【作者】 姜楠；

【导师】 魏晓辉；

【作者基本信息】 吉林大学 ， 计算机系统结构， 2022， 硕士

【摘要】 通用图形处理器(general purpose Graphics Processing Units,GPGPUs)由于其高并发性、高吞吐量以及不断提升的可编程能力,被广泛应用于高性能计算中心。制造工艺的发展导致芯片尺寸缩小、集成度提高,这会增加GPGPU平台受到高能粒子撞击产生软错误的概率。软错误可能会导致程序产生静默数据损坏(Silent Data Corruptions,SDCs),这是最难以检测的错误类型,因为其会影响输出的准确性,但却没有明确的迹象表明程序执行过程中产生了异常。因此,为了保证GPGPU程序的可靠性执行,亟需提出一种有效的SDC检测方法。全指令复制是检测SDC的一种常用方法,其通过指令冗余执行来判断是否有软错误发生。然而在GPGPU上实施全指令复制面临着以下挑战。对于每个线程来说,指令复制会额外增加其执行指令的数量,从而延长程序执行时间。此外,指令复制会占用额外的寄存器来存储副本数据,单个线程所需寄存器资源因此增加,这会影响线程的并发度,进而影响程序的性能。对此,现有工作提出了选择性指令复制的思想,即选择性地保护程序中容易产生SDC的指令。现有工作通常采用故障注入方法来分析指令的软错误弹性,从而选择出SDC倾向性较高的指令进行保护。然而为了准确判断指令的SDC倾向性,往往需要详尽的故障注入,这个过程通常是十分耗时的。此外,以往的工作往往对程序提供了过度保护。考虑到一些应用程序对轻微的SDC错误具有一定的容忍性,可以根据用户的精度需求进一步放松指令保护条件,即仅保护易产生严重SDC错误的指令。面向上述问题和挑战,本文提出了一种高能效的选择性指令复制方法,具体工作如下:1.提出一种基于机器学习模型的SDC倾向性指令预测方法,用以高效地甄别程序中需要保护的指令。本文发现指令本身属性、指令功能以及错误传播过程中的特征能有效地表征指令的SDC倾向性,并利用机器学习分类器探索这些特征与指令SDC倾向性之间的关系。通过对少量指令进行故障注入,为机器学习模型提供训练集,能够以较低的时间开销预测程序中所有的SDC倾向性指令。2.考虑到应用程序固有的错误容忍性和用户的精度要求,可以放宽对非SDC严重性指令的保护,在合理的范围内进一步地降低可靠性开销。本文提出一种基于启发式特征的SDC严重性指令识别方法。本文发现,指令的SDC严重性与发生数据损坏的初值、软错误传播范围以及是否能产生可检测症状有关。利用这些启发式特征,并借助决策树的思想,提出了指令SDC严重性判断方法。3.在工作1和工作2的基础上,提出一种GPGPU上的选择性指令复制方法。根据提出的SDC和严重性指令识别模型确定保护的指令集,据此在GPGPU编译过程的中间文件上部署副本指令,并设计了一致性判断模块以及软错误处理模块,从而完成了指令复制在GPGPU程序执行过程中的具体实现。本文选择了相关研究中常用的基准测试程序来评估所提出方法的性能。实验结果显示,指令复制达到了良好的SDC预测准确度,并能分别检测到程序中90.5%的SDC和86.2%的严重性SDC,带来的平均时间开销分别为1.4倍和1.29倍。更多还原

【Abstract】 General purpose Graphics Processing Units(GPGPUs)are widely used in highperformance computing centers due to their high concurrency,high throughput,and increasing programmability.The development of manufacturing process leads to shrinking chip size and higher integration,which increases the probability of soft errors caused by high-energy particle impact on GPGPU platforms.Soft errors can cause programs to generate silent data corruptions(SDCs),which are the most difficult type of errors to detect because they affect the accuracy of output,but there is no clear indication that an exception occurred during program execution.Therefore,in order to ensure the reliable execution of GPGPU programs,it is urgent to propose an effective SDC detection method.Full instruction duplication is an effective method to detect SDC,which judges whether there is a soft error by redundant execution of instructions.However,implementing full instruction duplication on GPGPU faces the following challenges.For each thread,instruction duplication increases the number of instructions it executes,thereby extending program execution time.In addition,instruction duplication will occupy additional registers to store copy data,so the register resources required by a single thread will increase,which will affect the concurrency of the thread,and thus affect the performance of the program.In this regard,the existing work proposes the idea of selective instruction copying,that is,selectively protecting the instructions in the program that are prone to SDC.Existing work usually adopts the fault injection method to analyze the soft error resilience of instructions,so as to select the instructions with higher SDC tendency for protection.However,in order to accurately judge the SDC tendency of an instruction,detailed fault injection is often required,and this process is usually very time-consuming.In addition,previous work has often provided over-protection to the program.Considering that some applications have a certain tolerance for minor SDC errors,the instruction protection conditions can be further relaxed according to the user’s precision requirements,that is,only instructions that are prone to serious SDC errors are protected.Focusing on the problems and challenges of implementing instruction replication on GPGPU,this paper proposes an energyefficient selective instruction replication method.The contributions of this work are as follows:1.A machine learning model-based SDC vulnerability instruction prediction method is proposed to efficiently identify the instructions in the program that need to be protected.We found that the attributes of the instruction itself,the function of the instruction,and the features in the error propagation process can effectively characterize the SDC proneness of the instruction.We then use a machine learning classifier to explore the relationship between these features and the SDC proneness of the instruction.By injecting faults into a small number of instructions and providing a training set for the machine learning model,all SDC vulnerable instructions in the program can be predicted with low time overhead.2.Considering that some applications have a certain tolerance for benign SDC errors,the protection of these SDCs can be relaxed,and the reliability overhead can be further reduced within a reasonable range.This paper proposes a heuristic feature-based SDC severity instruction identification method.We find that the SDC severity of an instruction is related to the initial value of corrupted data,the propagation range of soft errors,and whether it can produce detectable symptoms.Using these heuristic features and the idea of decision tree,a method for judging the severity of instruction SDC is proposed.3.Based on the above work,a selective instruction duplication mechanism on GPGPU is proposed.We determine the protected instruction set according to the proposed SDC vulnerability and severity instruction identification model,and accordingly deploy the copy instruction on the intermediate file of the GPGPU compilation process.We design the consistency judgment module and soft error processing module.Finally,we complete the instruction duplication process during the execution of the GPGPU program.This paper selects benchmarks that are commonly used in related research to evaluate the performance of the proposed method.The experimental results show that instruction duplication based on SDC vulnerability and SDC severity instructions achieves good SDC prediction accuracy,and can detect 90.5% SDC and 86.2% severity SDC in the program,respectively,resulting in time overhead of 1.4 times and 1.29 times respectively.更多还原