As the scale of current systems become larger and larger and their complexity is increasing gradually,research on executable models in the design phase becomes significantly important as it is helpful to simulate the ...As the scale of current systems become larger and larger and their complexity is increasing gradually,research on executable models in the design phase becomes significantly important as it is helpful to simulate the execution process and capture defects of a system in advance.Meanwhile,the capability of a system becomes so important that stakeholders tend to emphasize their capability requirements when developing a system.To deal with the lack of official specifications and the fundamental theory basis for capability requirement,we propose a cooperative capability requirements(CCR)meta-model as a theory basis for researchers to refer to in this research domain,in which we provide detailed definition of the CCR concepts,associations and rules.Moreover,we also propose an executable framework,which may enable modelers to simulate the execution process of a system in advance and do well in filling the inconsistency and semantic gaps between stakeholders’requirements and their models.The primary working mechanism of the framework is to transform the Alf activity meta-model into the communicating sequential process(CSP)process meta-model based on some mapping rules,after which the internal communication mechanism between process nodes is designed to smooth the execution of behaviors in a CSP system.Moreover,a validation method is utilized to check the correctness and consistency of the models,and a self-fixing mechanism is used to fix the errors and warnings captured during the validation process automatically.Finally,a validation report is generated and fed back to the modelers for system optimization.展开更多
随着大规模人工智能应用的普及与发展,工业界和学术界对于人工智能算力的需求逐渐提升,结合了异构计算技术与缓存一致性技术的异构一致性融合计算系统逐渐成为未来构建智算中心的重要解决方案.然而,由于异构计算和一致性互连技术尚不成...随着大规模人工智能应用的普及与发展,工业界和学术界对于人工智能算力的需求逐渐提升,结合了异构计算技术与缓存一致性技术的异构一致性融合计算系统逐渐成为未来构建智算中心的重要解决方案.然而,由于异构计算和一致性互连技术尚不成熟,现有工作难以实现对该系统进行性能建模,导致研究者无法以低成本完成异构一致性融合计算系统的建设方案评估、计算性能预测以及系统优化方法评测等工作.提出了一种面向异构一致性融合计算系统的性能建模工具HCSim,解决了现有建模仿真研究中对该系统拓扑架构建模困难、对一致性系统中工作负载建模不准确等问题,为研究者提供了一个可灵活建模、评估互连拓扑与AI计算任务的低成本、高效建模仿真工具.利用HCSim,建模了千卡互连的异构一致性融合计算系统,并在该系统上模拟了LLAMA2-13B大语言模型(large language model,LLM)的数据并行分布式训练任务,探究了异构算力分布、带宽、时延和任务规模等变量对系统性能与任务执行效率的影响.进一步地,针对异构一致性融合计算系统的通信问题,设计了相应的优化方案,并利用HCSim进行了效果验证.仿真结果说明HCSim不仅能够满足异构一致性融合计算系统的性能建模需求,同时也可以被应用于评估、验证异构一致性融合计算系统的优化方案.展开更多
针对长时间充放电后锂电池模组之间荷电状态(state of charge,SOC)不一致的问题,传统集中式均衡电路存在均衡速度过低的缺陷,以对称式开关阵列、Boost变换器与LC准谐振电路作为均衡主电路,提出了一种基于连续集模型预测控制(continuous ...针对长时间充放电后锂电池模组之间荷电状态(state of charge,SOC)不一致的问题,传统集中式均衡电路存在均衡速度过低的缺陷,以对称式开关阵列、Boost变换器与LC准谐振电路作为均衡主电路,提出了一种基于连续集模型预测控制(continuous control set model predictive control,CCS-MPC)的均衡控制策略。首先,对均衡系统进行建模,构建离散状态空间方程;然后,根据状态方程设计多步模型预测算法,并以SOC预测值和参考值、变换器开关管当前输入和上一时刻输入之间的误差作为价值函数;最后,对价值函数进行二次规划,在线求解出一组控制最优解,并应用于均衡系统,通过动态调整占空比以控制均衡电流的大小。相较于单步预测,多步预测需要考虑被控量在多个周期内保持最优,可以保证在每个均衡周期内均衡器都能输出最优的均衡电流,有效防止均衡器失稳。仿真结果表明,所提模型预测算法实现了各电池组SOC一致,保证了均衡电流的稳定输出,相比常规PI算法缩短了17%的均衡时间。展开更多
This paper examined how microstructure influences the homogenized thermal conductivity of cellular structures and revealed a surface-induced size-dependent effect.This effect is linked to the porous microstructural fe...This paper examined how microstructure influences the homogenized thermal conductivity of cellular structures and revealed a surface-induced size-dependent effect.This effect is linked to the porous microstructural features of cellular structures,which stems from the degree of porosity and the distri-bution of the pores.Unlike the phonon-driven surface effect at the nanoscale,the macro-scale surface mechanism in thermal cellular structures is found to be the microstructure-induced changes in the heat conduction path based on fully resolved 3D numerical simulations.The surface region is determined by the microstructure,characterized by the intrinsic length.With the coupling between extrinsic and intrinsic length scales under the surface mechanism,a surface-enriched multiscale method was devel-oped to accurately capture the complex size-dependent thermal conductivity.The principle of scale separation required by classical multiscale methods is not necessary to be satisfied by the proposed multiscale method.The significant potential of the surface-enriched multiscale method was demon-strated through simulations of the effective thermal conductivity of a thin-walled metamaterial struc-ture.The surface-enriched multiscale method offers higher accuracy compared with the classical multiscale method and superior efficiency over high-fidelity finite element methods.展开更多
基金supported by the National Key R&D Program of China(2018YFC0806900)the China Postdoctoral Science Foundation Funded Project(2018M633757)+1 种基金the Primary Research&Development Plan of Jiangsu Province(BE2016904,BE2017616,BE2018754,BE2019762)Jiangsu Province Postdoctoral Science Foundation Funded Project(2019K185).
文摘As the scale of current systems become larger and larger and their complexity is increasing gradually,research on executable models in the design phase becomes significantly important as it is helpful to simulate the execution process and capture defects of a system in advance.Meanwhile,the capability of a system becomes so important that stakeholders tend to emphasize their capability requirements when developing a system.To deal with the lack of official specifications and the fundamental theory basis for capability requirement,we propose a cooperative capability requirements(CCR)meta-model as a theory basis for researchers to refer to in this research domain,in which we provide detailed definition of the CCR concepts,associations and rules.Moreover,we also propose an executable framework,which may enable modelers to simulate the execution process of a system in advance and do well in filling the inconsistency and semantic gaps between stakeholders’requirements and their models.The primary working mechanism of the framework is to transform the Alf activity meta-model into the communicating sequential process(CSP)process meta-model based on some mapping rules,after which the internal communication mechanism between process nodes is designed to smooth the execution of behaviors in a CSP system.Moreover,a validation method is utilized to check the correctness and consistency of the models,and a self-fixing mechanism is used to fix the errors and warnings captured during the validation process automatically.Finally,a validation report is generated and fed back to the modelers for system optimization.
文摘随着大规模人工智能应用的普及与发展,工业界和学术界对于人工智能算力的需求逐渐提升,结合了异构计算技术与缓存一致性技术的异构一致性融合计算系统逐渐成为未来构建智算中心的重要解决方案.然而,由于异构计算和一致性互连技术尚不成熟,现有工作难以实现对该系统进行性能建模,导致研究者无法以低成本完成异构一致性融合计算系统的建设方案评估、计算性能预测以及系统优化方法评测等工作.提出了一种面向异构一致性融合计算系统的性能建模工具HCSim,解决了现有建模仿真研究中对该系统拓扑架构建模困难、对一致性系统中工作负载建模不准确等问题,为研究者提供了一个可灵活建模、评估互连拓扑与AI计算任务的低成本、高效建模仿真工具.利用HCSim,建模了千卡互连的异构一致性融合计算系统,并在该系统上模拟了LLAMA2-13B大语言模型(large language model,LLM)的数据并行分布式训练任务,探究了异构算力分布、带宽、时延和任务规模等变量对系统性能与任务执行效率的影响.进一步地,针对异构一致性融合计算系统的通信问题,设计了相应的优化方案,并利用HCSim进行了效果验证.仿真结果说明HCSim不仅能够满足异构一致性融合计算系统的性能建模需求,同时也可以被应用于评估、验证异构一致性融合计算系统的优化方案.
文摘针对长时间充放电后锂电池模组之间荷电状态(state of charge,SOC)不一致的问题,传统集中式均衡电路存在均衡速度过低的缺陷,以对称式开关阵列、Boost变换器与LC准谐振电路作为均衡主电路,提出了一种基于连续集模型预测控制(continuous control set model predictive control,CCS-MPC)的均衡控制策略。首先,对均衡系统进行建模,构建离散状态空间方程;然后,根据状态方程设计多步模型预测算法,并以SOC预测值和参考值、变换器开关管当前输入和上一时刻输入之间的误差作为价值函数;最后,对价值函数进行二次规划,在线求解出一组控制最优解,并应用于均衡系统,通过动态调整占空比以控制均衡电流的大小。相较于单步预测,多步预测需要考虑被控量在多个周期内保持最优,可以保证在每个均衡周期内均衡器都能输出最优的均衡电流,有效防止均衡器失稳。仿真结果表明,所提模型预测算法实现了各电池组SOC一致,保证了均衡电流的稳定输出,相比常规PI算法缩短了17%的均衡时间。
基金supported by the National Key Research and Development Program of China(Grant No.2021YFB1714600)the National Natural Science Foundation of China(Grant No.52175095)the Young Top-Notch Talent Cultivation Program of Hubei Province of China.
文摘This paper examined how microstructure influences the homogenized thermal conductivity of cellular structures and revealed a surface-induced size-dependent effect.This effect is linked to the porous microstructural features of cellular structures,which stems from the degree of porosity and the distri-bution of the pores.Unlike the phonon-driven surface effect at the nanoscale,the macro-scale surface mechanism in thermal cellular structures is found to be the microstructure-induced changes in the heat conduction path based on fully resolved 3D numerical simulations.The surface region is determined by the microstructure,characterized by the intrinsic length.With the coupling between extrinsic and intrinsic length scales under the surface mechanism,a surface-enriched multiscale method was devel-oped to accurately capture the complex size-dependent thermal conductivity.The principle of scale separation required by classical multiscale methods is not necessary to be satisfied by the proposed multiscale method.The significant potential of the surface-enriched multiscale method was demon-strated through simulations of the effective thermal conductivity of a thin-walled metamaterial struc-ture.The surface-enriched multiscale method offers higher accuracy compared with the classical multiscale method and superior efficiency over high-fidelity finite element methods.
