Exploiting soft computing for increased fault tolerance

Traditionally, fault tolerance researchers have made very strict assumptions about program correctness. Such strict notions of correctness are appropriate for workloads that are numerically oriented. However, a growing number of important workloads produce results that have a higher (often qualitative) user-level interpretation. We call these soft computations. Examples of soft computations include multimedia processing, as well as cognitive information processing [1]. While data corruptions can change the numerical result of soft computations, they often do not change the user’s interpretation of those numerical results. Hence, faults that would otherwise be deemed unacceptable from a strict numerical standpoint may in fact be tolerable (or even imperceptible) from the user’s standpoint. Systems that can identify and exploit such error resiliency at the user level offer new opportunities for fault tolerance optimization. This paper makes several contributions in the context of fault tolerance for soft computations. First, we identify three characteristics of soft computations that make them resilient to error: redundancy, adaptivity, and reduced precision. Second, we quantify the extent to which soft computations are fault tolerant by conducting fault-injection experiments. Our results show that using relaxed notions of program correctness, 82 % of injected faults are tolerable. Furthermore, when employing a lightweight recovery scheme enabled by soft computations, we find 96 % of all injected faults are acceptably tolerated in our soft computing benchmarks. Finally, we present a method for identifying soft computations at the instruction level using dynamic slicing analysis. We find that in the benchmarks we study, 62 % of all dynamic instructions participate in soft computations, and can tolerate single-bit errors up to a fault rate of 8 × 10 −6