Heterogeneous computing node scheduling method and device based on uniform resource model

Abstract

The application discloses a kind of heterogeneous computing node scheduling method and device based on uniform resource model, belong to computer resource scheduling technical field.The method includes: S1 uniform resource model construction, obtains heterogeneous computing node resource information and is mapped as standardization resource vector, introduces the equivalent of computing power as uniform dimension unit, generates uniform resource description;S2 intelligent matching scheduling, obtains computing task resource demand, calculates resource vector similarity and selects target computing node and allocates task, considers task completion delay, resource utilization and energy consumption balance and carries out multi-objective optimization, according to real-time load dynamic adjustment scheduling weight;S3 execution and feedback, monitor task execution state and node health state, according to real-time load data revision resource description, support computing task migration and dynamic rescheduling.The application also discloses corresponding scheduling device.The application eliminates resource island, improves heterogeneous resource utilization, reduces scheduling management complexity.

Description

Technical Field

[0001] This invention belongs to the field of computer resource scheduling technology, specifically referring to a method and apparatus for scheduling heterogeneous computing nodes based on a unified resource model. Background Technology

[0002] With the rapid development of emerging businesses such as 5G networks, artificial intelligence, and the industrial internet, computing tasks in edge computing and cloud computing scenarios are exhibiting highly diversified characteristics. Multiple heterogeneous computing nodes, such as CPUs, GPUs, FPGAs, and NPUs, are often deployed simultaneously within the same computing cluster, with significant differences in computing power type, memory bandwidth, power consumption characteristics, and applicable scenarios.

[0003] Traditional resource scheduling systems typically employ a "classified management, separate scheduling" model, maintaining independent resource pools and scheduling strategies for different types of computing nodes. This approach has the following prominent problems: (i) Severe resource silos. GPU nodes and CPU nodes are managed independently, making it impossible to uniformly perceive and allocate resources across different resource types. This results in some nodes being overloaded while others are idle, leading to low overall resource utilization.

[0004] (ii) Fragmented scheduling strategies. Each type of computing resource requires a separate scheduling algorithm, resulting in high system maintenance costs, poor strategy consistency, and difficulty in forming a unified scheduling optimization goal.

[0005] (iii) Low task matching efficiency. The resource requirements and node capabilities of heterogeneous tasks lack a unified expression framework. The scheduler needs to rely on a large number of hard-coded rules to match tasks and nodes, resulting in insufficient flexibility and scalability.

[0006] (iv) Weak elasticity and responsiveness. When faced with sudden traffic surges or changes in task patterns, the system struggles to quickly migrate and redistribute load among heterogeneous resources, leading to a decline in service quality.

[0007] To address the aforementioned issues, existing open-source scheduling systems (such as Kubernetes) still support heterogeneous resources through extension plugins, lacking native unified resource modeling capabilities and failing to fundamentally solve the problem of unified perception and scheduling of heterogeneous resources. Summary of the Invention

[0008] This invention aims to solve the technical problems of severe resource silos, fragmented scheduling strategies, low task matching efficiency, and weak elastic response capability in existing heterogeneous computing node scheduling technologies, and provides a heterogeneous computing node scheduling method and device based on a unified resource model.

[0009] To achieve the above functions, the technical solution adopted by this invention is as follows: a heterogeneous computing node scheduling method based on a unified resource model, comprising the following steps: S1 Unified Resource Model Construction: Obtain resource information of multiple heterogeneous computing nodes in a heterogeneous computing cluster, map the computing power, storage capacity, network bandwidth, and energy consumption characteristics of each heterogeneous computing node into a standardized resource vector, and generate a unified resource description for each heterogeneous computing node; introduce computing power equivalent as a unified unit of measurement, converting different types of computing power into comparable unified measurement values; support the expansion of new resource dimensions through resource description without modifying the core logic of the scheduling engine; S2 Intelligent Matching and Scheduling: Based on the unified resource descriptions of each heterogeneous computing node generated in the unified resource description step, it receives computing tasks to be scheduled and extracts the resource requirements of the computing tasks. It calculates the resource vector similarity between the unified resource descriptions of each heterogeneous computing node and the resource requirements of the computing tasks. Based on the resource vector similarity, it selects a target computing node from the multiple heterogeneous computing nodes and assigns the computing tasks to the target computing node. Simultaneously, it performs multi-objective optimization, taking into account task completion latency, resource utilization, and energy consumption balance indicators. The scheduling weights are dynamically adjusted based on the real-time load of the nodes. S3 Execution and Feedback: Monitor the execution status of the computing task on the target computing node, and correct the unified resource description generated in the unified resource description step based on the real-time load data obtained during the execution process; support the migration of the computing task and dynamic rescheduling; and continuously correct the resource model parameters based on the execution feedback.

[0010] As a preferred embodiment of the present invention, the resource vector includes a computing power dimension, a storage capacity dimension, a network bandwidth dimension, and an energy consumption characteristic dimension.

[0011] As a preferred technical solution of the present invention, the multiple heterogeneous computing nodes are matched and evaluated based on the calculation results of the resource vector similarity, and a target computing node is selected.

[0012] As a preferred embodiment of the present invention, it further includes: when adding a new computing node type, access is made by registering a resource description, without modifying the core logic of the scheduling engine.

[0013] As a preferred embodiment of the present invention, the execution status monitoring further includes monitoring the health status of heterogeneous computing nodes.

[0014] This invention also discloses a heterogeneous computing node scheduling device based on a resource model, including... A heterogeneous computing cluster, wherein the heterogeneous computing cluster comprises multiple heterogeneous computing nodes; A scheduler, communicatively connected to each of the heterogeneous computing nodes, comprises: A unified resource description unit is used to map the resource information of each heterogeneous computing node into a standardized format resource vector, generate a unified resource description, and extend the resource dimension based on the resource description. The intelligent matching and scheduling unit is used to calculate the resource vector similarity between the unified resource description generated by the unified resource description unit and the resource requirements of the computing task to be scheduled, and select the target computing node and allocate the task based on the resource vector similarity. The execution and feedback unit is used to monitor the task execution status, correct the unified resource description generated by the unified resource description unit according to real-time load data, and trigger task migration and dynamic rescheduling according to real-time load conditions.

[0015] As a preferred embodiment of the present invention, the scheduler is also used to: when adding a new computing node type, access it by registering a resource description, without modifying the core logic of the scheduling engine.

[0016] Compared with the prior art, the present invention achieves the following beneficial effects by adopting the above structure: Eliminate resource silos and improve resource utilization: By using a unified resource description layer, the resource information of all types of computing nodes is incorporated into the same expression framework, realizing unified perception and allocation across resource types, and improving the utilization of heterogeneous nodes; Accelerated scheduling response: Standardized matching replaces hard-coded rules, reducing scheduling decision latency; Reduced management complexity: A single model manages all types of nodes. When adding a new type of compute node, only the resource description needs to be registered, without modifying the scheduling engine, resulting in a significant reduction in operation and maintenance costs. Attached Figure Description

[0017] Figure 1 This is a flowchart of a heterogeneous computing node scheduling method based on a resource model proposed in this invention; Figure 2 This is a block diagram of a heterogeneous computing node scheduling device for a resource model proposed in this invention. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Scheduling Method like Figure 1As shown, the heterogeneous computing node scheduling method based on a unified resource model provided by this invention includes the following steps: S1 Unified Resource Model Construction In this step, the resource information of multiple heterogeneous computing nodes in the heterogeneous computing cluster is first obtained. The heterogeneous computing cluster simultaneously deploys various types of computing nodes, such as CPU nodes, GPU nodes, FPGA nodes, NPU nodes, etc.

[0020] For each computing node, its resource information is collected, including computing power information, storage capacity information, network bandwidth information, and energy consumption characteristics information of each node.

[0021] The above-mentioned multi-dimensional resource information is mapped into a standardized resource vector, which includes computing power dimension, storage capacity dimension, network bandwidth dimension and energy consumption characteristic dimension, so as to realize the unified description and comparability of resource information of different types of nodes.

[0022] Specifically, "computing power equivalent" is introduced as a unified unit of measurement. Different types of computing nodes have different performance characteristics and measurement standards. By using "computing power equivalent", the above-mentioned heterogeneous computing capabilities are converted into a unified measurement value that can be directly compared across nodes, and thus incorporated into the same measurement system for comparison.

[0023] In addition, the unified resource model supports the dynamic expansion of new resource dimensions. When a new type of computing node is added to a heterogeneous computing cluster or new resource consideration factors need to be included, the new dimension only needs to be included in the unified resource model without modifying the core logic of the scheduling engine.

[0024] S2 Intelligent Matching and Scheduling This step performs scheduling decisions based on the unified resource description generated by S1.

[0025] First, the system receives the computing task to be scheduled and extracts the resource requirement information of the computing task. The computing task may be an AI inference task, a video encoding and decoding task, a general data processing task, a signal processing task, etc. The types and quantities of heterogeneous resources required by various tasks are different. This invention expresses the resource requirements of the computing task as a resource requirement vector in the same format as the unified resource model.

[0026] Then, based on the unified resource description of each heterogeneous computing node generated by S1, the similarity between the unified resource vector of each node and the resource requirement vector of the current task is calculated using preset matching rules.

[0027] Based on the calculation results of resource vector similarity, multiple heterogeneous computing nodes are matched and evaluated. The heterogeneous computing node with the highest matching degree with the current computing task is selected as the target computing node, and the computing task is assigned to the target computing node for execution.

[0028] During the scheduling decision-making process, this invention also performs multi-objective optimization, taking into account three evaluation indicators: task completion delay, resource utilization rate, and energy consumption balance.

[0029] Meanwhile, this invention dynamically adjusts the scheduling weights based on the real-time load status of heterogeneous computing nodes. For nodes with high loads, the scheduling priority is automatically reduced to prevent new tasks from being assigned to overloaded nodes.

[0030] S3 Execution and Feedback This step continues to run after the task is assigned and executed, forming a closed-loop feedback mechanism.

[0031] When a computing task is executed on a target computing node, the present invention monitors the execution status of the computing task in real time, and at the same time monitors the health status of the target computing node.

[0032] Based on the real-time load data obtained during execution, the resource parameters of relevant nodes in the unified resource description generated in S1 are dynamically corrected so that the resource model can more accurately reflect the actual operating status of the nodes.

[0033] When an abnormal load or health status is detected on the target computing node, this invention supports migrating the current computing task to other suitable heterogeneous computing nodes and triggering a dynamic rescheduling process to ensure that the computing task can continue to execute and improve the stability of task execution.

[0034] Through continuous real-time feedback and parameter correction, the resource model parameters are continuously optimized, and the accuracy of subsequent scheduling decisions is gradually improved, forming a complete closed loop of model building → scheduling decision → execution feedback → model correction.

[0035] Example 2: Scheduling Device like Figure 2 As shown, the present invention also provides a heterogeneous computing node scheduling device based on a unified resource model, including a heterogeneous computing cluster and a scheduler.

[0036] The heterogeneous computing cluster contains multiple heterogeneous computing nodes, including but not limited to CPU nodes, GPU nodes, FPGA nodes, and NPU nodes. Each node provides different computing capabilities and is suitable for different types of computing tasks.

[0037] The scheduler is communicatively connected to each of the heterogeneous computing nodes and is used to uniformly manage the resources of all heterogeneous computing nodes in the cluster and perform task scheduling. The scheduler includes a unified resource description unit, an intelligent matching scheduling unit, and an execution and feedback unit.

[0038] It also includes a unified resource description unit: This unit is used to perform the unified resource model construction function in S1. Specifically, it obtains the resource information of each heterogeneous computing node, maps the computing power, storage capacity, network bandwidth and energy consumption characteristics of each node into a standardized resource vector, and generates a unified resource description corresponding to each heterogeneous computing node. This unit introduces computing power equivalent as a unified unit of measurement, and converts different types of computing power into comparable unified measurement values.

[0039] The unified resource description unit maps the computing capabilities of different types of computing nodes, such as CPU nodes, GPU nodes, FPGA nodes, and NPU nodes, into a resource vector format. The resource information of each node is stored in the resource pool in a standardized data structure for other units of the scheduler to access.

[0040] Meanwhile, the unified resource description unit supports extending resource dimensions through resource descriptions. When a new type of computing node is introduced into the cluster or a new resource consideration dimension needs to be added, the new resource description can be registered in the unified resource description unit, and the resource information of the new node can be mapped to the new dimension of the resource vector. It can then be incorporated into the unified resource model for scheduling and management without modifying the core logic of the scheduler.

[0041] Intelligent Matching and Scheduling Unit: This unit executes the intelligent matching and scheduling function in S2. Specifically, it receives the computing tasks to be scheduled and extracts the resource requirements of the tasks, generating a task resource requirement vector. Based on the unified resource descriptions of each heterogeneous computing node generated by the unified resource description unit, it calculates the similarity between the task resource requirement vector and the resource vectors of each node using preset matching rules. Based on the similarity calculation results, it selects the target computing node with the highest matching degree and assigns the computing task to that target computing node.

[0042] For example, when an AI inference task arrives, the intelligent matching and scheduling unit matches the task's requirements for GPU computing power, storage capacity, and bandwidth with the unified resource description of each node, selecting the GPU node or NPU node with the highest matching degree as the target node; when a general data processing task arrives, it selects a CPU node with sufficient CPU computing power resources as the target node. In multi-task concurrent scenarios, this unit performs multi-objective optimization, comprehensively balancing the latency requirements of each task with the overall resource utilization and energy consumption balance of the cluster.

[0043] In addition, the intelligent matching and scheduling unit dynamically adjusts the scheduling weight of each node according to the current load status of each heterogeneous computing node, and automatically reduces the probability of high-load nodes being selected.

[0044] Execution and Feedback Unit: This unit is used to perform the execution and feedback functions in S3. Specifically, during the execution of the computing task, it continuously monitors the execution status of the task and the health status of each heterogeneous computing node, obtains real-time load data, and corrects the resource description parameters of the corresponding node in the unified resource description unit based on the real-time load data, so that the resource model can more accurately reflect the actual operating status of the node.

[0045] The execution and feedback unit establishes a closed-loop feedback mechanism: the unified resource description unit provides basic data to the intelligent matching and scheduling unit, the intelligent matching and scheduling unit allocates tasks to target nodes, the execution and feedback unit monitors the execution process and feeds back real-time data to the unified resource description unit, the unified resource description unit corrects the resource model parameters accordingly, and the corrected model is then used for subsequent scheduling decisions. This closed loop continues to operate, continuously improving scheduling accuracy.

[0046] When an abnormal load or health status is detected on the target computing node, the execution and feedback unit triggers task migration and dynamic rescheduling to migrate the affected computing tasks to other available nodes for continued execution, ensuring that the computing tasks can continue to execute and improving the stability of task execution.

[0047] In one embodiment of the present invention, the scheduler also supports accessing new computing node types by registering resource descriptions, without modifying the core logic of the scheduling engine.

[0048] Example 3: Application Scenarios 5G edge computing (MEC) scenarios: The scheduling method and device are deployed at 5G MEC edge sites, which typically deploy CPU nodes, GPU nodes and FPGA nodes in a mixed manner, with various heterogeneous computing powers coexisting.

[0049] In this scenario, a unified resource description layer maps the resource information of three types of heterogeneous nodes into a unified resource vector expression. The intelligent matching and scheduling layer allocates different tasks to the corresponding node types based on the resource requirement type of each computing task and the similarity of resource vectors. The execution and feedback layer monitors the execution status of various tasks in real time and dynamically adjusts the scheduling weights according to the real-time load of edge nodes. Through the above mechanism, this invention can realize the unified scheduling of edge nodes and meet the business requirements of low latency and high reliability.

[0050] AI inference service scenarios: The scheduling method and device are deployed on the AI ​​inference service platform. AI inference tasks have large differences in their computing power requirements for GPU / NPU resources, and the business load fluctuates drastically over time.

[0051] Through this invention, the platform integrates the resource information of all GPU nodes and CPU nodes into a unified resource vector model for management. When an inference request arrives, the intelligent matching and scheduling layer selects a matching node to perform inference based on the task's requirements for computing power and other resource dimensions, and on the basis of resource vector similarity. Based on the unified resource model, the allocation and load balancing of inference tasks among different accelerator cards can be realized.

[0052] Cloud computing data center scenario The scheduling method and device are deployed in large-scale cloud computing data centers, which generally adopt a CPU+GPU+ASIC hybrid deployment architecture, with various computing resources deployed in different server clusters.

[0053] This invention integrates all heterogeneous computing resources within a data center into a single scheduling framework for management. When a sudden surge in traffic or a surge in a certain type of task occurs, the scheduler dynamically adjusts the scheduling weights based on the real-time load status, allocating some tasks to other types of nodes, thereby achieving dynamic load balancing and elastic response among heterogeneous resources within the cluster.

[0054] Industrial Internet scenarios The scheduling method and device are deployed on an industrial internet platform. Industrial scenarios involve a variety of computing tasks, including real-time control tasks, visual inspection tasks, and data analysis tasks. The need for unified scheduling of heterogeneous resources is becoming increasingly urgent. Through this invention, unified scheduling and dynamic allocation of various industrial computing tasks among heterogeneous computing resources can be achieved.

[0055] This invention can be widely applied to industries such as 5G edge computing (MEC), AI inference services, cloud computing data centers, and the industrial internet.

[0056] This invention can improve the utilization rate of heterogeneous resources, reduce the complexity of resource management, and improve the system's scalability.

[0057] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A heterogeneous computing node scheduling method based on a unified resource model, characterized in that: Includes the following steps: S1 Unified Resource Model Construction: Obtain resource information of multiple heterogeneous computing nodes in a heterogeneous computing cluster, map the computing power, storage capacity, network bandwidth and energy consumption characteristics of each heterogeneous computing node into a standardized resource vector, and generate a unified resource description for each heterogeneous computing node; introduce computing power equivalent as a unified unit of measurement, and convert different types of computing power into comparable unified measurement values. It supports extending new resource dimensions through resource descriptions without modifying the core logic of the scheduling engine; S2 Intelligent Matching and Scheduling: Based on the unified resource descriptions of each heterogeneous computing node generated in the unified resource description step, it receives computing tasks to be scheduled and extracts the resource requirements of the computing tasks. It calculates the resource vector similarity between the unified resource descriptions of each heterogeneous computing node and the resource requirements of the computing tasks. Based on the resource vector similarity, it selects a target computing node from the multiple heterogeneous computing nodes and assigns the computing tasks to the target computing node. Simultaneously, it performs multi-objective optimization, taking into account task completion latency, resource utilization, and energy consumption balance indicators. The scheduling weights are dynamically adjusted based on the real-time load of the nodes. S3 Execution and Feedback: Monitor the execution status of the computing task on the target computing node, and correct the unified resource description generated in the unified resource description step based on the real-time load data obtained during the execution process; support the migration of the computing task and dynamic rescheduling; and continuously correct the resource model parameters based on the execution feedback.

2. The heterogeneous computing node scheduling method based on a unified resource model according to claim 1, characterized in that: The resource vector includes dimensions of computing power, storage capacity, network bandwidth, and energy consumption characteristics.

3. The heterogeneous computing node scheduling method based on a unified resource model according to claim 2, characterized in that: Based on the calculation results of the resource vector similarity, the multiple heterogeneous computing nodes are matched and evaluated, and a target computing node is selected.

4. The heterogeneous computing node scheduling method based on a unified resource model according to claim 3, characterized in that: Also includes: When adding a new compute node type, it can be accessed by registering a resource description, without needing to modify the core logic of the scheduling engine.

5. The heterogeneous computing node scheduling method based on a unified resource model according to claim 4, characterized in that: The execution status monitoring also includes monitoring the health status of heterogeneous computing nodes.

6. The heterogeneous computing node scheduling device based on a unified resource model according to any one of claims 1 to 5, characterized in that: This includes a heterogeneous computing cluster, which contains multiple heterogeneous computing nodes; A scheduler, communicatively connected to each of the heterogeneous computing nodes, comprises: A unified resource description unit is used to map the resource information of each heterogeneous computing node into a standardized format resource vector, generate a unified resource description, and extend the resource dimension based on the resource description. The intelligent matching and scheduling unit is used to calculate the resource vector similarity between the unified resource description generated by the unified resource description unit and the resource requirements of the computing task to be scheduled, and select the target computing node and allocate the task based on the resource vector similarity. The execution and feedback unit is used to monitor the task execution status, correct the unified resource description generated by the unified resource description unit according to real-time load data, and trigger task migration and dynamic rescheduling according to real-time load conditions.

7. The heterogeneous computing node scheduling device based on a unified resource model according to claim 6, characterized in that: The scheduler is also used to: when adding a new computing node type, access it by registering a resource description, without needing to modify the core logic of the scheduling engine.

Images