Sources of Variations in Error Sensitivity of Computer Systems

Licentiate thesis, 2014

Technology scaling is reducing the reliability of integrated circuits. This makes it important to provide computers with mechanisms that can detect and correct hardware errors. This thesis deals with the problem of assessing the hardware error sensitivity of computer systems. Error sensitivity, which is the likelihood that a hardware error will escape detection and produce an erroneous output, measures a system’s inability to detect hardware errors. This thesis present the results of a series of fault injection experiments that investigated how er- ror sensitivity varies for different system characteristics, including (i) the inputs processed by a program, (ii) a program’s source code implementation, and (iii) the use of compiler optimizations. The study focused on the impact of tran- sient hardware faults that result in bit errors in CPU registers and main memory locations. We investigated how the error sensitivity varies for single-bit errors vs. double-bit errors, and how error sensitivity varies with respect to machine instructions that were targeted for fault injection. The results show that the in- put profile and source code implementation of the investigated programs had a major impact on error sensitivity, while using different compiler optimizations caused only minor variations. There was no significant difference in error sen- sitivity between single-bit and double-bit errors. Finally, the error sensitivity seems to depend more on the type of data processed by an instruction than on the instruction type.