A method and system for structural evolution path constraint and state continuity maintenance control for a heterogeneous intelligent computing system

Abstract

This invention discloses a method and system for controlling structural evolution path constraints and maintaining state continuity in heterogeneous intelligent computing systems, belonging to the field of heterogeneous intelligent computing system control technology. The method includes: detecting structural change requirements in the heterogeneous intelligent computing system; acquiring the current structural state and the target structural state, the structural state including resource composition state, physical connection state, computing domain combination state, and operational constraint state; determining candidate structural evolution paths based on the current and target structural states; determining the legality of state transition relationships in the candidate structural evolution paths based on preset structural constraints, and determining legal evolution paths; executing structural evolution control based on the legal evolution paths; generating state identity information based on task state instances during structural evolution, and maintaining the association between task states before and after structural evolution; verifying the continuity of task states before and after structural evolution; and performing state submission or state rollback based on the verification results. This invention establishes a structural state space model, a task state space model, and a state identity anchoring mechanism to achieve path constraints and state continuity maintenance in heterogeneous intelligent computing systems during structural evolution. It can be widely applied to heterogeneous intelligent computing environments that include GPUs, ASICs, NPUs, CPUs, FPGAs, CXL extended resources, and Chiplet architectures.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous intelligent computing system control technology, and in particular to a control method and system for constraining and controlling the evolution path of the system structure and maintaining the continuity of the computing task operation state when the system structure dynamically changes due to computing task requirements, resource status changes, performance optimization, power consumption constraints or computing architecture adjustments in a heterogeneous intelligent computing environment containing multiple types of computing resources.

[0002] This invention is applicable to, but is not limited to, the following heterogeneous intelligent computing environments: graphics processing unit (GPU) computing clusters, dedicated artificial intelligence accelerators (ASICs), neural network processors (NPUs), central processing units (CPUs), field-programmable gate arrays (FPGAs), extended computing and storage resources based on Compute Express Link (CXL), combinatorial computing modules based on the Chiplet architecture, and cloud-edge-device collaborative intelligent computing infrastructure.

[0003] This invention belongs to the category of underlying control technology for heterogeneous computing infrastructure. It focuses on solving the problems of uncontrollable structural change paths and difficulty in ensuring the continuity of task states during the evolution of heterogeneous computing systems from static resource combination to dynamic structure. Background Technology

[0004] As the scale of artificial intelligence models continues to expand, especially with the development of large-scale language models, multimodal models, and complex intelligent agent systems, artificial intelligence computing infrastructure is evolving from a traditional single computing architecture to a heterogeneous computing system in which multiple types of computing resources work together.

[0005] Traditional AI computing systems typically employ fixed computing architectures. In such systems, the execution environment for computing tasks, data storage locations, and the connections between computing resources generally remain relatively stable. However, future AI computing infrastructure will incorporate a variety of computing resources, including GPUs, ASICs, NPUs, CXL Memory Pools, Chiplet Computing Fabric, and Edge AIDevice. The system architecture may dynamically adjust based on changes in model training phases, inference phase requirements, energy consumption constraints, service level requirements, data location changes, and chip resource status.

[0006] Therefore, future AI computing systems will no longer be fixed structures, but dynamic computing systems with continuous evolution capabilities.

[0007] Currently, the main technologies for addressing dynamic changes in computing systems include the following categories:

[0008] (a) Resource scheduling technology

[0009] Existing resource scheduling technologies primarily select appropriate computing resources based on CPU / GPU utilization, task load, resource availability, and performance metrics, with the core objective of finding a resource allocation scheme under current conditions. However, these technologies mainly focus on the mapping from tasks to target resources, neglecting the existence of a legitimate evolutionary path between the current architecture and the target architecture. For example, during the system's transition from a GPU cluster to a hybrid ASIC+CXL+Chiplet architecture, it may encounter intermediate states such as insufficient communication bandwidth, power consumption exceeding limits, and data access paths failing to meet security policies. Existing scheduling technologies cannot constrain these intermediate structural states.

[0010] (ii) Task migration technology

[0011] Existing virtual machine migration, container migration, and checkpoint recovery technologies primarily achieve task migration from one environment to another by saving the program execution context, memory state, and file system state. However, the task state in heterogeneous intelligent computing systems includes not only the execution environment but also the location of model parameters, intermediate computation results, data dependencies, cache state, computational topology, and resource constraints. Therefore, traditional migration technologies struggle to describe the complete state of a computing task during changes in a heterogeneous computing architecture.

[0012] (III) Virtualization State Preservation Technology

[0013] Virtualization technology can maintain software environment consistency, but its state description is typically based on virtual machine instances, container instances, and virtual resource mappings. However, in heterogeneous computing environments, changes in physical computing resources (GPU → ASIC or HBM → CXLMemory) can lead to changes in execution paths, data layouts, and communication topologies. Existing virtualization state preservation mechanisms cannot address the issue of whether task execution semantics remain continuous after structural changes.

[0014] Existing heterogeneous computing systems typically focus on whether the target state is reachable after structural changes, or whether the target state meets resource constraints. However, they lack constraint mechanisms for the legality of the system's structural state evolution path during structural changes, as well as control mechanisms for the continuity of task running state identities before and after structural changes. When the system evolves through multiple intermediate structural states, existing technologies cannot ensure that the task state identities maintain a continuous correspondence between different structural stages.

[0015] Therefore, there is an urgent need for a joint control method and system that can constrain the structural evolution path and simultaneously maintain the continuity of task state when the structural evolution of a heterogeneous intelligent computing system occurs. Summary of the Invention

[0016] (a) Purpose of the invention

[0017] The purpose of this invention is to provide a method and system for controlling the structural evolution path constraints and maintaining state continuity in heterogeneous intelligent computing systems.

[0018] By establishing a structural state space model, an evolution path constraint mechanism, a state identity anchoring mechanism, a multi-dimensional state association and migration mechanism, and a continuous joint verification mechanism, the system can achieve secure evolution and continuous maintenance of task state in heterogeneous intelligent computing systems during dynamic structural changes.

[0019] (II) Technical Solution

[0020] This invention provides a method for controlling structural evolution path constraints and maintaining state continuity, comprising the following steps:

[0021] Step S1: Detect structural change requirements

[0022] Acquire structural change trigger events in a heterogeneous intelligent computing system. These structural change events are triggered by at least one of the following: system state change, task requirement change, resource state change, performance requirement change, power consumption constraint change, and operating environment change.

[0023] Step S2: Obtain the current structural state

[0024] Establish a structural state-space model:

[0025]

[0026] in:

[0027] R(t) (Resource Composition Status): Used to describe the number of GPUs, ASICs, CPU resources, NPU resources, and storage resources contained in the current system.

[0028] P(t) (Physical Connection State): Used to describe node connection relationships, communication links, data access paths, and interconnection topology.

[0029] C(t) (Computation Domain Combination State): Used to describe the attribution of computation tasks, the division of execution domains, and the relationship of resource combination.

[0030] O x (t) (Structure operation constraint state): Used to describe power consumption limits, security policies, service level requirements, and scheduling constraints.

[0031] Step S3: Determine the evolution path of candidate structures

[0032] Based on the current structure state X_current and the target structure state X_target, determine the set of candidate paths:

[0033]

[0034] Where X0 represents the current structural state, X n For the target structural state, X1 to X n ₋1 represents the intermediate structural state.

[0035] The evolution path determination in this invention is not for planning spatial paths or task access paths, but for constraining and selecting the legal state transition relationships between the structural states of heterogeneous computing systems.

[0036] Step S4: Determine the validity of the path.

[0037] For any state transition in the path Perform legality judgment .

[0038] If the hardware resources are available, communication conditions are met, power consumption constraints are met, and security policies are met, then Legal is set to True; otherwise, Legal is set to False. Paths containing illegal state transitions are deleted, and only legal evolution paths that meet the constraints are retained.

[0039] Step S5: State Identity Anchoring

[0040] Before the structural evolution is executed, the task's running state is collected. State definition:

[0041]

[0042] Where E(t) represents the execution state, D(t) represents the data state, T(t) represents the topology state, and O s (t) represents the task execution constraint state.

[0043] Based on the structural state, task characteristics, and runtime environment information, establish an association mapping relationship between the task runtime state and the structural state:

[0044]

[0045] Here, G represents the association mapping relationship determined based on the measured state information at the current system runtime, rather than obtained through learning or prediction. This mapping relationship is used to characterize the impact of structural state changes on the task's runtime state space.

[0046] Generate status identity identifier:

[0047]

[0048] And it satisfies the state version continuity relationship:

[0049]

[0050] By establishing state identity associations for task state instances under specific structural conditions through SVI, the continuous correspondence of states throughout the entire structural evolution process is maintained.

[0051] Step S6: Associated State Transition Control

[0052] The system is adjusted along the legal structure evolution path while maintaining the correlation between the dimensions of S(t), including execution state mapping, data state mapping, topology state mapping, and constraint state mapping.

[0053] Step S7: Continuity verification and commit / rollback

[0054] Perform joint verification on S(t) before the change and S(t+1) after the change. The verification includes execution continuity, data consistency, topology validity, and constraint compliance. If C=True, the operation is committed; if C=False, a rollback is performed.

[0055] (III) Beneficial Effects

[0056] Compared with the prior art, the present invention has at least the following beneficial effects:

[0057] (1) Improve the security of structural changes in heterogeneous computing systems. By establishing a structural state space model and an evolution path constraint mechanism, the system no longer directly executes unverifiable structural switching, thus avoiding illegal intermediate structures, states that do not meet power consumption constraints, and unrecoverable topology changes.

[0058] (2) Achieving controllable evolution of structural change processes. Existing technologies mainly focus on whether the target state is attainable. This invention further focuses on whether a legal path is followed before reaching the target state, achieving path-level control, state-level constraints, and traceable evolution processes.

[0059] (3) Maintain the continuous correspondence between the state instances of the corresponding computing tasks during the process of changes across heterogeneous resources. Establish state identity anchors so that computing task state instances can maintain a continuous state correspondence during changes such as GPU→ASIC, GPU→CXL expansion node, CPU→NPU, and Chiplet recombination.

[0060] (4) Reduce the risk of large-scale heterogeneous computing infrastructure reconstruction. Applicable to AI data centers, heterogeneous supercomputing systems, intelligent computing clusters, and Chiplet computing platforms, reducing downtime, state loss, and task re-initialization costs.

[0061] (5) Form a control mechanism for the evolution of heterogeneous computing infrastructure structure, which is different from the traditional scheduling layer, resource management layer and virtualization layer. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the control system architecture for structural evolution path constraints and state continuity maintenance in heterogeneous intelligent computing systems according to the present invention.

[0063] Figure 2 This is a diagram of the state-space model of the structure of the present invention.

[0064] Figure 3 This is a task state model diagram for the present invention.

[0065] Figure 4 This is a schematic diagram illustrating the continuous state association in the state identity anchoring mechanism and structural evolution process of the present invention.

[0066] Figure 5 This is a schematic diagram illustrating the structural evolution path constraints of the present invention.

[0067] Figure 6 This is a schematic diagram of the structural evolution path constraint and state continuity maintenance control process of the present invention.

[0068] in, Figure 1 The accompanying drawings are a summary of the specification of this invention. Detailed Implementation

[0069] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention.

[0070] Example 1: Path Constraints and State Continuity Maintenance Control in the Evolution of GPU Computing Clusters to ASIC Heterogeneous Computing Architectures

[0071] Application scenarios

[0072] During the training and inference of artificial intelligence models, the system initially uses GPU clusters as the primary computing resource. As the model size increases, energy consumption constraints rise, or business needs change, the system needs to gradually introduce dedicated ASIC accelerated computing resources, forming a GPU+ASIC hybrid cluster.

[0073] Traditional systems typically perform resource switching directly, failing to determine whether a valid intermediate structure exists between the GPU and ASIC, whether data dependencies are satisfied, whether the execution state is continuous, or whether the task needs to be reinitialized. This embodiment employs the method of the present invention for control.

[0074] Current structural state acquisition

[0075] The system first obtains the current structural state X0 = X(t). The resource composition state R0 includes 128 GPU computing nodes, HBM memory resources, and high-speed network switching nodes. The physical topology state P0 includes GPU node connection relationships, data communication paths, and storage access paths. The computation domain state C0 includes the current training task execution domain, data parallel groups, and model parallel groups. The runtime constraint state O... x0 This includes current power consumption limits, real-time task requirements, and service level requirements.

[0076] Target structure state generation

[0077] The system generates a target structure X_T based on computing requirements, including GPU nodes, ASIC acceleration nodes, and high-speed interconnect resources, forming a GPU+ASIC heterogeneous computing structure.

[0078] Candidate evolutionary path determination

[0079] The system does not directly execute X0→X_T, but determines multiple candidate paths. Path A: GPU → Add ASIC node → Adjust computing domain → GPU+ASIC target structure. Path B: GPU → Migrate some tasks → Repartition computing domain → GPU+ASIC target structure. Path C: GPU → Shut down some GPUs → Switch ASIC → Target structure.

[0080] Path legality constraint determination

[0081] The system executes for each path. Determine the condition. For path C, since disabling part of the GPU and then switching to the ASIC may cause the task to be interrupted, the execution constraint O is not satisfied. x (t), therefore, is judged to be illegal and deleted. Path A and path B pass the legality check and enter the set of legal paths.

[0082] State identity anchoring

[0083] Before performing structural changes, the task state is generated. .

[0084] The state association mapping module determines the corresponding task state space mapping relationship under the current runtime environment Env based on the current GPU cluster structure state X0 and task characteristics Task:

[0085]

[0086] This mapping is used to define the anchoring baseline for the task's operational state before structural changes. Task State Record the current training iteration position, instruction execution context, intermediate computation stage (execution state E0), model parameter position, gradient data, cache state (data state D0), GPU communication relationship, and data access path (topology state T0).

[0087] Structural evolution is performed along a legal path.

[0088] The system executes changes in the following order: GPU state → GPU+ASIC intermediate state → target heterogeneous state. During the change process, SVIs remain associated with SVI0 → SVI1, ensuring that the structure changes (GPU resources change, ASIC is added) but the task identity remains unchanged.

[0089] Continuity verification

[0090] After the structural changes are completed, C(S0,S1) verification is performed. Continuity verification confirms the continuity of training steps and the consistency of the current computation position; data continuity verification confirms parameter integrity and gradient matching; topology continuity verification confirms the validity of data communication paths; constraint continuity verification confirms that power consumption and SLA are met. If C=True, the GPU+ASIC structure state is submitted; otherwise, the original GPU structure is rolled back.

[0091] Example 2: Structural Evolution Path Constraint Control in the CXL Dynamic Extension Environment

[0092] Application scenarios

[0093] As the scale of computing tasks increases, the system needs to dynamically add CXL memory expansion, CXL compute nodes, and high-speed shared resources, expanding from a GPU + local HBM structure to a GPU + HBM + CXL Memory Pool structure.

[0094] Structural change detection

[0095] The system detects CXL device joining events and obtains X(t), which includes newly added resources, new topology connections, and new access paths.

[0096] Evolutionary path constraints

[0097] The system identifies candidate paths: Path 1 involves adding CXL resources → establishing mappings → updating access paths; Path 2 involves adding CXL resources → immediately migrating data → modifying the execution domain. The system then uses O... x (t) Determine if the current task requires low latency, then prohibit the immediate migration of large amounts of data along path 2 and select the progressive mapping path (path 1).

[0098] State continuously maintained

[0099] During the CXL expansion process, the data state D(t) changes. Based on the current CXL-expanded structural state X(t), the current task characteristics (Task), and the runtime environment information (Env), the system determines the mapping relationship between the task runtime state and the structural state:

[0100]

[0101] The mapping relationship is a deterministic mapping used to characterize the impact of structural state changes on the task's runtime state space. The original task state and the extended state are maintained through SVI. The system verifies data access consistency, address mapping consistency, and communication path consistency; after successful verification, the CXL extended structure is submitted.

[0102] Example 3: Structural Evolution Path Constraint Control in a Chiplet Heterogeneous Reconfiguration Environment

[0103] Application scenarios

[0104] The heterogeneous computing system adopts a chipplet architecture, consisting of CPU chipplets, AI accelerator chipplets, and memory chipplets. As tasks change, the chipplet structure needs to be reorganized, from a CPU+Memory structure to a CPU+AI chipplet+Memory structure.

[0105] Chiplet structural state modeling

[0106] Resource state R(t) describes the chiplet type, quantity, and function. Topology state P(t) describes the connections between dies, bus relationships, and high-speed interconnects. Computation domain C(t) describes which chiplets constitute the execution domain. Runtime constraint state O x (t) describes power consumption, thermal constraints, and bandwidth requirements.

[0107] Chiplet Evolutionary Path Constraints

[0108] The system identifies candidate path A: Add AI Chiplet → Establish interconnect → Repartition computational domain; Candidate path B: Dismantle old computational domain → Reload tasks → Establish new architecture. Based on O x (t) Determines that path B may cause state interruption and data reloading, therefore it is prohibited. Lock path A.

[0109] Chiplet state continuous verification

[0110] After the structural changes are completed, C(S(t),S(t+1)) verification is performed. Based on the reconstructed structural state X(t), the current task characteristics (Task), and the runtime environment information (Env), the system establishes a mapping relationship between the task runtime state and the structural state.

[0111]

[0112] Where G is a deterministic mapping function determined based on the current measured state. The execution state verifies whether the CPU task is continuous, the data state verifies whether the Memory Chiplet data is consistent, the topology state verifies whether the Chiplet interconnects are satisfied, and the constraint state verifies whether power consumption and safety are satisfied. If successful, a new Chiplet structure is submitted; otherwise, the original Chiplet combination is restored.

[0113] Industrial applicability

[0114] This invention can be widely applied to AI training clusters, autonomous driving computing platforms, Chiplet heterogeneous integrated systems, robot control systems, edge heterogeneous computing environments, and cloud computing infrastructure containing CXL interconnect resources. By establishing a structural state space model, a task state space model, and a state identity anchoring mechanism, this invention enables heterogeneous intelligent computing systems to maintain state continuity during continuous structural evolution, providing underlying control support for next-generation adaptive intelligent computing infrastructure.

Claims

1. A method for controlling the structural state evolution process and the continuous relationship of task operation states during structural state changes in a heterogeneous intelligent computing system, characterized in that, include: Detecting structural changes in heterogeneous intelligent computing systems; Obtain the current structural state and target structural state of the heterogeneous intelligent computing system, wherein the structural state is used to describe the resource composition state, physical connection state, computing domain combination state and operation constraint state of the heterogeneous intelligent computing system; Based on the current structural state and the target structural state, at least one candidate structural evolution path is determined, wherein each candidate structural evolution path includes at least one intermediate structural state; Based on preset structural constraints, the legality of the state transition relationships in the candidate structural evolution paths is determined to identify legal evolution paths that satisfy the structural constraints. Execute structural evolution control of the heterogeneous intelligent computing system based on the legal evolution path; During the structural evolution process, state identity information is generated based on the task state instance, and the association between the task states before and after the structural evolution is maintained based on the state identity information. Continuity verification of the task state before and after structural evolution; Perform a status commit operation or a status rollback operation based on the continuity verification results; Specifically, when the continuity verification result meets the preset continuity conditions, the task state after structural evolution is submitted; when the continuity verification result does not meet the preset continuity conditions, the task state before structural evolution is restored.

2. A structural evolution path constraint and state continuity maintenance control system for heterogeneous intelligent computing systems, characterized in that, include: The structural state perception module is used to acquire the current structural state and the target structural state of the heterogeneous intelligent computing system. The evolution path determination module is used to determine a set of candidate structure evolution paths based on the current structural state and the target structural state; The path constraint determination module is used to determine the legality of the state transition relationships in the candidate structure evolution path according to preset structural constraints, and to determine the legal evolution path. The structural evolution control module is used to control the heterogeneous intelligent computing system to perform structural evolution according to the legal evolution path; The state identity anchoring module is used to generate state identity identification information based on task state instances and establish the association between task states before and after structural evolution. The state continuity verification module is used to jointly verify the task state before and after the structure evolution; The commit rollback control module is used to perform state commit or state rollback based on the joint verification results.

3. The method according to claim 1 or the system according to claim 2, characterized in that: The structural state is represented as follows: in: R(t) represents the resource composition state, used to describe the type, quantity, and resource configuration relationship of heterogeneous computing resources; P(t) represents the physical connection state, which is used to describe the physical interconnection relationship and communication path between computing resources; C(t) represents the computation domain combination state, which is used to describe the logical computation domain partitioning relationship formed by computation resources; O x (t) represents the structural operational constraint state, used to describe the corresponding constraint conditions during the structural changes of the system.

4. The method according to claim 1 or the system according to claim 2, characterized in that: The task running status is represented as follows: in: E(t) represents the execution state; D(t) represents the data state; T(t) represents the topological state; O s (t) represents the task execution constraint state; The method or system establishes a mapping relationship between the task running state and the structural state based on the structural state, task characteristics, and operating environment information: in: X(t) represents the structural state; Task represents the characteristics of a computational task; Env represents runtime environment information; Where G represents the association mapping relationship determined based on the measured state information at the current system operating moment, rather than obtained through learning or prediction.

5. The method according to claim 1 or the system according to claim 2, characterized in that: The state identity identification information is a state identity anchor point generated based on task state instance, structural state, runtime environment information, and time information. The status identification information is used for: Identify corresponding task state instances during structural changes in heterogeneous intelligent computing systems, and establish a continuous correspondence between state instances before and after structural changes; Maintain the relationships between execution status, data status, topology status, and runtime constraint status; The status identity information is represented as follows: And it satisfies the state version continuity relationship: Where SVI0 represents the state identity identifier before the structural change, SVI n The state identity identifier represents the state after structural changes, and the version continuity relationship is used to represent the continuous evolution identity of the same computing task under different structural states.

6. The method according to claim 1 or the system according to claim 2, characterized in that: The legality determination includes determining the resource composition state, physical connection state, computational domain combination state, and operational constraint state of the continuous structural states one by one. When all state dimensions between the continuous structural states satisfy the preset constraints, the state transition between the continuous structural states is determined to be legal. When any state dimension does not meet the preset constraints, the state transition between the continuous structural states is determined to be illegal.

7. The method according to claim 1 or the system according to claim 2, characterized in that: The continuity verification includes: Continuity verification of execution states before and after structural evolution; Verify the consistency of data states before and after structural evolution; Verify the legality of the topological state before and after structural evolution; Perform compliance verification on the task operation constraint states before and after structural evolution; When all the above verification results meet the preset conditions, the task status is determined to be continuous. Otherwise, the execution status will be rolled back.

8. The method according to claim 1 or the system according to claim 2, characterized in that: The structural evolution process includes multiple consecutive evolution stages, each of which corresponds to a structural state instance and its corresponding state identity information. By leveraging the version continuity relationship among multiple state identity information, continuous state tracking and identity association can be achieved throughout the entire process of structural evolution.

9. The method according to claim 1 or the system according to claim 2, characterized in that: The structural change requirement is triggered by at least one of the following: system state change, task requirement change, resource state change, performance requirement change, power consumption constraint change, and operating environment change.

10. The method or system according to any one of claims 1 to 9, characterized in that: The heterogeneous intelligent computing system includes one or more heterogeneous computing environments: GPU computing environment, ASIC accelerated computing environment, CPU-GPU heterogeneous computing environment, NPU artificial intelligence computing environment, FPGA reconfigurable computing environment, computing and storage expansion environment based on CXL interconnect technology, and heterogeneous integrated computing environment based on Chiplet architecture.

Images