Core Concepts
FastFlip enables efficient error injection analysis for evolving programs, optimizing protection against Silent Data Corruptions.
Abstract
FastFlip introduces a novel approach combining empirical error injection and symbolic SDC propagation analyses.
The method allows for compositional and incremental error analysis of evolving software programs.
By selecting static instructions to protect against SDCs, FastFlip minimizes runtime costs while ensuring effective protection.
The study evaluates the utility of protecting selected instructions using various metrics across different benchmarks.
FastFlip demonstrates significant speedups in analysis time when compared to traditional approaches.
The methodology involves sub-analyses such as Approxilyzer for error injection and Chisel for SDC propagation.
Detailed experiments on benchmarks like LU decomposition showcase the effectiveness and efficiency of FastFlip.
I. INTRODUCTION
Error-prone hardware necessitates protection against Silent Data Corruptions (SDCs).
Software techniques like instruction duplication offer viable solutions but require selective application.
II. BACKGROUND
Error injection analyses simulate errors to assess their impact on program outputs.
SDC propagation analyses calculate the effects of SDCs on final program outputs.
III. EXAMPLE
Illustrates how LU decomposition can be vulnerable to hardware errors causing SDCs.
IV. THE FASTFLIP APPROACH
A. Preliminaries
Inputs, data flow specifications, and target values are crucial for FastFlip's operation.
B. Error injection analysis of program sections
C. SDC propagation analysis of program sections
D. End-to-end SDC propagation analysis
Stats
FastFlipはプログラムの進化に対する効率的なエラーインジェクション分析を可能にします。
FastFlipは、経験的なエラーインジェクションと象徴的SDC伝播分析を組み合わせた新しいアプローチを導入します。