随着人工智能大模型训练与推理业务的快速发展,智算中心面临算网协同调度的新挑战。为优化分布式智算Ring Allreduce业务的部署问题,首先,通过扩展传统波平面,开发了算力波平面技术,实现了算力与网络资源的一体化虚拟管理。然后,基于算...随着人工智能大模型训练与推理业务的快速发展,智算中心面临算网协同调度的新挑战。为优化分布式智算Ring Allreduce业务的部署问题,首先,通过扩展传统波平面,开发了算力波平面技术,实现了算力与网络资源的一体化虚拟管理。然后,基于算力波平面,提出了一种高效路由、波长、算力和时隙分配(routing,wavelength,computing power and time slot assignment,RWCTA)算法用于环规约(Ring Allreduce)业务部署。仿真结果表明,与传统基于波平面的部署算法相比,基于算力波平面的RWCTA算法能有效降低62.4%的总业务完成时间和54.5%的平均业务计算时间。展开更多
The evolution of threats and scenarios requires continuous performance improvements of ballistic protections for armed forces.From a modeling point of view,it is necessary to use sufficiently precise material behavior...The evolution of threats and scenarios requires continuous performance improvements of ballistic protections for armed forces.From a modeling point of view,it is necessary to use sufficiently precise material behavior models to accurately describe the phenomena observed during the impact of a projectile on a protective equipment.In this context,the goal of this paper is to characterize the behavior of a small caliber steel jacket by combining experimental and numerical approaches.The experimental method is based on the lateral compression of ring specimens directly machined from the thin and small ammunition.Various speeds and temperatures are considered in a quasi-static regime in order to reveal the strain rate and temperature dependencies of the tested material.The Finite Element Updating Method(FEMU)is used.Experimental results are coupled with an inverse optimization method and a finite element numerical model in order to determine the parameters of a constitutive model representative of the jacket material.Predictions of the present model are verified against experimental results and a parametric study as well as a discussion on the identified material parameters are proposed.The results indicate that the strain hardening parameter can be neglected and the behavior of the thin steel jacket can be described by a modeling without strain hardening sensitivity.展开更多
文摘随着人工智能大模型训练与推理业务的快速发展,智算中心面临算网协同调度的新挑战。为优化分布式智算Ring Allreduce业务的部署问题,首先,通过扩展传统波平面,开发了算力波平面技术,实现了算力与网络资源的一体化虚拟管理。然后,基于算力波平面,提出了一种高效路由、波长、算力和时隙分配(routing,wavelength,computing power and time slot assignment,RWCTA)算法用于环规约(Ring Allreduce)业务部署。仿真结果表明,与传统基于波平面的部署算法相比,基于算力波平面的RWCTA算法能有效降低62.4%的总业务完成时间和54.5%的平均业务计算时间。
基金co-funded by the Direction Générale de l'Armement (DGA)the French-German Institute of Saint Louis (ISL)。
文摘The evolution of threats and scenarios requires continuous performance improvements of ballistic protections for armed forces.From a modeling point of view,it is necessary to use sufficiently precise material behavior models to accurately describe the phenomena observed during the impact of a projectile on a protective equipment.In this context,the goal of this paper is to characterize the behavior of a small caliber steel jacket by combining experimental and numerical approaches.The experimental method is based on the lateral compression of ring specimens directly machined from the thin and small ammunition.Various speeds and temperatures are considered in a quasi-static regime in order to reveal the strain rate and temperature dependencies of the tested material.The Finite Element Updating Method(FEMU)is used.Experimental results are coupled with an inverse optimization method and a finite element numerical model in order to determine the parameters of a constitutive model representative of the jacket material.Predictions of the present model are verified against experimental results and a parametric study as well as a discussion on the identified material parameters are proposed.The results indicate that the strain hardening parameter can be neglected and the behavior of the thin steel jacket can be described by a modeling without strain hardening sensitivity.
