Improving transient memory fault resilience of an H.264 decoder

Traditionally, fault-tolerance has been the domain of expensive, hard real-time critical systems. However, the rates of transient faults occurring in semiconductor devices will increase significantly due to shrinking structure sizes and reduced operating voltages. Thus, even consumer-grade embedded applications with soft real-time requirements, like audio and video players, will require error detection and correction methods to ensure reliable everyday operation. Cost, timing and energy considerations, however, prevent the embedded system developer from correcting every single error. In many situations, however, it will not be required to create a totally error-free system. In such a system, only perceptible errors will have to be corrected. To distinguish between perceptible and non-perceptible errors, a classification of errors according to their relevance to the application is required. When real-time conditions have to be observed, the current timing properties of the system will provide additional contextual information. In this paper, we present a structure for an error-correcting embedded system based on a real-time aware classification. Using a cross-layer approach utilizing application annotations of error classifications as well as information available inside the operating system, the error correction overhead can be significantly reduced. This is shown in a first evaluation by analyzing the achievable improvements in an H.264 video decoder under error injection and simulated error correction.