Instruction-Level Fault Tolerance Configurability

Fault tolerance (FT) is becoming increasingly important in computing systems. FT features are based on some form of redundancy, which adds a significant cost to a system, either increasing the required amount of hardware resources or degrading performance. To enable a user to choose between stronger FT or performance, some schemes have been proposed, which can be configured for each application to use the available redundancy to increase either reliability or performance. We propose to have an instruction-level, rather than application-level, configurability of this kind, since some applications (for example, multimedia) can have different reliability requirements for their different parts. We propose to apply weaker (or no) FT techniques to the less critical parts. This yields a certain time or resource gain, which can be used to apply stronger FT techniques to the more critical parts, thereby, increasing the overall FT. We show how some existing FT techniques can be adapted to support instruction-level FT configurability, and how a programmer can specify the desired FT of particular instructions or blocks of instructions in assembly or in a high-level programming language. In some cases compiler can assign the FT level to instructions automatically. Experimental results demonstrate that reducing the FT of non-critical instructions can lead to significant performance gains compared to a redundant execution of all the instructions. The fault coverage of this scheme is also evaluated, demonstrating that it is very application-specific. For some applications the fault coverage is very admissible, but unacceptable for others.