With the steady advance of high performance computing systems featuring smaller and smaller hardware components, the systems and algorithms used for numerical simulations increasingly contend with disruptions caused by hardware failures and bit-levels misrepresentations of computing data. In numerical frameworks exploiting massive processing power, the solution of linear systems often represents the most computationally intensive component. Given the large amount of repeated operations involved, iterative solvers are particularly vulnerable to bit-flips. A new method named FT-GCR is proposed here that supplies the preconditioned Generalized Conjugate Residual Krylov solver with detection of, and recovery from, soft faults. The algorithm tests on the monotonic decrease of the residual norm and, upon failure, restarts the iteration within the local Krylov space. Numerical experiments on the solution of an elliptic problem arising from a stationary flow over an isolated hill on the sphere show the skill of the method in addressing bit-flips on a range of grid sizes and data loss scenarios, with best returns and detection rates obtained for larger corruption events. The simplicity of the method makes it easily extendable to other solvers and an ideal candidate for algorithmic fault tolerance within integrated model resilience strategies.
翻译:随着高性能计算系统以较小硬件组件为主的稳步推进,用于数字模拟的系统和算法日益与硬件故障和计算数据的比特级误差造成的干扰相抗衡。在利用大规模处理能力的数字框架中,线性系统的解决方案往往代表着最计算密集的部分。鉴于涉及的大量重复操作,迭代求解器特别容易受到位翻的伤害。在此建议采用称为FT-GCR的新方法,为通用通用Conjugate剩余Krylov软件提供先决条件的通用Conjugate剩余Krylov软件,以探测和从软故障中回收。该方法的简单化使得其剩余规范的单调降法测试,一旦失败,将重新启动本地Krylov空间的迭代。关于因地表上一个偏僻的山上的固定流动而产生的椭圆问题的解决方案的数值实验显示了在一系列网格大小和数据损失假设中处理小滴问题的方法的技巧,为更大的腐败事件获得最佳回报和检测率。该方法的简单化使其易于推广到其他解算器中,并且成为了一种理想的抗错能力。