive the state change information of the subtask and modify and edit the stored data; Some of the code for this system is described below: class TaskDecompositionModule: """Task decomposition adaptation module""" def __init__(self): pass class TaskAnalyzer: """Task Analysis Unit: Analyze the characteristics of large model task requirements""" @staticmethod def analyze(task): print(f"Analyzing task: {task}") #Simulate return task requirements return {"type": "compute", "resources": ["CPU", "GPU"]} class TaskSplitter: """Task decomposition unit: decompose tasks to generate subtasks""" @staticmethod def split(task): print(f"Splitting task: {task}") #Simulate returning a subtask list return [{"subtask_id": 1}, {"subtask_id": 2}] class ResourceMatcher: """Resource adaptation unit: matching appropriate computing resources""" @staticmethod def match(subtask): print(f"Matching resources for subtask: {subtask}") #Simulate returning matching resources return {"resource": "GPU"} class HeterogeneousResourceManager: """Heterogeneous resource management module""" def __init__(self): self.resources = {} class ResourceMonitor: """Resource monitoring unit: obtain resource status in real time""" @staticmethod def monitor(): print("Monitoring resources...") # Simulate resource status return {"CPU": "busy", "GPU": "idle"} class ResourceRegistrar: """Resource registration unit: register resource registration information""" @staticmethod def register(resource): print(f"Registering resource: {resource}") class ResourceScheduler: """Resource scheduling unit: assign subtasks to computing resources""" @staticmethod def schedule(subtask, resource): print(f"Scheduling subtask {subtask} to resource{resource}") class TaskOptimizationModule: """Task scheduling optimization module""" def __init__(self): pass class BaseStrategyGenerator: """Basic strategy generation unit: Generates initial scheduling plan""" @staticmethod def generate(task_info): print(f"Generating base strategy for: {task_info}") #Simulation return scheduling plan return {"strategy": "basic"} class TypeOptimizer: """Type optimization and adjustment unit: convert resource adaptation type""" @staticmethod def optimize(strategy): print(f"Optimizing strategy: {strategy}") # Simulation Optimization return {"strategy": "optimized"} class HistoricalAnalyzer: """Historical data analysis unit: collect optimization results and analyze them""" @staticmethod def analyze(): print("Analyzing historical data...") def transmit(data, resource).

[0020] Now we use 10 large model tasks as samples to test the processing time in this system and the ordinary system respectively, and get Figure 6 The effect comparison chart is shown.

[0021] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.

Claims

1. A large-scale heterogeneous computing resource scheduling system, characterized in that: It includes task decomposition and adaptation module, heterogeneous resource management module, task scheduling optimization module and data communication coordination module; The task decomposition and adaptation module is used to decompose the tasks of the large model and adapt the corresponding computing resources, the heterogeneous resource management module is used to manage and monitor heterogeneous computing resources, the task scheduling optimization module is used to provide an efficient task scheduling strategy, and the data communication coordination module is used to perform data communication between different resource nodes; The task decomposition and allocation module includes a task analysis unit, a task disassembly unit and a resource adaptation unit. The task analysis unit is used to analyze the demand characteristics of the large model task, the task disassembly unit is used to disassemble the task into subtasks, and the resource adaptation unit is used to match the subtask with a suitable computing resource type; The heterogeneous resource management module includes a resource monitoring unit, a resource registration unit and a resource scheduling unit, wherein the resource monitoring unit is used to obtain the usage status of computing resources in real time, the resource registration unit is used to register the registration information of resources, and the resource scheduling unit is used to send subtasks to corresponding computing resources; The task scheduling optimization module includes a basic strategy generation unit, a type optimization adjustment unit and a historical data analysis unit. The basic strategy generation unit generates a scheduling plan based on the adaptation information of the task. The type optimization adjustment unit is used to convert the adaptation resource type of the subtask to generate a scheduling plan. The historical data analysis unit is used to collect optimization effect information for analysis and processing; The data communication coordination module includes a task association management unit, a data identification processing unit and a directional transmission processing unit. The task association management unit is used to register and manage the association between the data generated by the tasks, the data identification processing unit is used to identify the data to be transmitted, and the directional transmission processing unit transmits the data to the corresponding resource node based on the identification information and the association information.

2. A large-model heterogeneous computing resource scheduling system as claimed in claim 1, characterized in that: The basic strategy generation unit includes a policy judgment processor, a node selection processor and a solution output processor. The policy judgment processor is used to judge and select a basic strategy or an optimization strategy to process a subtask. The node selection processor selects a resource node to execute the subtask based on the basic strategy. The solution output processor is used to output the scheduling solution to the resource scheduling unit.

3. A large-model heterogeneous computing resource scheduling system as claimed in claim 2, characterized in that: The node selection processor calculates the matching index P of the resource node according to the following formula: ； in, is the utilization rate of the corresponding type of resources of the resource node, Ve is the number of idle resources of the corresponding type of resources of the resource node, and Vd is the number of resources required by the subtask for the corresponding type; The node selection processor selects a resource node with a maximum matching index as an execution node of the subtask.

4. A large-model heterogeneous computing resource scheduling system as claimed in claim 3, characterized in that: The type optimization and adjustment unit includes a type conversion evaluation processor, a waiting comparison processor and an adjustment output processor. The type conversion evaluation processor is used to evaluate the execution performance of the subtask after the resource type conversion, the waiting comparison processor is used to compare the efficiency of waiting for the original resource type to be executed and the efficiency after the resource type is converted, and the adjustment output processor is used to output the scheduling plan after the type conversion to the resource scheduling unit.

5. A large-model heterogeneous computing resource scheduling system as claimed in claim 4, characterized in that: The evaluation process of the type conversion evaluation processor includes the following steps: S1, select the resource type after conversion, called the target type; S2, select the resource node with the largest amount of target type resources, called the target node; S3. Calculate the performance index Q of the target node according to the following formula: ； in, To preserve the proportion, is the adaptation coefficient of the subtask to the target type resource, is the idle amount of target type resources in the target node, V0 is the resource base, and e is a natural constant; S4, repeating steps S1 to S3 until the performance index of the target node under all resource types is obtained; The waiting comparison processor calculates the comparison judgment value according to the following formula: : ； Among them, Tw is the time that the subtask waits for the adaptation resource node, and Tr is the running time of the subtask under the adaptation resource processing; When the comparison judgment value is greater than 0, the target node with the largest performance index is used as the node for executing the subtask, and the corresponding target type resource is used as the resource for processing the subtask; When the comparison judgment value is not greater than 0, the subtask is returned to the basic strategy generation unit and waits for the emergence of a node that can meet the adaptation resources of the subtask.

Images