A Comparison of Inject-on-Read and Inject-on-Write in ISA-Level Fault Injection

Abstract

ISA-level fault injection, i.e. the injection of bit-flip faults in Instruction Set Architecture (ISA) registers and main memory words, is widely used for studying the impact of transient and intermittent hardware faults in computer systems. This paper compares two techniques for ISA-level fault injection: inject-on-read, and inject-on-write. The first technique injects bit-flips in a data-item (the content of a register or memory word) just before the data-item is read by a machine instruction, while the second one injects bit-flips in a data-item just after it has been updated by a machine instruction. In addition, the paper compares two variants of inject-on-read, one where all faults are given the same weight and one where weight factors are used to reflect the time a data-item spends in a register or memory word. The weighted injected-on-read aims to accurately model soft errors that occur when an ionizing particle perturbs a data-item while it resides in an ISA register or a memory word. This is in contrast to inject-on-write, which emulates errors that propagate into an ISA register or a memory word. Our experiments show significant differences in the results obtained with the three techniques.