Management, modeling and simulation method and system for heterogeneous computing task scheduling

Abstract

The invention provides a management, modeling and simulation method and system for heterogeneous computing task scheduling, and the method comprises the steps: generating a global logic task identifier set according to the input parameters of a computing problem, and dividing the set into a plurality of logic task blocks based on the internal attributes of logic task identifiers; mapping each logic task block into an atomic computing unit, and packaging a computing attribute for each atomic computing unit; based on a preset screening rule and the resource constraint information, the atomic computing units are recombined, and schedulable task instances are formed; and allocating the schedulable task instances to the computing resources for execution, and monitoring performance data of the execution to update the quantifiable computing attributes of the atomic computing units by using the performance data. According to the method, the problems that the utilization efficiency of heterogeneous computing resources is low and simulation prediction is distorted due to the fact that a traditional task scheduling scheme lacks a quantifiable task abstraction and dynamic recombination mechanism across hardware platforms are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of quantum chemistry calculation and molecular simulation calculation, and particularly relates to a management, modeling and simulation method and system for heterogeneous computing task scheduling. BACKGROUND

[0002] In the field of quantum chemistry calculation, the number of electronic repulsion integral tasks is extremely large, and the calculation scale thereof increases exponentially with the increase of the complexity of a molecular system and a basis set, which can easily cause uneven load distribution and low calculation efficiency.

[0003] A traditional task scheduling scheme usually adopts a direct allocation mode with coarse granularity to distribute integral tasks to computing nodes as a whole, and lacks the consideration of the coordination of task calculation characteristics and hardware resource constraints. Such a rigid scheduling mechanism cannot dynamically split and recombine tasks according to the memory requirements and parallelism characteristics of the tasks, and is difficult to adapt to the heterogeneous characteristics of different architecture computing units (such as CPU, GPU and special accelerators), resulting in low utilization efficiency of computing resources. More seriously, due to the lack of precise modeling of the task execution process, the simulation platform cannot effectively predict the performance of different scheduling strategies in the real hardware environment, resulting in a significant deviation between the simulation results and the actual running effect, which seriously restricts the effectiveness of scheduling strategy optimization and system performance improvement.

[0004] Therefore, it is necessary to improve the existing scheduling scheme to solve the above technical problems. SUMMARY

[0005] The present application aims to provide a management, modeling and simulation method and system for heterogeneous computing task scheduling to solve the problem that the traditional task scheduling scheme lacks quantifiable task abstraction and dynamic recombination mechanism across hardware platforms, resulting in low utilization efficiency of heterogeneous computing resources and simulation prediction distortion.

[0006] To achieve the above-mentioned purpose, the present application is implemented as follows:

[0007] In a first aspect, the present application provides a task scheduling management method for a heterogeneous computing environment, comprising:

[0008] generating a set of global logical task identifiers according to input parameters of a computing problem, so as to divide the set into a plurality of logical task blocks based on one or more intrinsic attributes of the logical task identifiers;

[0009] mapping each of the logical task blocks to one or more atomic computing units, and encapsulating one or more quantifiable computing properties for each of the atomic computing units;

[0010] reorganize the atomic computing units based on a predetermined filtering rule and resource constraint information of a target execution environment to form at least one schedulable task instance;

[0011] allocate the schedulable task instance to a computing resource in the target execution environment for execution and monitor performance data of the execution to update the quantifiable computing properties of the atomic computing units with the performance data.

[0012] In a second aspect, a task abstraction modeling method for a heterogeneous computing environment is provided, comprising:

[0013] creating a logical layer model comprising a plurality of logical task blocks partitioned from a global logical task identifier set generated according to input parameters of a computing problem, and assigning a first set of descriptive properties to each logical task block;

[0014] creating a computing layer model by mapping the logical task blocks to a plurality of atomic computing units, and assigning a second set of quantifiable properties to each atomic computing unit;

[0015] creating a scheduling layer model by reorganizing the atomic computing units based on a predetermined filtering rule and resource constraint information of a target execution environment to form at least one schedulable task instance, and assigning a third set of aggregated properties to each schedulable task instance.

[0016] In a third aspect, a task scheduling simulation method for a heterogeneous computing environment is provided, executed in a discrete event driven simulation environment, comprising:

[0017] generating a global logical task identifier set and partitioning it to establish a logical layer model comprising a plurality of logical task blocks based on input molecular and basis set information;

[0018] mapping the logical task blocks to atomic computing units, and encapsulating quantifiable computing properties for each atomic computing unit to establish a computing layer model;

[0019] reorganizing the atomic computing units based on a predetermined filtering rule and resource constraint information of a simulated target hardware environment to form at least one schedulable task instance to establish a scheduling layer model;

[0020] advancing a time of the simulation environment to a predicted completion timestamp of the schedulable task instance and simulating completion of its execution to generate a task completion event;

[0021] updating the quantifiable computing properties of the atomic computing units in response to the task completion event.

[0022] In a fourth aspect, a task scheduling management system for a heterogeneous computing environment is provided, comprising:

[0023] a task planning module configured to generate a set of global logical task identifiers according to input parameters of a computing problem, and divide the set into a plurality of logical task blocks based on one or more intrinsic attributes of the logical task identifiers;

[0024] a task abstraction module configured to map each of the logical task blocks into one or more atomic computing units, and encapsulate one or more quantifiable computing attributes for each of the atomic computing units;

[0025] a task reorganization module configured to reorganize the atomic computing units according to predetermined screening rules and resource constraint information of a target execution environment, and form at least one schedulable task instance;

[0026] a task execution and feedback module configured to assign the schedulable task instance to a computing resource in the target execution environment for execution, monitor performance data of the execution, and update the quantifiable computing attributes of the atomic computing units in the task abstraction module using the performance data.

[0027] In a fifth aspect, the present application further provides a computer readable storage medium, wherein a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to implement the steps of the method according to the first aspect or the second aspect or the third aspect.

[0028] The present application has the following beneficial effects:

[0029] The task scheduling management method for a heterogeneous computing environment provided by the present application establishes a unified task description and scheduling basis across platforms by intelligent division based on intrinsic attributes of tasks and attribute encapsulation of atomic computing units; effectively realizes adaptive optimization and load balancing of task granularity in a heterogeneous computing environment by using a dynamic reorganization mechanism driven by attributes and aware of resource constraints, greatly improves resource utilization; and through closed-loop feedback and continuous updating of atomic unit computing attributes by performance monitoring data, the task estimation accuracy and scheduling strategy are constantly self-optimized, thereby ensuring high consistency between simulation deduction results and real hardware execution performance, and providing reliable technical support for efficient scheduling of large-scale computing tasks. Thus, the present application successfully solves the problems of low utilization efficiency of heterogeneous computing resources and simulation prediction distortion caused by lack of quantifiable task abstraction and dynamic reorganization mechanism across hardware platforms in the traditional scheme.

[0030] Furthermore, this invention proactively eliminates redundant computing units before task reorganization through a pre-screening mechanism, fundamentally reducing overall computational complexity. The hierarchical abstract model (logic layer -> computation layer -> scheduling layer) naturally supports flexible expansion and adaptation to heterogeneous hardware platforms, allowing migration to new computing environments without refactoring the scheduling logic. In addition, the task block preprocessing mechanism based on attribute sorting optimizes data locality, reduces cache misses and communication overhead during computation, and further improves overall execution efficiency. Attached Figure Description

[0031] Figure 1 This is a schematic flowchart of a task scheduling and management method for a heterogeneous computing environment according to an embodiment of the present invention;

[0032] Figure 2 This is a schematic flowchart of a task scheduling and management method for a heterogeneous computing environment according to another embodiment of the present invention;

[0033] Figure 3 This is a schematic flowchart of a task scheduling and management method for a heterogeneous computing environment according to another embodiment of the present invention;

[0034] Figure 4 A schematic flowchart illustrating a task abstraction modeling method for a heterogeneous computing environment according to an embodiment of the present invention;

[0035] Figure 5 A schematic flowchart of a task abstraction modeling method for a heterogeneous computing environment according to another embodiment of the present invention;

[0036] Figure 6 This is a schematic flowchart of a task scheduling simulation method for a heterogeneous computing environment according to an embodiment of the present invention.

[0037] Figure 7 This is a schematic structural block diagram of a task scheduling and management system for a heterogeneous computing environment according to an embodiment of the present invention;

[0038] Figure 8 This is a topology diagram of a computer-readable storage medium disclosed in this invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent changes or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0040] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Example One

[0042] like Figure 1 As shown, this embodiment provides a task scheduling management method (hereinafter referred to as "task scheduling management method" or "management method" or "method") for heterogeneous computing environments, used in a task scheduling management system (hereinafter referred to as "task scheduling management system" or "management system" or "system") for heterogeneous computing environments. The method includes:

[0043] Step 102. Generate a global set of logical task identifiers based on the input parameters of the computation problem, and divide the set into multiple logical task blocks based on one or more intrinsic attributes of the logical task identifiers. For example... Figure 2 As shown, the operations following step 102 also include:

[0044] Step 103. Sort the set based on the inherent attributes of the logical task identifier to optimize the load balancing efficiency of subsequent partitioning and reorganization steps.

[0045] Step 104. Map each logical task block to one or more atomic computation units, and encapsulate one or more quantifiable computational properties for each atomic computation unit.

[0046] Continue with Figure 2 The operations following step 104 include:

[0047] Step 105. Based on the integral symmetry property of atomic computing units, redundant atomic computing units that do not require computation are pre-screened and removed. That is, by utilizing the inherent mathematical symmetry of electronic integrals, redundant computing units with calculation results of zero or negligible values ​​are proactively identified and removed before the computation task is executed, directly reducing the total computational scale at the source of the problem. In this way, valuable computing resources and scheduling overhead are avoided from being wasted on invalid computations, laying an efficient foundation for subsequent task reorganization and scheduling execution, thereby significantly improving the efficiency of the overall computation process.

[0048] Step 106. Based on the predetermined screening rules and the resource constraint information of the target execution environment, the atomic computing units are reorganized to form at least one schedulable task instance.

[0049] It should be noted that the target execution environment can include various hardware computing units with specific memory capacity, number of parallel computing cores, and communication bandwidth, such as CPU clusters, GPU accelerators, or FPGA heterogeneous computing platforms. Of course, it can also be expanded into a dynamic, elastic virtual resource pool. This resource pool can include: local heterogeneous computing nodes (existing CPUs, GPUs, FPGAs), cloud virtualization accelerator instances (such as AWS EC2 F1 instances, Azure NVv4 series, Google CloudTPU), and computing resources from other remote collaborative organizations. Therefore, a Unified Resource Abstraction Layer can be added to the system, which is responsible for:

[0050] Resource discovery and registration involves dynamically discovering and registering available computing units from local and cloud environments, and standardizing their resource constraints (cost, bandwidth, memory, computing power type) into attributes that the system can understand.

[0051] Resource status monitoring: Real-time monitoring of the health status, load, real-time pricing (for cloud resources), and network latency of each resource in the pool.

[0052] Cost-aware scheduling has become a new basis for decision-making in the "task reorganization module". When reorganizing tasks, in addition to considering technical constraints, the "cost" attribute can also be added to achieve performance-cost trade-off optimization (for example, scheduling non-urgent tasks to run on lower-cost cloud Spot instances).

[0053] This setup breaks through the limitations of local physical resources, allowing for automatic and seamless expansion using cloud resources when demand surges. It also introduces an economic model to minimize the total cost of ownership (TCO) while meeting performance targets. Furthermore, if a single local node or cloud instance fails, the resource abstraction layer can mark it as unavailable and reassign the task to other resources in the pool, thus achieving fault tolerance.

[0054] The operation of "reorganizing atomic computing units based on predetermined screening rules and resource constraint information of the target execution environment" in step 106 includes: reorganizing atomic computing units using a differentiated strategy based on the quantifiable computing attributes and resource constraint information of the atomic computing units.

[0055] Specifically, aggregation operations are performed on atomic computing units with low computational load, while isolation operations are performed on atomic computing units with high resource requirements. For example... Figure 3 As shown, the specific implementation process is as follows:

[0056] Step 301. Aggregate multiple atomic computing units whose computation time is lower than the first threshold and whose memory usage is lower than the second threshold to form a composite schedulable task instance.

[0057] Step 302. Treat a single atomic computing unit whose memory usage exceeds the third threshold as an independent schedulable task instance.

[0058] It should be understood that this embodiment intelligently aggregates lightweight tasks and processes heavyweight tasks independently through a differentiated reorganization mechanism, achieving a high degree of matching between computing resources and task requirements. The aggregation operation packages multiple small tasks for execution, effectively reducing task scheduling overhead and improving the throughput efficiency of computing cores; while the isolation operation ensures the exclusive use of resources by tasks with large memory requirements, avoiding performance bottlenecks and memory overflow risks caused by resource contention, thereby maximizing load balancing and resource utilization efficiency in a heterogeneous computing environment as a whole.

[0059] Step 108. Assign schedulable task instances to computing resources in the target execution environment and monitor execution performance data to update the quantifiable computational attributes of atomic computing units using the performance data. This achieves adaptive optimization of the task management model.

[0060] The step of "updating the quantifiable computational attributes of atomic computing units using performance data" includes: updating the quantifiable computational attributes of atomic computing units based on monitored performance data using a statistical learning model to optimize the prediction accuracy and reorganization decisions for subsequent task instances.

[0061] The specific operation of updating the quantifiable computational attributes of atomic computing units through statistical learning models is as follows: record the actual execution time of each atomic computing unit in the schedulable task instance, and update the value of the "estimated computation time" attribute in the atomic computing unit using a weighted average algorithm.

[0062] This embodiment continuously integrates actual execution data through a closed-loop feedback mechanism ("monitoring -> comparison -> decision -> update" cycle), transforming task attribute prediction from static empirical values ​​to dynamic, adaptive, and accurate predictions. Furthermore, it continuously calibrates the computational time model of atomic computing units using algorithms such as weighted averaging, ensuring that task reorganization and scheduling decisions are based on increasingly realistic data. This significantly improves the credibility of simulations and the accuracy of resource allocation, ultimately achieving continuous self-optimization of system performance.

[0063] Therefore, the task scheduling and management method for heterogeneous computing environments in this embodiment establishes a unified cross-platform task description and scheduling foundation through intelligent partitioning based on the inherent attributes of tasks and attribute-based encapsulation of atomic computing units. Utilizing an attribute-driven and resource-constraint-aware dynamic reorganization mechanism, it effectively achieves adaptive optimization and load balancing of task granularity in heterogeneous computing environments, significantly improving resource utilization. Furthermore, through closed-loop feedback and continuous updating of atomic unit computing attributes using performance monitoring data, task prediction accuracy and scheduling strategies are continuously self-optimized, ensuring a high degree of consistency between simulation results and actual hardware execution performance. This provides reliable technical support for the efficient scheduling of large-scale computing tasks. Thus, it successfully solves the problems of low heterogeneous computing resource utilization efficiency and simulation prediction distortion caused by the lack of a cross-hardware platform quantifiable task abstraction and dynamic reorganization mechanism in traditional solutions.

[0064] In one specific embodiment, the task scheduling and management method for heterogeneous computing environments is used for task abstraction modeling of electron repulsion integral (ERI) computation and its implementation scheme in a two-level master-slave task distribution simulation platform. This embodiment's method, by layer-by-layer "deconstructing and packaging" the complex structural features of electron integration computation tasks, enables tasks at different computational levels to be generated, filtered, packaged, and distributed in the simulation environment, thereby supporting load balancing and performance optimization for large-scale quantum chemical integration computation tasks. The specific implementation steps are as follows:

[0065] I. Task Abstraction Modeling Levels

[0066] 1. Top-level tasks (Host layer)

[0067] (1) The Host first reads the basis set information of the molecular system and generates the index of the shell quartet as the top-level unit of task abstraction.

[0068] (i,j,k,l)i,j,k,l∈{1,2,3,...,N basis}(1)

[0069] As shown in Equation 1, (i, j, k, l) are the indices of the four basis sets, corresponding to a shell quartet. Each basis function is represented by a Gaussian Type Orbital (GTO), and the basic form of the basis set is a cGTO (contracted Gaussian Type Orbital), as shown in Equation 2:

[0070]

[0071] Where K is the contraction degree, Ti It is the Gaussian coefficient, v GTO This is the original Gaussian orbit. The basic form of the original Gaussian orbit is shown in Equation 3:

[0072]

[0073] in, α is the atomic coordinate, N is the normalization coefficient, l, m, n are quantum numbers, and α is the exponential coefficient.

[0074] (2) Each shell quartet corresponds to a set of contracted ERIs, which have attributes such as angular momentum, Gaussian coefficient, and memory usage requirements.

[0075]

[0076] As shown in Equation 4, (i, j, k, l) are basis set indices, (ij|kl) are a set of contracted electron repulsion integrals (ERI), and T a i T b j T c k T d l These are the Gaussian coefficients corresponding to different pGTOs, and (ab|cd) is a set of primitive electron repulsion integrals (primitive ERI).

[0077] (3) The host first sorts the index based on angular momentum and Gaussian coefficients, and divides the shell quartet index space into multiple task blocks according to configurable scheduling rules; according to the task requests of different SoCs, the divided quadruple index space is distributed to the SoC to build the entry point of the global task of the simulation system.

[0078] (4) The simulation platform only needs to input the atomic information to be simulated and the Gaussian group information used. The host can automatically generate the task index and distribute it in blocks, which significantly simplifies the task input complexity of the simulation platform.

[0079] 2. Intermediate Tasks (SoC Layer)

[0080] (1) The SoC generates a batch of shellquartet sets based on the four-fold loop start and end index issued by the Host and the base set information.

[0081] (2) Each shell quartet can be decomposed into multiple primitive ERIs. Each primitive ERI is assigned abstract task attributes, including angular momentum, Gaussian coefficient expansion, measured runtime, parallelizability, and estimated memory footprint. A group of primitive ERIs with the same coordinates and angular momentum can be packaged into a contracted ERI.

[0082]

[0083] As shown in Equation 5, (ab|cd) is a set of original electron repulsion integrals, χ pGTO These are primitive Gaussian orbitals, where r1 and r2 are atomic coordinates, and r... 12 It is an atomic coordinate vector.

[0084] (3) When generating tasks, the SoC first performs a pre-screening operation (such as removing tasks with an ERI value lower than a specified threshold) to eliminate redundant tasks that do not require computation.

[0085] (4) Based on hardware simulation constraints (such as memory limitations, parallelism limitations, etc.), the SoC further performs packet assembly or disassembly operations on ERIbatch to form task packages suitable for lower-level execution. Task packages with different angular momentum properties exhibit significant differences in processing:

[0086] a) For tasks with very small total angular momentum (e.g., (ss|ss)), memory is usually not a constraint (i.e., memory usage is less than a certain implicit second threshold). The number of primary ERIs in the task package is mainly limited by FMA capability. They can be aggregated and packaged according to the actual hardware parallel capability. That is, the computation time is short and many can be packaged, which is equivalent to less than a certain implicit first threshold.

[0087] b) For tasks with high total angular momentum (e.g., (gg|gg)), the memory requirement of a primary ERI is large (i.e., the memory usage is higher than a certain implicit third threshold), and such tasks usually need to be computed serially. Therefore, they do not need to be unpacked and cannot be parallelized with other tasks. They should be treated as separate task instances (isolated).

[0088] c) For tasks with moderate total angular momentum (e.g., (gs|fd)), the memory requirement of a primary ERI batch is small, allowing multiple primary ERI batches to be processed in parallel. The degree of parallelism depends on the Gaussian coefficient, which is used to group or unpack the batches.

[0089] 3. Low-level tasks (CP / APC layer)

[0090] (1) The CP maintains the task packet queue and dynamically and evenly distributes the task packets to idle or low-load APCs according to the real-time requests of APCs and the load status of each unit.

[0091] (2) APC uses the task package after SoC packaging as the basic computing unit and executes the primary ERI contained therein sequentially or in parallel.

[0092] (3) During the execution process, APC provides real-time feedback on the task status and completion time, which is used to drive the discrete event scheduling mechanism in the simulation platform.

[0093] Referring to Table 1, this embodiment decomposes the ERI task through a three-layer abstract model of logic layer, computation layer and scheduling layer: the top layer is handled by the Host entity for Shell Quartet indexing and block division, the middle layer is handled by the SoC entity for attribute encapsulation and reorganization of Primitive ERI, and the bottom layer is handled by the APC entity for specific computation and feedback of Task Package, thus forming a precise mapping and efficient scheduling closed loop from task abstraction to hardware execution.

[0094] Table 1: Mapping Relationship Between Task Abstract Model and Simulation Entity

[0095]

[0096] The task scheduling and management method for Electron Repulsion Integral (ERI) proposed in this embodiment decomposes the complex computational process, which is traditionally performed directly at the integration level, into multiple abstract task units layer by layer. These task units form a hierarchical mapping relationship with the simulation computation entities (Host). Top-level tasks, SoC Intermediate Tasks, APC (Underlying task). This hierarchical mapping mechanism not only significantly simplifies the complexity of dual-electron integral calculations, but also helps to achieve efficient matching and load-balanced scheduling between abstract tasks and physical computing resources.

[0097] II. Simulation and Scheduling Mechanism

[0098] Based on the abstract model of the electronic integration task, this embodiment not only supports high-fidelity simulation of the task scheduling process in a simulation environment, but also ensures the reusability and consistency of the model across different computing platforms. The specific mechanisms include the following three aspects:

[0099] 1. Discrete event-driven simulation

[0100] The lifecycle of all e-integration tasks, including generation, selection, distribution, execution, and completion, is modeled as discrete events and managed through an event priority queue. The simulation platform advances sequentially based on event timestamps, thereby accurately simulating the scheduling dynamics of large-scale e-integration tasks on real hardware within a virtual environment.

[0101] 2. Layered interaction mechanism

[0102] This embodiment employs a combination of top-down task indexing and bottom-up task requesting: APC, as the underlying computing entity, actively requests tasks to avoid idle computing resources; the Host layer distributes task indexes in batches according to real-time requests to prevent resource skew caused by excessive one-time distribution. This mechanism is not only applicable to simulation environments but can also be directly ported to various hardware architectures (such as CPU clusters, GPU arrays, and various heterogeneous accelerators).

[0103] 3. Resource Constraint Modeling

[0104] The task abstract model explicitly encapsulates configurable system attributes such as memory capacity, parallelism threshold, and communication overhead, ensuring that the packet assembly / disassembly strategy matches the actual hardware capabilities. This allows the scheduling strategy to flexibly adapt to different hardware environments. The task model can drive discrete event simulations and can also be directly deployed to run on actual hardware platforms.

[0105] It should be understood that this embodiment employs a hierarchical standardized abstraction method to process ERI tasks: from the top-level shellquartet index, to the middle-level ERI batch, and then to the bottom-level primitive Gaussian orbital (pGTO) combination. In the model, each task is encapsulated with structured attributes such as angular momentum type, Gaussian coefficients, symmetry conditions, memory usage estimates, and computation time estimates, giving each task a unique identifier and a quantifiable description of its resource requirements. The simulation platform can perform pre-screening, packet assembly, or packet splitting operations based on these attributes, and leverage these parameters in the simulation environment to achieve cross-level, cross-platform load balancing scheduling and high-fidelity performance prediction.

[0106] In summary, the method for the electron repulsion integral (ERI) task in this embodiment achieves consistent task management that can be reused across platforms by decomposing and mapping the complex task layer by layer and encapsulating computational attributes and resource constraints in each layer. This method not only provides a unified simulation optimization framework for large-scale quantum chemical simulations, but also establishes a consistent foundation for task migration and strategy reuse between heterogeneous platforms, significantly improving computational efficiency and resource utilization, and possessing high engineering feasibility and practical guiding value.

[0107] Example Two

[0108] likeFigure 4 As shown, this embodiment provides a task abstraction modeling method for heterogeneous computing environments, used in the task scheduling and management method for heterogeneous computing environments and the task scheduling and management system described in Embodiment 1. The method includes:

[0109] Step 402. Create a logical layer model, which includes a set of global logical task identifiers generated based on the input parameters of the computation problem, and assign a first set of descriptive attributes to each logical task identifier.

[0110] Step 404. Create a computational layer model, which is implemented by decomposing the logical task identifiers in the logical layer model into multiple atomic computational units and assigning a second set of quantifiable attributes to each atomic computational unit.

[0111] Step 406. Create a scheduling layer model, which reorganizes atomic computing units based on predetermined filtering rules and resource constraint information of the target execution environment to form at least one scheduled task instance, and assigns a third set of aggregation attributes to each scheduled task instance.

[0112] It is worth noting that the logic layer model, computation layer model, and scheduling layer model together constitute a hierarchical task abstraction model, which is used to provide a consistent task description between the simulation environment and the actual execution environment.

[0113] This embodiment achieves a unified description of the entire process from problem definition to hardware execution based on a hierarchical task abstraction model that combines a logic layer, a computation layer, and a scheduling layer. The logic layer establishes a task logical view through global identifiers with descriptive attributes, the computation layer accurately characterizes the computational granularity using atomic units with quantifiable attributes, and the scheduling layer dynamically reorganizes task instances and assigns aggregation attributes according to resource constraints and selection rules. This allows the same task model to drive discrete event simulations with high precision and to be seamlessly deployed to heterogeneous hardware platforms such as CPUs, GPUs, and FPGAs for execution, significantly improving the reliability of simulation results, the efficiency of hardware resource utilization, and the consistency of cross-platform task scheduling.

[0114] In the above embodiments, such as Figure 5 As shown, the "Create computational layer model" step 404 further includes: pre-screening the decomposed atomic computational units based on predetermined computational value screening rules, wherein the pre-screening includes:

[0115] Step 501. Based on the second set of quantifiable attributes of the atomic computing unit, determine whether its estimated computational contribution is lower than a predetermined threshold, and mark the atomic computing units that are lower than the threshold as redundant objects and remove them.

[0116] Step 502. Only atomic computation units that are not marked as redundant are used for the creation of subsequent scheduling layer models.

[0117] It should be understood that this embodiment actively eliminates redundant atomic units with low estimated contribution at the front-end computing layer through pre-screening, which significantly reduces the overall computing scale from the source, avoids the occupation of subsequent scheduling and execution resources by invalid computing, thereby greatly improving the efficiency of simulation and actual execution while ensuring computing accuracy, and optimizing the utilization rate of system resources.

[0118] It should be noted that the scheme or principle involved in the task abstract modeling method in this embodiment is the same as that in Embodiment 1, and the same or similar content will not be described in detail.

[0119] Example Three

[0120] like Figure 6 As shown, this embodiment proposes a task scheduling simulation method (hereinafter referred to as the "simulation method") for heterogeneous computing environments, which is executed in a discrete event-driven simulation environment, including:

[0121] Step 602. Based on the input molecule and basis set information, generate a global logical task identifier set and divide it to establish a logical layer model.

[0122] Step 604. Decompose the logical task identifier into atomic computational units, and encapsulate quantifiable computational attributes for each atomic computational unit to establish a computational layer model.

[0123] Step 606. Based on predetermined screening rules and the resource constraint information of the simulated target hardware environment, the atomic computing units are reorganized to form at least one schedulable task instance to establish a scheduling layer model.

[0124] Step 608. Advance the simulation environment time to the estimated completion timestamp of the schedulable task instance and simulate its completion to generate a task completion event.

[0125] Step 610. In response to the task completion event, update the quantifiable computational properties of the atomic computational unit.

[0126] It should be understood that the simulation method in this embodiment achieves high-fidelity simulation of the electronic integration task scheduling process by constructing logic layer, computation layer and scheduling layer models in a discrete event-driven environment and performing task reorganization and timestamp advancement based on simulated hardware constraints. Furthermore, by updating the computational attributes of atomic units in a closed loop through the event-driven mechanism, the simulation system has continuous self-optimization capabilities, which significantly improves the accuracy and reliability of scheduling strategy evaluation and provides an efficient and low-cost verification method for performance prediction and optimization of heterogeneous computing platforms.

[0127] It should be noted that the scheme or principle involved in the task scheduling simulation method of this embodiment is the same as that of Embodiment 1, and the same or similar contents will not be described in detail.

[0128] Example Four

[0129] like Figure 7 As shown, this embodiment provides a task scheduling and management system 700 for heterogeneous computing environments, including a task planning module 701, used to generate a global set of logical task identifiers based on the input parameters of the computation problem, and divide the set into multiple logical task blocks based on one or more intrinsic attributes of the logical task identifiers; a task abstraction module 702, used to map each logical task block to one or more atomic computing units, and encapsulate one or more quantifiable computational attributes for each atomic computing unit; a task reorganization module 703, used to reorganize the atomic computing units according to predetermined filtering rules and resource constraint information of the target execution environment to form at least one schedulable task instance; and a task execution and feedback module 704, used to allocate the schedulable task instance to the computing resources in the target execution environment for execution, monitor the execution performance data, and use the performance data to update the quantifiable computational attributes of the atomic computing units in the task abstraction module.

[0130] The system 700 in this embodiment adopts a hierarchical control architecture to achieve efficient task scheduling management: the task planning module 701 in the top-level control node first generates a set of logical tasks based on the computational problem and divides them into blocks, and then sends them down to the intermediate layer control node; the task abstraction module 702 in the intermediate layer decomposes the tasks into atomic units and encapsulates computational attributes, and the task reorganization module 703 groups or unpacks the atomic units according to resource constraints to form schedulable task instances; finally, the task execution and feedback module 704 allocates the task instances to the bottom-level computing unit 705 for execution, and collects performance data in real time to feed back to the intermediate layer, forming a closed-loop optimization mechanism, enabling the system to continuously calibrate task attributes and optimize scheduling decisions, thereby achieving efficient resource utilization and high-fidelity simulation across platforms.

[0131] It should be noted that the scheme or principle involved in the task scheduling management system 700 for heterogeneous computing environments in this embodiment is the same as the scheme or principle of the task scheduling management method for heterogeneous computing environments, and the same or similar contents will not be described in detail.

[0132] Example Five

[0133] This invention also provides a terminal device, which may include a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the above-described functionality. Figures 1-4The various processes shown in the embodiment of the task scheduling and management method for heterogeneous computing environments can achieve the same technical effect, and will not be described again here to avoid repetition.

[0134] Combination Figure 8 As shown, this embodiment also discloses a specific implementation of a computer-readable storage medium 800. This computer-readable storage medium 800 can be configured wholly or partially in a physical computer, server, cluster server, or data center.

[0135] In this embodiment, the computer-readable storage medium 800 stores computer program instructions 801. The computer program instructions 801 are read and executed by a processor 802 to perform the steps in the task scheduling and management method for heterogeneous computing environments disclosed in Embodiment 1.

[0136] Optionally, the computer-readable storage medium 800 can be configured as a server, and the server runs on a physical device used to build a private cloud, hybrid cloud, or public cloud. The computer-readable storage medium 800 can also be configured as random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0137] The computer-readable storage medium 800 is used to store a program, and the processor 802, upon receiving an execution instruction, executes the task scheduling and management method for heterogeneous computing environments disclosed in Embodiment 1.

[0138] Meanwhile, the processor 802 disclosed in this embodiment may be an integrated circuit chip with signal processing capabilities. The processor 802 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0139] The technical solution of the same part in the computer-readable storage medium 800 disclosed in this embodiment as in Embodiment 1 and / or Embodiment 2 is described in Embodiment 1 and / or Embodiment 2, and will not be repeated here.

[0140] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

[0141] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0142] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A task scheduling and management method for heterogeneous computing environments, characterized in that, include: A global set of logical task identifiers is generated based on the input parameters of the computation problem, and the set is divided into multiple logical task blocks based on one or more intrinsic attributes of the logical task identifiers. Each of the logical task blocks is mapped to one or more atomic computation units, and each atomic computation unit is encapsulated with one or more quantifiable computational attributes; Based on predetermined filtering rules and resource constraint information of the target execution environment, the atomic computing units are reorganized to form at least one schedulable task instance. The schedulable task instance is assigned to computing resources in the target execution environment for execution, and the performance data of the execution is monitored to update the quantifiable computing attributes of the atomic computing unit using the performance data.

2. The method according to claim 1, characterized in that, The reorganization of the atomic computing units based on predetermined filtering rules and resource constraint information of the target execution environment includes: Based on the quantifiable computational attributes and resource constraint information of the atomic computing units, a differentiated strategy is adopted to reorganize the atomic computing units; wherein, an aggregation operation is performed on atomic computing units with low computational load, and an isolation operation is performed on atomic computing units with high resource requirements.

3. The method according to claim 2, characterized in that, The process of performing aggregation operations on atomic computing units with low computational load and isolation operations on atomic computing units with high resource requirements includes: Multiple atomic computing units whose computation time is lower than the first threshold and whose memory usage is lower than the second threshold are aggregated to form a composite schedulable task instance; Individual atomic computation units with memory usage exceeding the third threshold are treated as independent schedulable task instances.

4. The method according to claim 1, characterized in that, Updating the quantifiable computational attributes of the atomic computing unit using the performance data includes: Based on the monitored performance data, the quantifiable computational attributes of the atomic computing units are updated through a statistical learning model to optimize the prediction accuracy and reorganization decisions for subsequent task instances.

5. The method according to claim 4, characterized in that, The step of updating the quantifiable computational attributes of the atomic computing unit through a statistical learning model includes: Record the actual execution time of each atomic computing unit in the schedulable task instance, and update the value of the "estimated computation time" attribute in the atomic computing unit using a weighted average algorithm.

6. The method according to claim 1, characterized in that, Following the step of "mapping each of the logical task blocks to one or more atomic computing units", the method further includes: Based on the integral symmetry property of the atomic computing units, redundant atomic computing units that do not require computation are pre-screened and eliminated.

7. The method according to claim 1, characterized in that, Following the step of "generating a global set of logical task identifiers", the following is also included: Based on the inherent attributes of the logical task identifiers, the set is sorted to optimize the load balancing efficiency of subsequent partitioning and reorganization steps.

8. A task abstraction modeling method for heterogeneous computing environments, characterized in that, include: Create a logical layer model, which includes dividing a global set of logical task identifiers generated based on the input parameters of the computation problem into multiple logical task blocks, and assigning a first set of descriptive attributes to each logical task block; A computational layer model is created by mapping the logical task blocks to multiple atomic computational units and assigning a second set of quantifiable attributes to each atomic computational unit; A scheduling layer model is created, which reorganizes the atomic computing units based on predetermined filtering rules and resource constraint information of the target execution environment to form at least one scheduled task instance, and assigns a third set of aggregate attributes to each scheduled task instance.

9. A task scheduling simulation method for a heterogeneous computing environment, characterized in that, Executed in a discrete event-driven simulation environment, including: Based on the input molecular and basis set information, a global set of logical task identifiers is generated and divided to establish a logical layer model containing multiple logical task blocks. The logical task blocks are mapped to atomic computing units, and each atomic computing unit is encapsulated with quantifiable computing attributes to establish a computing layer model. Based on predetermined screening rules and resource constraint information of the simulated target hardware environment, the atomic computing units are reorganized to form at least one schedulable task instance to establish a scheduling layer model. The time of the simulation environment is advanced to the estimated completion timestamp of the schedulable task instance, and its execution is simulated to generate a task completion event; In response to the task completion event, the quantifiable computational properties of the atomic computing unit are updated.

10. A task scheduling and management system for heterogeneous computing environments, characterized in that, include: The task planning module is used to generate a global set of logical task identifiers based on the input parameters of the computation problem, and to divide the set into multiple logical task blocks based on one or more intrinsic attributes of the logical task identifiers. The task abstraction module is used to map each of the logical task blocks to one or more atomic computing units, and to encapsulate one or more quantifiable computing attributes for each of the atomic computing units. The task reorganization module is used to reorganize the atomic computing units according to predetermined filtering rules and resource constraint information of the target execution environment to form at least one schedulable task instance. The task execution and feedback module is used to allocate the schedulable task instance to the computing resources in the target execution environment for execution, monitor the performance data of the execution, and use the performance data to update the quantifiable computing attributes of the atomic computing units in the task abstraction module.

Images