A state migration and deterministic execution control system and method across execution domains
By implementing a cross-domain state transition and deterministic execution control system, the problems of task scheduling instability and cross-domain synchronization error accumulation under heterogeneous communication media are solved. This system achieves adaptability and stability when communication conditions change, and enhances the deterministic execution capability of large-scale AI systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 樊雪梅
- Filing Date
- 2026-06-22
- Publication Date
- 2026-07-28
AI Technical Summary
Existing distributed computing systems suffer from problems such as task scheduling instability, cross-domain synchronization error accumulation, and inability to adapt to changes in communication conditions under heterogeneous communication media.
A cross-execution domain state transition and deterministic execution control system is adopted. The system acquires physical layer telemetry data through the communication monitoring module, generates communication constraint vectors, divides the execution domain, and uses a closed-loop feedback controller to update the domain division parameters, thereby realizing cross-cycle irreversible execution domain control.
It achieves task scheduling stability and system adaptability under heterogeneous communication media conditions, enhances the deterministic execution capability of large-scale AI systems, and ensures that the system maintains cross-node consistency when communication conditions change.
Smart Images
Figure CN122476027A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of distributed computing system technology, specifically to a cross-execution domain state transition and deterministic execution control system and method, applicable to the execution location control of computing tasks in AI training clusters, heterogeneous computing environments, and data center scenarios. Background Technology
[0002] As the scale of training large-scale artificial intelligence models continues to expand, the communication infrastructure of distributed computing systems is undergoing profound changes. The traditional performance optimization paradigm centered on GPU computing power is being broken, and the bandwidth, latency, and energy consumption of communication interconnects are becoming bottlenecks restricting the overall efficiency of the system. At the same time, data center communication media are showing a significant trend of heterogeneity, with multiple technical routes such as copper backplane interconnects, pluggable optical modules, linear pluggable optics, and co-packaged optics coexisting, and the system may dynamically switch between different communication media.
[0003] Existing distributed computing systems typically employ resource scheduling methods based on static topology or pre-defined rules. For example, schedulers like Kubernetes rely on static node labels for task allocation. While these methods are effective in scenarios with relatively stable communication conditions, they suffer from the following drawbacks:
[0004] (1) It is impossible to perceive the real-time physical state changes of the communication link. Existing scheduling methods only focus on the computational load of nodes or the network layer congestion, and do not involve the structural migration of communication constraints caused by changes in the physical layer communication medium.
[0005] (2) Lack of a unified modeling mechanism based on physical layer telemetry. Existing solutions treat “computing” and “communication” as independent resource dimensions, and lack a method to uniformly model the physical characteristics of communication links as schedulable constraint vectors.
[0006] (3) It lacks the ability to control closed-loop consistency across execution domains. When the system re-divides execution domains due to changes in communication conditions, the migration of tasks between different domains may lead to the accumulation of cross-domain synchronization errors. Existing methods lack a control mechanism to iteratively update domain partitioning parameters based on closed-loop feedback of execution results.
[0007] (4) Scheduling instability in scenarios with coexistence of heterogeneous communication media. In data centers where copper backplanes, linear pluggable optics, and co-packaged optics are deployed in a mixed manner, different domains have drastically different communication characteristics, and existing static scheduling schemes cannot effectively utilize this heterogeneity. Summary of the Invention
[0008] (a) Technical problems to be solved
[0009] This invention provides a cross-execution domain state transition and deterministic execution control system and method to solve the technical problems of task scheduling instability, cross-domain synchronization error accumulation, and the inability of the system to adapt to changes in communication conditions under heterogeneous communication media.
[0010] (II) Technical Solution
[0011] The present invention adopts the following technical solution:
[0012] A cross-execution domain state transition and deterministic execution control system, comprising:
[0013] The communication monitoring module (110) is used to obtain physical layer telemetry data from the communication links between multiple computing nodes. The telemetry data includes at least link delay, bandwidth utilization, energy consumption indicators and bit error rate.
[0014] A constraint vector generator (120) is used to generate a communication constraint vector based on the physical layer telemetry data, wherein the communication constraint vector includes at least a delay parameter, a bandwidth parameter, and an energy consumption parameter.
[0015] An execution domain partitioning engine (130) is used to divide multiple computing nodes into at least two execution domains based on the matching relationship between the computing task feature vector and the communication constraint vector, with each execution domain corresponding to different communication physical constraints.
[0016] The task execution topology controller (140) is used to map computation tasks to computation nodes in the corresponding execution domain for execution based on the partitioning results of the execution domain;
[0017] A closed-loop feedback controller (150) is used to update the partitioning parameters of the execution domain partitioning engine based on the cross-node synchronization error generated during task execution;
[0018] The execution domain partitioning result is locked as the unique execution location constraint of the computing task within the current scheduling cycle, and the closed-loop feedback controller only acts on the execution domain reconstruction of the next scheduling cycle, thereby forming a cross-cycle irreversible execution domain control architecture.
[0019] A method for state transition and deterministic execution control across execution domains includes the following steps:
[0020] S1: Acquire physical layer telemetry data;
[0021] S2: Generate communication constraint vectors;
[0022] S3: Divide the execution domain;
[0023] S4: Mapping and executing tasks;
[0024] S5: Closed-loop feedback update.
[0025] (III) Beneficial Effects
[0026] Compared with the prior art, the present invention has the following advantages:
[0027] (1) Realize the dynamic mapping from communication state to computing structure. This invention is the first to model physical layer communication telemetry data as constraint vectors and drive the execution domain partitioning of computing nodes based on this, thus solving the technical problem of the disconnect between "communication state" and "computation scheduling" in traditional schemes.
[0028] (2) Supports unified modeling and adaptive scheduling of heterogeneous communication media. This invention uses constraint vectors to perform unified modeling of various communication media such as copper interconnects, linear pluggable optical devices, and co-packaged optical devices. The system can automatically identify the step change of constraint vectors caused by media switching and trigger execution domain reconstruction.
[0029] (3) Introducing a closed-loop feedback mechanism to improve system stability. This invention continuously monitors cross-domain synchronization deviation through a closed-loop feedback controller and iteratively updates the execution domain partitioning parameters based on the deviation, enabling the system to quickly converge to a stable state after changes in communication conditions.
[0030] (4) Enhance the deterministic execution capability of large-scale AI systems. This invention ensures that the system maintains cross-node consistency when communication conditions change through a non-switched control loop that includes telemetry acquisition, constraint vector generation, execution domain partitioning, task mapping, and closed-loop feedback updates. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of the overall architecture of the cross-execution domain state transition and deterministic execution control system of the present invention;
[0032] Figure 2 This is a schematic diagram of the communication constraint vector generation structure of the present invention;
[0033] Figure 3 This is a schematic diagram of the execution domain partitioning and task matching mechanism of the present invention;
[0034] Figure 4 This is a schematic diagram of the communication medium state transition triggering mechanism of the present invention;
[0035] Figure 5 This is a schematic diagram of the closed-loop feedback control structure of the present invention;
[0036] Figure 6 This is a schematic diagram of the three-layer architecture of the topology control layer of this invention. Detailed Implementation
[0037] 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.
[0038] Example 1: System Overall Architecture
[0039] like Figure 1 As shown, the cross-execution domain state transition and deterministic execution control system of the present invention includes five core modules: a communication monitoring module (110), a constraint vector generator (120), an execution domain partitioning engine (130), a task execution topology controller (140), and a closed-loop feedback controller (150).
[0040] The communication monitoring module (110) collects physical layer telemetry data from the communication links between multiple computing nodes. The telemetry data is obtained through in-band network telemetry protocol, switch port mirroring or network card hardware counter. The sampling period is 100 microseconds to 10 milliseconds. The telemetry data includes at least link delay, bandwidth utilization, energy consumption indicators and bit error rate.
[0041] The constraint vector generator (120) generates communication constraint vectors based on the telemetry data. For example... Figure 2 As shown, the constraint vector generator extracts four dimensions from the telemetry data: delay L, bandwidth B, energy consumption E, and bit error rate ρ, and constructs a four-dimensional constraint vector V=[L,B,E,ρ]. Here, L is the link delay, B is the bandwidth utilization, E is the energy consumption index, and ρ is the bit error rate; all parameters are derived from physical layer hardware measurements.
[0042] The execution domain partitioning engine (130) divides multiple computing nodes into at least two execution domains based on the matching relationship between the computation task feature vector and the communication constraint vector. Each execution domain corresponds to different communication physical constraints. Figure 3 As shown, the task feature vector T=[CCR,SG] contains at least two dimensions: communication computation ratio and state synchronization granularity. The execution domain partitioning engine calculates the weighted distance between T and V, dividing the node into multiple execution domains, each corresponding to a specific set of communication performance constraints.
[0043] The task execution topology controller (140) maps computational tasks to computational nodes in the corresponding execution domains based on the division of execution domains. Tasks with high communication-to-computation ratios are preferentially assigned to low-latency domains, while tasks with low communication-to-computation ratios are assigned to low-cost domains.
[0044] The closed-loop feedback controller (150) updates the partitioning parameters of the execution domain partitioning engine based on the cross-node synchronization errors generated during task execution. For example... Figure 5As shown, the closed-loop feedback controller takes the cross-domain synchronization deviation as input and outputs the adjustment offset of the execution domain division boundary, and only applies to the execution domain reconstruction in the next scheduling cycle.
[0045] Each module is executed in the order of telemetry acquisition, constraint vector generation, execution domain partitioning, task mapping, and closed-loop feedback update, forming a non-exchangeable control closed loop. That is, the execution domain partitioning result of the current cycle is locked before task scheduling, and the feedback update takes effect in the next cycle, prohibiting parallel execution across stages.
[0046] In a preferred embodiment, the execution domain is defined as a set of computing resources determined in real time by communication constraint vectors. Its member nodes are dynamically generated based on the current physical layer communication state, rather than on a static topology or pre-configured logical groups. The partitioning of the execution domain directly corresponds to the actual execution location of the computing task in the distributed computing system and is used to control the physical execution migration of the computing task between computing nodes.
[0047] Example 2: Generation of Communication Constraint Vectors
[0048] like Figure 2 As shown, the constraint vector generator (120) constructs a four-dimensional communication constraint vector based on telemetry data. In one specific embodiment, the system obtains real-time physical layer data of each communication link through an in-band network telemetry protocol.
[0049] The delay parameter L is calculated based on a weighted average of round-trip time and jitter, reflecting the time sensitivity of the link.
[0050] The bandwidth parameter B is calculated based on the ratio of peak bandwidth to average bandwidth, reflecting the bandwidth stability of the link.
[0051] The energy consumption parameter E is calculated based on the energy consumption per bit transmitted.
[0052] The bit error rate parameter ρ is based on link-layer bit error rate statistics and reflects the signal integrity of the link.
[0053] The constraint vector generator combines the above parameters into V=[L,B,E,ρ] and outputs it to the execution domain partitioning engine (130).
[0054] Example 3: Dynamic Partitioning of Execution Domains and Task Mapping
[0055] like Figure 3 As shown, the execution domain partitioning engine (130) performs node grouping based on the similarity measure between the task feature vector and the communication constraint vector.
[0056] The task feature vector T=[CCR,SG] contains at least two dimensions: the communication computation ratio, which represents the proportion of time spent on cross-node communication during task execution, and the state synchronization granularity, which represents the frequency and scale of state synchronization.
[0057] The execution domain partitioning engine calculates the weighted distance between the task feature vector T and the communication constraint vector V, and uses a clustering algorithm to divide the nodes into multiple execution domains. The three execution domains are arranged in ascending order of communication distance and bandwidth: copper interconnect domain, linear pluggable optics domain, and co-packaged optics domain, each adapted to tasks with different communication computation ratios.
[0058] In one specific embodiment, the node is divided into a copper interconnect domain, a linear pluggable optics domain, and a co-packaged optics domain, and the domain boundaries are reconstructed in real time driven by communication constraint vectors.
[0059] Example 4: Communication Medium State Transition Triggering Mechanism
[0060] like Figure 4 As shown in the figure, this embodiment demonstrates the system's automatic detection and response process when the communication medium state changes.
[0061] The communication monitoring module (110) continuously collects physical layer telemetry data. When the system detects that the energy consumption change exceeds the first preset threshold or the delay change exceeds the second preset threshold, the system determines that the communication medium has undergone a state transition and triggers the execution domain reconstruction.
[0062] After detecting structural changes in the constraint vector within multiple consecutive sampling periods, the constraint vector generator (120) updates the constraint vector. The execution domain partitioning engine (130) completes execution domain repartitioning within several scheduling periods, re-aggregating nodes in the communication domain after state transition into the corresponding execution domain, and migrating the corresponding tasks to that domain. The closed-loop feedback controller (150) completes convergence verification in subsequent periods.
[0063] In a hybrid deployment environment, the system simultaneously supports copper interconnect domains, linear pluggable optics domains, and co-packaged optics domains, all scheduled through a unified control plane. When a state change occurs in an execution domain, the system only performs local repartitioning of that domain, rather than global recomputation, ensuring global consistency in heterogeneous communication environments.
[0064] Example 5: Closed-loop feedback control and cross-cycle reconfiguration
[0065] like Figure 5 As shown, the closed-loop feedback controller (150) updates the execution domain partitioning parameters based on the cross-node synchronization error.
[0066] Cross-domain synchronization deviation is defined as the degree of inconsistency between the global parameter version vector and the local confirmation vectors of each execution domain. The closed-loop feedback controller uses a proportional-integral-derivative control law to adjust the execution domain boundaries. It should be noted that this invention is not limited to proportional-integral-derivative control; it can also be implemented using model predictive control, rule-based controllers, or other feedback control methods, as long as the constraint of irreversible reconfiguration across cycles is met.
[0067] The cross-cycle characteristic of closed-loop feedback control is that the execution state within the current scheduling cycle cannot be modified by feedback, and the feedback signal only affects the execution domain reconstruction of the next scheduling cycle. If the current cycle detects a cross-domain synchronization deviation exceeding the threshold, the feedback controller calculates the adjustment amount, which takes effect during the execution domain construction process of the next cycle.
[0068] If the cross-domain synchronization deviation exceeds the preset threshold, the system triggers a fast reconvergence mechanism to temporarily expand the state synchronization time window and prevent execution jitter during state changes.
[0069] In a preferred embodiment, the physical layer telemetry data directly participates in the control path generation process and serves as one of the input bases for execution domain partitioning. After being converted into multi-dimensional communication constraint vectors by a constraint vector generator, the telemetry data is used not only to analyze communication status but also directly to determine the execution location and resource ownership of computational tasks. The execution domain partitioning engine outputs control decision results based on the constraint vectors. These results are locked within the current scheduling cycle and serve as constraints for task execution in the next scheduling cycle, forming a control link of "telemetry data → constraint vector → execution domain partitioning → task execution location change".
[0070] Example 6: Execution of a three-layer architecture for the topology control layer
[0071] like Figure 6 As shown, the complete form of this invention in a distributed computing system deployment environment includes a three-layer structure.
[0072] The physical communication layer contains a variety of heterogeneous communication media, including copper backplane interconnects, linear pluggable optical interconnects, and co-packaged optical interconnects, each with different delay characteristics and bandwidth capabilities.
[0073] The execution topology control layer is the core control layer of the system, comprising five modules: a communication monitoring module (110), a constraint vector generator (120), an execution domain partitioning engine (130), a task execution topology controller (140), and a closed-loop feedback controller (150). This layer generates constraint vectors based on physical layer telemetry signals, drives the dynamic reconstruction of execution domain boundaries, and controls the physical execution location of computational tasks.
[0074] The computation execution layer is the final execution layer for computation tasks, and the tasks are executed within the execution domain determined by the execution topology control layer.
[0075] The execution topology control layer achieves system stability through the following mechanisms: physical layer telemetry serves as the control input signal rather than monitoring data; execution domain boundaries are reconstructed in real time driven by communication constraint vectors; and cross-cycle closed-loop feedback only affects the execution domain reconstruction in the next cycle. The control objective is the physical execution topology of the computational task.
[0076] In a preferred embodiment, the execution domain partitioning and task scheduling process of the present invention strictly follows a fixed scheduling cycle constraint, wherein the execution domain partitioning must be completed within a preset time window and locked before the start of the next scheduling cycle. Cross-cycle delay optimization, historical state backtracking, or offline recalculation are not allowed in the current execution domain decision-making process, thereby ensuring that the system always makes control decisions based on the latest communication physical state at runtime. The execution domain partitioning, task scheduling, and closed-loop feedback update together constitute a control closed-loop structure.
[0077] Instruction manual with accompanying drawings
[0078] like Figure 1 As shown in the figure, this diagram illustrates the five core modules (110-150) of the system and their control flow relationships: 110 Communication Monitoring Module → 120 Constraint Vector Generator → 130 Execution Domain Partitioning Engine → 140 Task Execution Topology Controller → 150 Closed-Loop Feedback Controller. Dashed arrows represent closed-loop feedback loops (150→130), and solid arrows represent the main control flow, labeled "Non-commutative Control Loop".
[0079] like Figure 2 As shown in the figure, this figure illustrates the structure of the constraint vector generator (120): the input is physical layer telemetry data (delay L / bandwidth B / power consumption E / bit error rate ρ), the hardware sources include INT / Port Mirroring / NIC Counter, and the output is the communication constraint vector V=[L,B,E,ρ].
[0080] like Figure 3 As shown in the figure, this diagram illustrates the working mechanism of the execution domain partitioning engine (130): the input is the task feature vector T=[CCR,SG] and the communication constraint vector V, and the output is the CU copper interconnect domain, LPO domain, and CPO domain. The core annotation is "Execution domain = physical execution topology boundary", and the execution domain boundary is reconstructed in real time driven by the communication constraint vector.
[0081] like Figure 4As shown in the figure, this diagram illustrates the triggering mechanism for communication medium state transition: copper interconnect domain (L≈2.5μs, E reference) → LPO domain (L≈1.0μs, E↓60%) → CPO domain (L≈0.7μs, E↓50%), with the triggering condition being ΔE>30% or ΔL>20%.
[0082] like Figure 5 As shown in the figure, this diagram illustrates the structure of the closed-loop feedback controller (150): the input is the cross-domain synchronization error E(t), the control law is PID control, and the output is the execution domain adjustment u(t). The core annotation is "cross-cycle closed-loop feedback: execution state locking in the current cycle → execution domain reconstruction in the next cycle".
[0083] like Figure 6 As shown in the figure, this diagram illustrates the three-layer architecture of the system: Layer 1 is the physical communication layer (copper interconnect / LPO / CPO), Layer 2 is the execution topology control layer (five modules from 110 to 150), and Layer 3 is the computation execution layer, with the control objective being the Execution Topology.
Claims
1. A cross-execution domain state transition and deterministic execution control system, characterized in that, include: The communication monitoring module (110) is used to obtain physical layer telemetry data from the communication links between multiple computing nodes. The telemetry data includes at least link delay, bandwidth utilization, energy consumption indicators and bit error rate. A constraint vector generator (120) is used to generate a communication constraint vector based on the physical layer telemetry data, wherein the communication constraint vector includes at least a delay parameter, a bandwidth parameter, and an energy consumption parameter. An execution domain partitioning engine (130) is used to divide multiple computing nodes into at least two execution domains based on the matching relationship between the computing task feature vector and the communication constraint vector, with each execution domain corresponding to different communication physical constraints. The task execution topology controller (140) is used to map computation tasks to computation nodes in the corresponding execution domain for execution based on the partitioning results of the execution domain; A closed-loop feedback controller (150) is used to update the partitioning parameters of the execution domain partitioning engine based on the cross-node synchronization error generated during task execution; The execution domain partitioning result is locked as the unique execution location constraint of the computing task within the current scheduling cycle, and the closed-loop feedback controller only acts on the execution domain reconstruction of the next scheduling cycle, thereby forming a cross-cycle irreversible execution domain control architecture.
2. The system according to claim 1, characterized in that, The physical layer telemetry data is acquired through at least one of the following hardware mechanisms: Switch port mirroring; In-band network telemetry protocol; or Network interface card hardware counter.
3. The system according to claim 1, characterized in that, The communication constraint vector includes: Delay parameter is used to represent the time characteristics of a communication link; Bandwidth parameters are used to represent the resource characteristics of a communication link; Energy consumption parameters are used to represent the energy consumption characteristics of a communication link; And the bit error rate parameter, used to represent the signal integrity characteristics of the communication link.
4. The system according to claim 1, characterized in that, The execution domain includes at least: Execution domain based on copper interconnect; Execution domain based on linearly pluggable optics; And execution domains based on co-packaged optics.
5. The system according to claim 1, characterized in that, The execution domain partitioning engine performs partitioning based on a multidimensional similarity metric between the computation task feature vector and the communication constraint vector, wherein the similarity metric includes at least weighted distance calculation.
6. The system according to claim 1, characterized in that, The task feature vector includes at least: The communication-to-computation ratio is used to represent the ratio of communication overhead to computational overhead. And state synchronization granularity, used to represent the frequency and scale of cross-node synchronization.
7. The system according to claim 1, characterized in that, The closed-loop feedback controller updates the execution domain partitioning parameters based on the cross-node synchronization error. The update depends on historical execution error information and is applied to the execution domain partitioning process in subsequent scheduling cycles.
8. The system according to claim 1, characterized in that, The reconstruction of the execution domain is triggered by the change of the communication constraint vector within a continuous time window, the change including link delay, bandwidth or bit error rate exceeding a preset threshold range within a preset time window.
9. The system according to claim 1, characterized in that, The closed-loop feedback controller is configured to cause the cross-node synchronization error to converge to a stable range within a preset scheduling period when communication conditions change.
10. A method for state transition and deterministic execution control across execution domains, characterized in that, Includes the following steps: S1: Obtain physical layer telemetry data of the communication link between multiple computing nodes through the communication monitoring module. The telemetry data includes at least link delay, bandwidth utilization, energy consumption index and bit error rate. S2: Using a constraint vector generator, a communication constraint vector is generated based on the physical layer telemetry data. The communication constraint vector includes at least a delay parameter, a bandwidth parameter, and an energy consumption parameter. S3: By using the execution domain partitioning engine, based on the matching relationship between the computation task feature vector and the communication constraint vector, multiple computing nodes are divided into at least two execution domains, each execution domain corresponding to different communication physical constraints. S4: Through the task execution topology controller, the computation tasks are mapped to the computation nodes in the corresponding execution domain for execution based on the partitioning results of the execution domain; S5: The partitioning parameters of the execution domain partitioning engine are updated based on the cross-node synchronization error generated during task execution through the closed-loop feedback controller; The execution domain partitioning result is locked as the unique execution location constraint of the computing task within the current scheduling cycle, and the closed-loop feedback controller only acts on the execution domain reconstruction of the next scheduling cycle, thereby forming a cross-cycle irreversible execution domain control architecture.