Center-free adaptive distributed joint simulation scheduling system

Abstract

A centerless adaptive distributed joint simulation scheduling system comprises a plurality of local scheduling domains, data communication among the local scheduling domains is realized through'inter-domain interaction interfaces', each local scheduling domain comprises a local master node and a plurality of local slave nodes forming a subsystem, the local master node communicates with the local slave nodes, and the local slave nodes communicate with the local master node. The method is used for automatically carrying out local master and slave node dynamic division, intelligent model distribution based on calculation complexity, time sequence scheduling of self-adaptive propulsion step length, step length optimization and data interaction based on the least common multiple, and subsequent simulation data prediction based on a fault breakpoint.

Description

Technical Field

[0001] This invention relates to simulation scheduling systems, and more particularly to a decentralized adaptive distributed joint simulation scheduling system, belonging to the field of simulation scheduling. Background Technology

[0002] For joint simulation systems of large-scale equipment (such as aerospace equipment, large ships and equipment), in core application scenarios characterized by "large model scale, significant differences in simulation scale, and complex subsystem coupling," existing centralized scheduling schemes and simple distributed scheduling schemes cannot coordinate the technical pain points of "high-precision timing synchronization, high-performance simulation advancement, and high system reliability." Specifically, centralized scheduling suffers from a sharp drop in simulation efficiency due to overload of the global central master node, and the fixed step size cannot adapt to the needs of multi-scale models. At the same time, the failure of the central node can easily lead to system interruption. Simple distributed scheduling, on the other hand, suffers from node load imbalance due to static pre-allocation, waiting and blocking caused by fixed data interaction nodes, and a high task failure rate due to the lack of a dynamic fault takeover mechanism. Ultimately, both schemes fall into the inherent contradiction of "sacrificing efficiency to ensure synchronization" or "disrupting synchronization to pursue efficiency," failing to meet the comprehensive requirements of "efficient advancement, accurate synchronization, and stable reliability" for joint simulation systems in the development of large-scale equipment, thus restricting the efficiency of equipment development and testing accuracy.

[0003] Furthermore, both schemes handle faults differently. The centralized distributed master node monitors the status of computing nodes via "heartbeat packets" (sent every 100ms), but lacks redundancy. If the master node fails (e.g., hardware crash, software failure), the system is directly interrupted, requiring manual restart of the master node and reloading of simulation data; the recovery process relies on manual operation. The simple distributed scheme employs a stateless monitoring and dynamic takeover mechanism between nodes. If a node fails (e.g., network disconnection, computational anomaly), the modules it is responsible for immediately stop, and related nodes remain blocked due to the inability to receive data, ultimately leading to simulation task failure. Therefore, how to handle faults and achieve seamless fault continuity is another pressing technical problem that needs to be solved. Summary of the Invention

[0004] 1. The core technology of this invention First, the decentralized architecture distributes computing and communication loads. Second, dynamic step size adaptation of multi-scale simulation models. Third, by combining the dynamic changes in the model's computational complexity with the coupling strength of the subsystems, distributed node load balancing is achieved, coordinating the timing synchronization and efficient advancement of distributed nodes. Fourth, simulate the fault node for the faulty module, and embed the simulation results into the above-mentioned decentralized system in real time until the fault repair time node.

[0005] 2. Specific technical solutions To achieve the core of the above technical solution and address the dilemmas of existing technologies, this invention provides a decentralized adaptive distributed joint simulation scheduling system, comprising multiple local scheduling domains. These local scheduling domains communicate with each other via an "inter-domain interaction interface." Each local scheduling domain includes one local master node and several local slave nodes that make up the subsystem. The local master node communicates with the local slave nodes to automatically perform dynamic partitioning of local master and slave nodes, intelligent model allocation based on computational complexity, time-series scheduling with adaptive step size, step size optimization based on least common multiple and data interaction, and prediction of subsequent simulation data based on fault breakpoints.

[0006] Optionally, the method of dynamically dividing local master-slave nodes is based on the coupling strength between the models of the system and subsystems and the difference in simulation scale, and dynamically divides local master-slave nodes, specifically including coupling relationship determination and scale adaptation division.

[0007] Optionally, the coupling relationship determination includes: quantifying the degree of coupling between models through a "coupling strength calculation model", with input parameters including the frequency of data interaction between models. Unit: times / second, data dependency The value ranges from 0 to 1, and is determined by the correlation between the model's input and output, and the amount of interactive data. Unit: MB / time, the coupling strength calculation formula is as follows:

[0008] Where α, β, and γ are weighting coefficients.

[0009] Optionally, scale adaptation partitioning: Define the simulation scale matching degree within the same local scheduling domain. (k is the node number), the calculation formula is as follows:

[0010] in, This represents the total number of models within the local scheduling domain. Let k be the set of models that node k can support. The simulation scale of model m, in seconds.

[0011] Optionally, the intelligent model allocation based on computational complexity includes Model computational complexity assessment: Establish a "quantitative index system for model computational complexity," and calculate the complexity weight of each model from three dimensions: "computational load," "memory requirements," and "time sensitivity," using the Analytic Hierarchy Process (AHP).

[0012]

[0013] in, Weights determined for AHP Node resource status awareness: Each distributed node reports its current resource utilization (CPU utilization) to its local master node in real time. Memory usage Network bandwidth utilization The formula for calculating resource redundancy is as follows:

[0014] Intelligent allocation algorithm execution: Local master nodes are allocated based on the "complexity-resource matching principle" model. like (Lightweight model) and If node resource redundancy meets the requirements, then the model will be assigned to a single node; like (Heavyweight model) or The "module splitting algorithm" is then used to split the model into K sub-modules, based on the internal functional boundaries of the model, and assigned to different local slave nodes.

[0015] The timing scheduling method with adaptive propagation step size specifically includes: Dynamic time-series progression table establishment: Based on the coupling relationship between models in the domain, the local master node generates a "dynamic time-series progression table" to identify data interaction points between models (i.e., the time points when the output of one model is the input of another model) and defines the interaction strength. ,when At that time, mark the time node as a "keyframe node". ”; The timing progression table records the requirements for issuing progression commands for key frame nodes, including... The timestamps, the model pairs that need to be interacted with, and the data interaction format are all determined by the local slave nodes. The advancement of non-critical frame nodes is autonomously controlled by the local slave nodes. Master-slave node collaborative advancement: including local master nodes: in "Data exchange and advance instructions" are constantly issued to local slave nodes. These instructions include... The time base and interactive data verification rules are established. After receiving instructions from the slave node, the current simulation results are uploaded to the master node. The master node verifies the data validity through the "data consistency verification algorithm." After the verification is passed, the next round of keyframe nodes is confirmed. Model interaction requirements are dynamically determined, including local slave nodes: and Between, according to the simulation scale of its own model Choose step size independently , Must meet and , To advance the number of simulations, multi-step continuous simulations are performed; at the same time, the multiple models that a node is responsible for are sorted according to time sensitivity priority.

[0016] Preferably, priority Prioritize A high-precision model ensures the simulation accuracy of a highly sensitive model.

[0017] Inter-domain timing coordination: Different local scheduling domains achieve timing synchronization through an "inter-domain keyframe alignment" mechanism, defining the inter-domain synchronization degree. (a and b are adjacent domains):

[0018] when At that time, the inter-domain interaction coordination node (determined through negotiation between the two domain master nodes) adjusts the keyframe node to make... At the same time, ensure the consistency of data interaction timing between domains, whereby, It is the greatest common divisor.

[0019] Distinguishing between intra-node and inter-node interactions: like and Within the same local slave node, the slave node is ordered as follows: It can complete data interaction on its own, and the interaction process is realized through "intra-node shared memory". There is no need to report to the local master node, and the communication latency is reduced to less than 1ms. like and In different nodes, the local master node is responsible for... Data interaction commands are triggered at all times, and the commands are transmitted through a "low-latency communication protocol" to ensure that the interaction latency is ≤10ms.

[0020] Optionally, methods for predicting subsequent simulation data based on the fault breakpoint include: Methods for predicting subsequent simulation data based on fault breakpoints include: The S1 search module retrieves the second historical data from the first data point at a given time point under the same operating parameters, up to the time point of fault recovery. S2 divides all data between the first and second historical data into multiple time nodes according to the time-series scheduling of the adaptive advancement step size, acquires the data at the multiple time nodes, and finally performs step size optimization based on the least common multiple to train the long short-term memory model. Based on the trained long short-term memory model, subsequent simulation data predicted after the first historical data are used for data interaction.

[0021] The operating parameter environment refers to the operating data environment of modules in the simulation system other than the fault module.

[0022] 3. Beneficial effects

[0023] Solving the bottlenecks of traditional centralized scheduling centers and the difficulty of synchronization in simple distributed systems: This invention features a decentralized local master-slave architecture. By dynamically dividing local scheduling domains (based on the matching degree between model coupling strength and simulation scale), master nodes within a domain coordinate scheduling and inter-domain collaborative interaction. This not only distributes the computing and communication load (avoiding the risk of overload in centralized global centers) but also ensures synchronization within the domain through local master nodes (solving the problem of difficulty in synchronization in simple distributed systems with complete dispersion).

[0024] Improving the load imbalance problem of simple distributed static allocation: The multi-dimensional intelligent model allocation mechanism of this invention quantifies the model complexity based on "computational load - memory requirements - time sensitivity", combines the real-time resource status allocation model of nodes, and also supports dynamic migration when the load is too high, avoiding the uneven node load caused by simple distributed static pre-allocation and improving resource utilization.

[0025] Overcoming the adaptability defects of existing fixed step size / synchronization point solutions: The key frame driven adaptive timing scheduling of this invention identifies key frame nodes through interaction intensity (the master node only controls synchronization for key frames) and autonomously selects the step size from non-key frames of the node. This not only adapts to the step size requirements of multi-scale models (solving the accuracy / resource waste problem of centralized fixed step size), but also avoids node waiting and blocking caused by simple distributed fixed synchronization points.

[0026] Optimize the problems of one-size-fits-all interaction step size and communication latency in existing technologies: The LCM step size optimization and differentiated interaction mechanism of this invention determines the exchange step size according to the least common multiple of the autonomous step size of the interaction model, and uses shared memory within nodes and low-latency protocols between nodes, which not only ensures the timing matching of data interaction, but also reduces communication latency and improves synchronization accuracy.

[0027] Resolving the inherent contradiction between "synchronization and efficiency": This invention integrates the above-mentioned architecture, allocation, scheduling, and interaction mechanisms. The synergistic effect of multiple mechanisms breaks through the common problems of existing technologies that "sacrifice efficiency for synchronization" or "destroy synchronization for efficiency", and achieves a synergistic improvement in simulation efficiency and timing synchronization. Attached Figure Description

[0028] Figure 1 Adaptive region partitioning diagram, Figure 2 A schematic diagram illustrating a method for predicting subsequent simulation data based on fault breakpoints. Detailed Implementation

[0029] Figure 1The paper presents a decentralized adaptive distributed co-simulation scheduling system with different regional distribution states before and after adaptive regional division. Taking region A in the figure as an example, it has a local master node and other local child nodes to form a subsystem.

[0030] Multiple local scheduling domains communicate with each other through an "inter-domain interaction interface". Each local scheduling domain contains one local master node and several local slave nodes that make up the subsystem. The local master node communicates with the local slave nodes to automatically perform dynamic partitioning of local master and slave nodes, intelligent model allocation based on computational complexity, time-series scheduling with adaptive step size, step size optimization and data interaction based on least common multiple, and prediction of subsequent simulation data based on fault breakpoints.

[0031] The method of dynamically partitioning local master-slave nodes is based on the coupling strength between the models of the system and subsystems and the difference in simulation scale. It dynamically partitions local master-slave nodes, specifically including coupling relationship determination and scale adaptation partitioning.

[0032] The coupling relationship determination includes: quantifying the degree of coupling between models through a "coupling strength calculation model", with input parameters including the frequency of data interaction between models. Unit: times / second, data dependency The value ranges from 0 to 1, and is determined by the correlation between the model's input and output, and the amount of interactive data. Unit: MB / time, the coupling strength calculation formula is as follows:

[0033] Where α, β, and γ are weight coefficients, α + β + γ = 1, determined by the analytic hierarchy process (AHP), with α = 0.4, β = 0.3, and γ = 0.3; max(F) is the maximum interaction frequency of all model pairs, and max(V) is the maximum amount of interaction data of all model pairs. Model i and model j are grouped into the same local scheduling domain.

[0034] Optionally, scale adaptation partitioning: Define the simulation scale matching degree within the same local scheduling domain. (k is the node number), the calculation formula is as follows:

[0035] in, This represents the total number of models within the local scheduling domain. Let k be the set of models that node k can support. The simulation scale of model m, in seconds. (Select) And calculate resource redundancy The node is designated as the local master node, and the remaining nodes are designated as local slave nodes; during the simulation, if the coupling relationship of the subsystem or the model scale changes (e.g., ...), the node is designated as the local master node, and the remaining nodes are designated as local slave nodes. Drop to below 0.5 or If the value drops below 0.6, the local scheduling domain and master-slave nodes will be automatically re-divided.

[0036] The computationally complex-based intelligent model allocation mechanism includes: Model computational complexity assessment: Establish a "quantitative index system for model computational complexity," and calculate the complexity weight of each model from three dimensions: "computational load," "memory requirements," and "time sensitivity," using the Analytic Hierarchy Process (AHP).

[0037] in, For the weights determined by AHP, take , , , .

[0038] a. Computational workload indicators Measured by the number of instructions executed in a single simulation (unit: million instructions, MIPS), the standardized formula is: ; b. Memory requirement indicators Measured by the memory space required for model execution (unit: GB), the standardized formula is: ; c. Time sensitivity index Measured as the reciprocal of the maximum allowable simulation delay (in milliseconds), the standardized formula is: .

[0039] Node resource status awareness: Each distributed node reports its current resource utilization (CPU utilization) to its local master node in real time. Memory usage Network bandwidth utilization The formula for calculating resource redundancy is as follows:

[0040] Intelligent allocation algorithm execution: Local master nodes are allocated based on the "complexity-resource matching principle" model. like (Lightweight model) and If node resource redundancy meets the requirements, then the model will be assigned to a single node; like (Heavyweight model) or Then, the "module splitting algorithm" is used to split the model into K sub-modules. The average resource redundancy of slave nodes within the domain is determined by splitting the submodules based on the functional boundaries within the model, ensuring the coupling strength between submodules. And assign them to different local slave nodes.

[0041] During the simulation, the local master node is recalculated every 5 minutes (configurable). and ,right or Nodes with excessive load will be dynamically migrated using the model. The target node to be migrated must meet certain requirements. And between submodules To achieve resource balance.

[0042] The timing scheduling method with adaptive propagation step size specifically includes: Dynamic time-series progression table establishment: Based on the coupling relationship between models in the domain, the local master node generates a "dynamic time-series progression table" to identify data interaction points between models (i.e., the time points when the output of one model is the input of another model) and defines the interaction strength. ,when At that time, mark the time node as a "keyframe node". ”; The timing progression table records the requirements for issuing progression commands for key frame nodes, including... The timestamps, the model pairs that need to be interacted with, and the data interaction format are all determined by the local slave nodes. The advancement of non-critical frame nodes is autonomously controlled by the local slave nodes. Master-slave node collaborative advancement: including local master nodes: in "Data exchange and advance instructions" are constantly issued to local slave nodes. These instructions include... The time base and interactive data verification rules are established. After receiving instructions from the slave node, the current simulation results are uploaded to the master node. The master node verifies the data validity through the "data consistency verification algorithm." After the verification is passed, the next round of keyframe nodes is confirmed. , Model interaction requirements are dynamically determined, including local slave nodes: and Between, according to the simulation scale of its own model Choose step size independently , Must meet and , To advance the number of simulations, multi-step continuous simulations are performed; at the same time, the multiple models that a node is responsible for are sorted according to time sensitivity priority.

[0043] Priority Prioritize A high-precision model ensures the simulation accuracy of a highly sensitive model.

[0044] Inter-domain timing coordination: Different local scheduling domains achieve timing synchronization through an "inter-domain keyframe alignment" mechanism, defining the inter-domain synchronization degree. (a and b are adjacent domains):

[0045] when At that time, the inter-domain interaction coordination node (determined through negotiation between the two domain master nodes) adjusts the keyframe node to make... At the same time, ensure that the timing of data interaction between domains is consistent.

[0046] Step size optimization and data interaction based on least common multiple include: Interaction step size calculation: For any two models that have data interaction and Let their autonomous propulsion step lengths be respectively and The local master node calculates the least common multiple of the two. Use it as a data exchange frame node Right now:

[0047] in, The greatest common divisor; Distinguishing between intra-node and inter-node interactions: like and Within the same local slave node, the slave node is ordered as follows: It can complete data interaction on its own, and the interaction process is realized through "intra-node shared memory". There is no need to report to the local master node, and the communication latency is reduced to less than 1ms. like and In different nodes, the local master node is responsible for... Data interaction commands are triggered at all times, and the commands are transmitted through a "low-latency communication protocol" to ensure that the interaction latency is ≤10ms.

[0048] like Figure 2 Methods for predicting subsequent simulation data based on fault breakpoints include: S1 searches (on the timeline) the fault module for the second historical data after the first data point at a given time point under the same operating parameters, up to the fault recovery time point. S2 divides all data between the first and second historical data into multiple time nodes according to the time-series scheduling of the adaptive advancement step size, acquires the data at the multiple time nodes, and finally performs step size optimization based on the least common multiple to train the long short-term memory model. Based on the trained long short-term memory model, subsequent simulation data predicted after the first historical data are used for data interaction.

Claims

1. A central-less adaptive distributed joint simulation scheduling system, characterized in that, The system comprises a plurality of local scheduling domains, the local scheduling domains are in data communication through an "inter-domain interaction interface", each local scheduling domain comprises a local master node and a plurality of local slave nodes constituting a subsystem, the local master node communicates with the local slave nodes, and is used for automatically performing local master-slave node dynamic division, intelligent model allocation based on calculation complexity, timing scheduling of adaptive advancing step, step optimization and data interaction based on least common multiple, and prediction of subsequent simulation data based on a fault breakpoint.

2. The system of claim 1, wherein, The method for local master-slave node dynamic division is based on the coupling strength relationship between the system and the subsystem and the simulation scale difference, and dynamically divides the local master-slave nodes, and specifically includes coupling relationship judgment and scale adaptation division.

3. The system of claim 1, wherein, The coupling relationship determination includes: quantifying the degree of inter-model coupling through a "coupling strength calculation model", input parameters include inter-model data interaction frequency , unit: times / second, data dependency , value 0-1, determined by model input-output correlation, interaction data volume , unit: MB / time, coupling strength calculation formula as follows: ， Wherein, α, β, γ are weight coefficients, α+β+γ=1, determined by the analytic hierarchy process, α=0.4, β=0.3, γ=0.3, max(F) is the maximum interaction frequency of all model pairs, max(V) is the maximum interaction data volume of all model pairs, When model i and model j are grouped into the same local dispatch domain.

4. The system of claim 3, wherein, Scale adaptation division: in the same local scheduling domain, the simulation scale matching degree is defined k is the node number, and the calculation formula is as follows: ， wherein, is the total number of local scheduling domains, is the set of models that node k can host, is the simulation scale of model m, in seconds, selected from and the redundancy of computing resources , the node with the largest is the local master node and the rest are local slave nodes; during the simulation, if the coupling relationship of the subsystem or the model scale changes, the local scheduling domain and the master-slave nodes are automatically re-divided.

5. The system of claim 3 or 4, wherein, The intelligent model allocation mechanism based on calculation complexity includes: Model calculation complexity evaluation: a "model calculation complexity quantitative index system" is established, from three dimensions of "calculation amount", "memory requirement" and "time sensitivity", the complexity weight of each model is calculated by the analytic hierarchy process (AHP) ， wherein The weight determined for AHP is taken , , , ; a. Compute the metrics : the number of instructions executed in a single simulation of the model, in millions of instructions, MIPS, measured using the formula ; b. Memory requirement index : Memory space required for model running, unit: GB, measurement, standardization formula is ; c. Time sensitivity indicator : reciprocal of the maximum simulation delay allowed by the model, in milliseconds, measured in normalized formula ; Node resource state awareness: Each distributed node reports the "current resource utilization" to the local master node in real time, that is, CPU usage , memory usage , network bandwidth utilization , and "resource redundancy", the formula for calculating resource redundancy is as follows: ， Intelligent allocation algorithm execution: the local master node allocates models based on the "complexity-resource matching principle": If and then assign the model to a single node; If or then the model is split into K sub-modules using the "module splitting algorithm". To average resource redundancy of nodes in the domain, split according to the model internal function boundary, ensure the coupling strength between sub-modules , and allocate to different local slave nodes.

6. The system of claim 5, wherein, During the simulation process, the local master node recalculates every 5 minutes and ,right or Nodes with excessive load will be dynamically migrated using the model. The target node to be migrated must meet certain requirements. And between submodules To achieve resource balance.

7. The system of claim 6, wherein, The timing scheduling method of adaptive advancing step specifically includes: Dynamic timing advancement table establishment: the local master node generates a "dynamic timing advancement table" to identify the data interaction points between the models based on the coupling relationship between the master-slave nodes in the domain, and defines the interaction intensity When , the time node is marked as a "key frame node "; The time sequence advancing table records the advancing instruction issuing requirements of the key frame node, including The timestamp, model pair to be interacted, data interaction format, and the advancing of the non-key frame node is autonomously controlled by the local slave node. Master-slave node cooperation: including local master node: in the moment to local slave node issued "data exchange and promotion instructions", the instruction contains time reference, interactive data verification rules; after receiving the instruction, the slave node uploads the current simulation result to the master node, and the master node verifies the data validity through "data consistency verification algorithm". After verification, the next round of key frame node model interaction demand is dynamically determined, and the local slave node includes: between and , according to the simulation scale of its own model , need to meet and , is the number of times to promote, and multiple-step continuous simulation is performed; at the same time, the slave node sorts the multiple models responsible for itself according to the time sensitivity priority, and the priority , preferentially promote high sensitivity model, ensure the simulation accuracy of high sensitivity model; Inter-domain timing coordination: Different local scheduling domains achieve timing synchronization through the "inter-domain key frame alignment" mechanism, and define the inter-domain synchronization degree , a, b are adjacent domains: , When , the inter-domain interaction coordination node adjusts the key frame node, so that the inter-domain data interaction timing is consistent.

8. The system of claim 6 or 7, wherein, The step optimization and data interaction based on the least common multiple includes: Interaction step length calculation: for any two models that exist data interaction With , let its autonomous propulsion step length be With , the local master node calculates the least common multiple of the two , which is used as the data exchange frame node That is: , wherein is the greatest common divisor; Intra-node and inter-node interaction distinction: like and Within the same local slave node, the slave node is ordered as follows: It can complete data interaction on its own, and the interaction process is realized through "intra-node shared memory". There is no need to report to the local master node, and the communication latency is reduced to less than 1ms. If With At different nodes, the local master node triggers data interaction instructions at the moment, and the instructions pass through "low power" Delay communication protocol transmission, ensuring that the interaction delay is ≤10ms.

9. The system of claim 8, wherein, The method for predicting subsequent simulation data based on a fault breakpoint includes: S1 searches the first data after the time node of the fault module in the history under the same running parameter environment, and the second historical data from the time node when the fault is recovered, S2 divides all data between the first and second historical data into a corresponding plurality of time nodes according to the timing scheduling of adaptive advancing step, obtains the data on the plurality of time nodes, finally performs step optimization based on the least common multiple, trains the long short-term memory model, and performs data interaction on the subsequent simulation data predicted after the first historical data based on the trained long short-term memory model.

10. The system of claim 9, wherein, The running parameter environment refers to the running data environment of the modules in the simulation system except the fault module.

Images