Towards Accurate Estimation of Error Sensitivity in Computer Systems

Abstract: Fault injection is an increasingly important method for assessing, measuringand observing the system-level impact of hardware and software faults in computer systems. This thesis presents the results of a series of experimental studies in which fault injection was used to investigate the impact of bit-flip errors on program execution. The studies were motivated by the fact that transient hardware faults in microprocessors can cause bit-flip errors that can propagate to the microprocessors instruction set architecture registers and main memory. As the rate of such hardware faults is expected to increase with technology scaling, there is a need to better understand how these errors (known as ‘soft errors’) influence program execution, especially in safety-critical systems. Using ISA-level fault injection, we investigate how five aspects, or factors, influence the error sensitivity of a program. We define error sensitivity as the conditional probability that a bit-flip error in live data in an ISA-register or main-memory word will cause a program to produce silent data corruption (SDC; i.e., an erroneous result). We also consider the estimation of a measure called SDC count , which represents the number of ISA-level bit flips that cause an SDC. The five factors addressed are (a) the inputs processed by a program, (b) the level of compiler optimization, (c) the implementation of the program in the source code, (d) the fault model (single bit flips vs double bit flips) and (e)the fault-injection technique (inject-on-write vs inject-on-read). Our results show that these factors affect the error sensitivity in many ways; some factors strongly impact the error sensitivity or SDC count whereas others show a weaker impact. For example, our experiments show that single bit flips tend to cause SDCs more than double bit flips; compiler optimization positively impacts the SDC count but not necessarily the error sensitivity; the error sensitivity varies between 20% and 50% among the programs we tested; and variations in input affect the error sensitivity significantly for most of the tested programs.