Cross-cluster scheduling method and system for heterogeneous computing power resources

Abstract

The invention provides a cross-cluster scheduling method and system for heterogeneous computing power resources, and the method comprises the steps: carrying out the unified modeling of different types of hardware resources in a plurality of heterogeneous computing power clusters according to a preset resource description specification, and generating a global unified virtual resource pool model; collecting dynamic state information of internal resources of the heterogeneous computing power clusters from the heterogeneous computing power clusters in real time, and performing mapping and updating according to the virtual resource pool model; receiving a user job, analyzing a resource demand specification of the user job, and retrieving an available resource set meeting the resource demand specification among a plurality of clusters based on the updated virtual resource pool model; and selecting a target resource from the available resource set according to a preset scheduling strategy, converting the user job into a job instruction which can be identified by a target cluster, and distributing the job instruction to a corresponding cluster for execution. According to the invention, the overall utilization rate and scheduling efficiency of computing power resources can be improved.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a cross-cluster scheduling method and system for heterogeneous computing resources. Background Technology

[0002] With the rapid development of fields such as artificial intelligence and scientific computing, the demand for computing power has increased dramatically, giving rise to various heterogeneous computing clusters such as supercomputing centers and intelligent computing centers based on different hardware architectures. However, these clusters are often built independently and operate autonomously, exhibiting significant differences in hardware type, system software, and scheduling strategies, forming "computing silos." Existing scheduling technologies are mainly geared towards single clusters or limited types of resources, lacking the ability to provide a unified description and global view of heterogeneous computing power across regions and technology systems, thus failing to achieve effective resource integration and collaborative scheduling.

[0003] Currently, although some research attempts to connect multiple computing centers through cloud platforms or customized interfaces, they still generally face key challenges: First, there is a lack of a unified resource abstraction model, and the status data formats of each cluster are inconsistent, making it difficult to standardize aggregation and real-time synchronization; second, cross-cluster scheduling strategies often ignore multi-dimensional factors such as dynamic resource changes, network overhead, and job costs, resulting in insufficient intelligence in decision-making; third, job submission interfaces are incompatible, requiring users to adapt to different clusters, making usage complex. These problems lead to low utilization of computing resources, making it difficult to support the "one-stop" intelligent scheduling needs for diverse computing tasks. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-cluster scheduling method and system for heterogeneous computing resources, aiming to solve at least one of the problems in the background art.

[0005] In a first aspect, the present invention provides a cross-cluster scheduling method for heterogeneous computing resources, the method comprising: Based on the preset resource description specifications, different types of hardware resources in multiple heterogeneous computing power clusters are uniformly modeled to generate a globally unified virtual resource pool model. The dynamic status information of the internal resources of each heterogeneous computing power cluster is collected in real time and mapped and updated according to the virtual resource pool model. Receive user jobs, parse their resource requirement specifications, and based on the updated virtual resource pool model, retrieve a set of available resources that meet the resource requirement specifications across multiple clusters; From the set of available resources, a target resource is selected according to a preset scheduling strategy, and the user job is converted into a job instruction that can be recognized by the target cluster and distributed to the corresponding cluster for execution.

[0006] In some embodiments, the step of uniformly modeling different types of hardware resources in multiple heterogeneous computing power clusters according to a preset resource description specification to generate a globally unified virtual resource pool model includes: Different architectures of CPU, GPU, and FPGA hardware resources are abstracted into computing unit objects with unified attribute fields, which include at least architecture type, computing power value, memory, and interconnect bandwidth.

[0007] In some embodiments, the step of mapping and updating according to the virtual resource pool model includes: By deploying adaptation agents in each cluster, the local resource monitoring interface of each cluster is actively polled or listened to at a preset period to dynamically collect its original status data. Based on the standardized attribute set defined by the virtual resource pool model, the raw state data collected in real time is dynamically mapped into standard fields and values.

[0008] In some embodiments, the step of retrieving a set of available resources that meet the resource requirement specifications across multiple clusters based on the updated virtual resource pool model includes: The resource requirements of the user job are matched and compared with the resource status data recorded in the resource status database, which has been formatted according to the unified attribute fields, to filter out all computing units and their respective clusters that meet the requirements in the current state.

[0009] In some embodiments, the step of selecting a target resource from the available resource set according to a preset scheduling strategy includes: From the set of available resources, the comprehensive cost of each candidate resource is calculated according to the optimization objective function. The comprehensive cost is calculated by weighting the execution cost and the expected queuing time, and the candidate resource with the lowest comprehensive cost is selected as the target resource.

[0010] In some embodiments, the step of converting the user job into job instructions recognizable by the target cluster includes: Based on the attributes of the target resource and the type of its cluster, the corresponding job template and adapter are invoked to convert the general description of the user job into a specific job script that conforms to the syntax of the target cluster's local job scheduling system.

[0011] In some embodiments, the method further includes: Define and deploy unified resource discovery and status reporting interfaces for the various heterogeneous hardware resources and local scheduling systems to be connected.

[0012] Secondly, the present invention provides a cross-cluster intelligent scheduling system for heterogeneous computing resources, the system comprising: The virtual model building module is used to uniformly model different types of hardware resources in multiple heterogeneous computing power clusters according to the preset resource description specifications, and generate a globally unified virtual resource pool model. The data mapping module is used to collect dynamic status information of the internal resources of each heterogeneous computing power cluster in real time, and to map and update the data according to the virtual resource pool model. The resource parsing module is used to receive user jobs, parse their resource requirement specifications, and, based on the updated virtual resource pool model, retrieve a set of available resources that meet the resource requirement specifications across multiple clusters. The resource scheduling module is used to select target resources from the available resource set according to a preset scheduling strategy, and convert the user job into job instructions that can be recognized by the target cluster, and distribute them to the corresponding cluster for execution.

[0013] Thirdly, the present invention provides a storage medium that stores one or more programs, which, when executed by a processor, implement the above-described method for cross-cluster scheduling of heterogeneous computing resources.

[0014] Fourthly, the present invention provides an electronic device, the electronic device comprising a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the above-described method for cross-cluster scheduling of heterogeneous computing resources.

[0015] Compared with the prior art, the present invention has the following advantages: This invention effectively solves the technical challenge of resource integration and scheduling across heterogeneous computing clusters by constructing a unified virtual resource pool model and implementing dynamic intelligent scheduling. Specifically, firstly, by establishing a unified resource description specification, hardware such as CPUs and GPUs with different architectures are abstracted into standardized computing units, solving the problem of difficult unified management caused by resource heterogeneity; secondly, by deploying an adaptation agent to collect the dynamic status of each cluster in real time and map it to standard data, accurate synchronization of the global resource view is achieved; based on this, the system can intelligently match the optimal resources according to job requirements and make comprehensive decisions based on multiple factors such as execution cost and queuing time; finally, by automatically converting general tasks into instructions recognizable by the target cluster through job templates, seamless distribution of cross-cluster jobs is achieved. This solution fundamentally breaks down "computing silos" and significantly improves the overall utilization rate and scheduling efficiency of computing resources. Attached Figure Description

[0016] Figure 1 This is a flowchart of a cross-cluster scheduling method for heterogeneous computing resources proposed in an embodiment of the present invention; Figure 2This is a schematic diagram of the structure of a cross-cluster intelligent scheduling system for heterogeneous computing resources proposed in an embodiment of the present invention.

[0017] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0019] like Figure 1 As shown, an embodiment of the present invention proposes a cross-cluster scheduling method for heterogeneous computing resources, the method comprising steps S101 to S104, wherein: Step S101: Based on the preset resource description specifications, perform unified modeling of different types of hardware resources in multiple heterogeneous computing power clusters to generate a globally unified virtual resource pool model; It should be noted that due to significant differences in hardware architecture, operating systems, and management interfaces among heterogeneous computing clusters, the resources of each cluster will exist in an "information silo" state without a unified description standard, making it impossible to form a global understanding. This step constructs a virtual resource pool model based on standardized attribute fields, abstracting various heterogeneous hardware such as CPUs and GPUs into unified logical resource objects. This establishes a common semantic foundation for cross-cluster resource discovery, description, and comparison, which is a prerequisite for realizing large-scale computing resource pooling integration and unified scheduling.

[0020] Specifically, CPU, GPU, and FPGA hardware resources with different architectures are abstracted into computing unit objects with unified attribute fields, which at least include architecture type, computing power value, memory, and interconnect bandwidth.

[0021] By abstracting and processing hardware resources of different architectures such as CPUs, GPUs, and FPGAs, they are transformed into computing unit objects. These objects have unified attribute fields, where the architecture type clarifies the hardware's architectural characteristics, such as x86 or ARM; the computing power value measures the hardware's computing capabilities; the memory attribute reflects the hardware's memory capacity; and the interconnect bandwidth reflects the data transfer capabilities between hardware components. Through these unified attribute fields, various heterogeneous hardware resources are presented in a unified form, facilitating subsequent processing.

[0022] Furthermore, prior to this step, to enable the scheduling system to access various heterogeneous hardware resources and local scheduling systems, a unified interface specification needs to be defined. Specifically, a unified resource discovery interface and a unified status reporting interface need to be defined and deployed for the various heterogeneous hardware resources and local scheduling systems to be accessed. The unified resource discovery interface allows the scheduling system to promptly discover newly accessed hardware resources, while the unified status reporting interface ensures that hardware resources report their status information to the scheduling system in a timely and accurate manner. By deploying these unified interfaces, the system's compatibility and scalability can be improved, and the difficulty and cost of accessing new resources can be reduced.

[0023] Step S102: Collect dynamic status information of internal resources from each heterogeneous computing power cluster in real time, and map and update it according to the virtual resource pool model; In this step, the status of computing resources (such as load, queues, and availability) changes dynamically over time. Static or outdated resource views will lead to scheduling decision errors and resource allocation conflicts. This step deploys a lightweight adaptation agent to collect native state data from each cluster in real time using polling or event-driven methods. This data is then standardized and updated in real time based on a unified model, ensuring that the global resource view used by the scheduling system is accurate, consistent, and up-to-date. This provides the information guarantee for efficient and fair scheduling.

[0024] Specifically, firstly, by deploying adaptation agents in each cluster, the local resource monitoring interface of each cluster is actively polled or listened to at a preset period to dynamically collect its original state data; then, based on the standardized attribute set defined by the virtual resource pool model, the original state data collected in real time is dynamically mapped into standard fields and values.

[0025] Adaptive agents are deployed in each cluster. These agents actively poll or listen to the local resource monitoring interfaces of each cluster at preset intervals to dynamically collect the raw state data of the cluster. Then, based on the standardized attribute set defined by the virtual resource pool model, the collected raw state data is converted into standard fields and values, realizing standardized data processing and ensuring that the data can be accurately populated into the virtual resource pool model to complete the model update.

[0026] Step S103: Receive user jobs, parse their resource requirement specifications, and based on the updated virtual resource pool model, retrieve a set of available resources that meet the resource requirement specifications across multiple clusters; Because user jobs have specific requirements for computing power type, scale, and performance, simple random or round-robin allocation can lead to a mismatch between resources and tasks, resulting in low computational efficiency or job failure. By transforming the abstract requirements of user jobs into quantifiable matching conditions and quickly retrieving all candidate resources that meet the conditions globally to form a set of available resources, a crucial leap from resource awareness to task matching is achieved.

[0027] After updating the virtual resource pool model, to allocate appropriate resources to user jobs, it is necessary to search for a set of available resources that meet the resource requirements across multiple clusters. Only by accurately selecting resources that meet the requirements can we ensure the smooth execution of user jobs, improve resource utilization, and increase user satisfaction. By matching and comparing the user job resource requirements with the data in the resource status database, available resources can be found quickly and accurately. Based on this, the resource requirements of the user job are first matched and compared with the resource status data recorded in the resource status database, which has been formatted according to the unified attribute fields, to filter out all computing units and their respective clusters that meet the requirements in the current state.

[0028] By matching the resource requirements specified by user jobs with the resource status data recorded in the resource status database, which has been formatted according to uniform attribute fields, and by comparing each attribute one by one, all computing units whose current status meets the user's requirements are selected. At the same time, the clusters to which these computing units belong are identified, forming a set of available resources, which provides a basis for subsequent resource selection.

[0029] Step S104: Select a target resource from the available resource set according to a preset scheduling strategy, and convert the user job into a job instruction that the target cluster can recognize, and distribute it to the corresponding cluster for execution.

[0030] It's important to note that simply finding a set of available resources is insufficient for optimal scheduling; a final selection must be made based on strategies such as cost, efficiency, and fairness. Furthermore, heterogeneous clusters have different job submission interfaces, and direct distribution would render the job unexecutable. This step first intelligently selects the optimal target from candidate resources using a pre-defined optimization objective function. Then, through job templates and adapter technology, the general job description is automatically "translated" into specific instructions recognizable by the target cluster's local scheduling system. This ensures the global optimization of the scheduling results.

[0031] In some embodiments, the comprehensive cost of each candidate resource is first calculated from the available resource set according to an optimization objective function. This comprehensive cost is calculated by weighting execution cost and expected queuing time. The candidate resource with the lowest comprehensive cost is then selected as the target resource. Selecting a target resource from the available resource set requires considering multiple factors to achieve optimal resource allocation. Different candidate resources may differ in execution cost and expected queuing time. By using a preset scheduling strategy, calculating the comprehensive cost of each candidate resource according to the optimization objective function, and selecting the resource with the lowest comprehensive cost as the target resource, it is possible to ensure that user job requirements are met while reducing costs, minimizing waiting time, and improving overall scheduling efficiency.

[0032] Furthermore, based on the attributes of the target resource and the type of its cluster, the corresponding job template and adapter must be invoked to convert the general description of the user job into a specific job script that conforms to the syntax of the target cluster's local job scheduling system. More specifically, based on the attributes of the target resource and the type of its cluster, a suitable job template is invoked from a pre-prepared job template library, and an adapter is used to process the general description of the user job. The parameters and instructions in the general description are converted according to the syntax requirements of the target cluster's local job scheduling system to generate a specific job script that conforms to the requirements of the target cluster, so that the target cluster can recognize and execute the job.

[0033] In summary, based on the aforementioned method for cross-cluster scheduling of heterogeneous computing resources, a unified virtual resource pool model is constructed and dynamic intelligent scheduling is implemented, effectively solving the technical challenges of resource integration and scheduling across heterogeneous computing clusters. Specifically, firstly, by establishing a unified resource description specification, hardware such as CPUs and GPUs with different architectures are abstracted into standardized computing units, solving the problem of difficult unified management caused by resource heterogeneity; secondly, by deploying an adaptation agent to collect the dynamic status of each cluster in real time and map it to standard data, accurate synchronization of the global resource view is achieved; on this basis, the system can intelligently match the optimal resources according to job requirements and make comprehensive decisions based on multiple factors such as execution cost and queuing time; finally, by automatically converting general tasks into instructions recognizable by the target cluster through job templates, seamless distribution of cross-cluster jobs is achieved. This solution fundamentally breaks down "computing silos" and significantly improves the overall utilization and scheduling efficiency of computing resources.

[0034] like Figure 2 As shown, one embodiment of the present invention proposes a cross-cluster intelligent scheduling system for heterogeneous computing resources, the system comprising: The virtual model construction module 10 is used to uniformly model different types of hardware resources in multiple heterogeneous computing power clusters according to the preset resource description specifications, and generate a globally unified virtual resource pool model. The data mapping module 20 is used to collect dynamic status information of the internal resources of each heterogeneous computing power cluster in real time, and to map and update it according to the virtual resource pool model. The resource parsing module 30 is used to receive user jobs, parse their resource requirement specifications, and search for a set of available resources that meet the resource requirement specifications among multiple clusters based on the updated virtual resource pool model. The resource scheduling module 40 is used to select target resources from the available resource set according to a preset scheduling strategy, and convert the user job into job instructions that can be recognized by the target cluster and distribute them to the corresponding cluster for execution.

[0035] In another aspect, the present invention also proposes a storage medium on which one or more programs are stored, which, when executed by a processor, implement the above-described method for cross-cluster scheduling of heterogeneous computing resources.

[0036] In another aspect, the present invention also proposes an electronic device, including a memory and a processor, wherein the memory is used to store computer programs and the processor is used to execute the computer programs stored in the memory, so as to realize the above-mentioned method for cross-cluster scheduling of heterogeneous computing resources.

[0037] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain stored, communicated, propagated, or transmitted programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0038] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0039] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0040] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A cross-cluster scheduling method for heterogeneous computing resources, characterized in that, The method includes: Based on the preset resource description specifications, different types of hardware resources in multiple heterogeneous computing power clusters are uniformly modeled to generate a globally unified virtual resource pool model. The dynamic status information of the internal resources of each heterogeneous computing power cluster is collected in real time and mapped and updated according to the virtual resource pool model. Receive user jobs, parse their resource requirement specifications, and based on the updated virtual resource pool model, retrieve a set of available resources that meet the resource requirement specifications across multiple clusters; From the set of available resources, a target resource is selected according to a preset scheduling strategy, and the user job is converted into a job instruction that can be recognized by the target cluster and distributed to the corresponding cluster for execution.

2. The cross-cluster scheduling method for heterogeneous computing resources according to claim 1, characterized in that, The step of uniformly modeling different types of hardware resources in multiple heterogeneous computing power clusters according to a preset resource description specification to generate a globally unified virtual resource pool model includes: Different architectures of CPU, GPU, and FPGA hardware resources are abstracted into computing unit objects with unified attribute fields, which include at least architecture type, computing power value, memory, and interconnect bandwidth.

3. The cross-cluster scheduling method for heterogeneous computing resources according to claim 2, characterized in that, The steps of mapping and updating based on the virtual resource pool model include: By deploying adaptation agents in each cluster, the local resource monitoring interface of each cluster is actively polled or listened to at a preset period to dynamically collect its original status data. Based on the standardized attribute set defined by the virtual resource pool model, the raw state data collected in real time is dynamically mapped into standard fields and values.

4. The cross-cluster scheduling method for heterogeneous computing resources according to claim 3, characterized in that, The step of retrieving a set of available resources that meet the resource requirement specifications across multiple clusters based on the updated virtual resource pool model includes: The resource requirements of the user job are matched and compared with the resource status data recorded in the resource status database, which has been formatted according to the unified attribute fields, to filter out all computing units and their respective clusters that meet the requirements in the current state.

5. A cross-cluster scheduling method for heterogeneous computing resources according to claim 4, characterized in that, The step of selecting a target resource from the available resource set according to a preset scheduling strategy includes: From the set of available resources, the comprehensive cost of each candidate resource is calculated according to the optimization objective function. The comprehensive cost is calculated by weighting the execution cost and the expected queuing time, and the candidate resource with the lowest comprehensive cost is selected as the target resource.

6. The cross-cluster scheduling method for heterogeneous computing resources according to claim 1, characterized in that, The step of converting the user job into job instructions recognizable by the target cluster includes: Based on the attributes of the target resource and the type of its cluster, the corresponding job template and adapter are invoked to convert the general description of the user job into a specific job script that conforms to the syntax of the target cluster's local job scheduling system.

7. A cross-cluster scheduling method for heterogeneous computing resources according to claim 6, characterized in that, The method further includes: Define and deploy unified resource discovery and status reporting interfaces for the various heterogeneous hardware resources and local scheduling systems to be connected.

8. A cross-cluster intelligent scheduling system for heterogeneous computing resources, characterized in that, The system includes: The virtual model building module is used to uniformly model different types of hardware resources in multiple heterogeneous computing power clusters according to preset resource description specifications, and generate a globally unified virtual resource pool model. The data mapping module is used to collect dynamic status information of the internal resources of each heterogeneous computing power cluster in real time, and to map and update the data according to the virtual resource pool model. The resource parsing module is used to receive user jobs, parse their resource requirement specifications, and, based on the updated virtual resource pool model, retrieve a set of available resources that meet the resource requirement specifications across multiple clusters. The resource scheduling module is used to select target resources from the available resource set according to a preset scheduling strategy, and convert the user job into job instructions that can be recognized by the target cluster, and distribute them to the corresponding cluster for execution.

9. A storage medium, characterized in that, The storage medium stores one or more programs, which, when executed by a processor, implement a cross-cluster scheduling method for heterogeneous computing resources as described in any one of claims 1-7.

10. An electronic device comprising a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements a cross-cluster scheduling method for heterogeneous computing resources as described in any one of claims 1-7.

Images