Analyzing the effects of compiler optimizations on application reliability

Abstract

As transistor sizes decrease, transient faults are becoming a significant concern for processor designers. A rich body of research has focused on ways to estimate the vulnerability of systems to transient errors and on techniques to reduce their sensitivity to soft errors. In this research, we analyze how compiler optimizations impact the expected number of failures during the execution of an application. Typically, optimizations have two effects. First, they increase structures occupancies by allowing more instructions in flight, which in turn increases their susceptibility to soft errors. Additionally, they decrease execution time, decreasing the time during which the application is exposed to transient errors. In particular, we focus on how optimizations impact occupancies in three processor structures, namely the Reorder Buffer, the Instruction Fetch Queue and the Load Store Queue. We explain the interplay between compiler and reliability by studying the changes in the code made by the compiler and the resulting responses at the microarchitectural level. Results from this research allow us to make decisions to keep an application within its performance goals and its vulnerability during its runtime within a well defined FIT target.