Research On Pin-based Fault Injection Technique

In recent years,as computer systems have grown in size and component density,transient faults caused by cosmic rays and some high-energy particles have begun to be found in large numbers in terrestrial computer systems.These transient faults,known as Soft Errors,occur primarily in registers and memory,and pose a significant threat to the reliability of current computer systems.Due to the lack of regularity and the short duration of soft errors,it becomes very difficult to record their occurrence and behavior.Therefore,it is very important to study the fault injection method for soft errors to study the impact of soft errors on hardware and software systems.In order to assess the system's response to a fault,errors are usually injected into the target program to observe the behavior of the program under error.A common method of fault injection is to modify the hardware or inject hardware faults th rough software emulation hardware.Hardware modifications are often costly and uncertain,and software simulation methods often fail to simulate the behavior of hardware failures accurately.To this end,this paper proposes a binary-level dynamic fault injection method based on Intel's dynamic instrumentation tool Pin.This method can perform fault injection operations on binary target programs at the machine code level which can simulate the fault behavior of soft errors more accurately.In this paper,we analyze the instruction-level instrumentation framework of Pin in detail,and design the fault model and injection method on machine code level for soft error.In order to evaluate the impact of soft errors on the program,a Pin-based binary-level fault injection tool B-SEFI is designed and implemented on this basis to simulate soft errors in registers and memory units.B-SEFI can customize the location and the number of fault injections and analyze the fault injection results statistically.Finally,this paper selects five classic machine learning algorithms as the target program.By performing soft error injection on specific registers and memory units during the execution of these programs,the validity of the B-SEFI tool is verified.The effects of soft errors to classical machine learning algorithms on performance and accuracy are analyzed in this paper.