ASAR: Application-Specific Approximate Recovery to Mitigate Hardware Variability

Technology scaling in microelectronics has reached limits that are resulting in increasing variation in component design and performance characteristics. Chips and systems comprising of such components are starting to exhibit a rise in process-induced failures and soft errors. Conventional design time solutions such as conservative guardbands to hide such variations are increasingly not viable for cost and performance reasons. As an alternative, researchers have sought to expose hardware fault information to the software stack and enable a programmer to use the fault information during software development. In this work, we propose the use of Software Recovery Blocks (SRB) as a programming construct that enables a programmer to provide application-specific error recovery code. Recovery comes in two modes: a) rerunning or b) discarding the erroneous computation. While rerunning comes at a performance overhead, discarding erroneous computations could result in degraded output quality, giving a user two extreme operating points on the performance-quality trade-off curve. In order to exploit intermediate performance-quality trade-off points, this work proposes approximate recovery which is particularly beneficial to approximate-computing applications. Such applications offer a natural tolerance to errors and the work introduces a SRB extension called Application-Specific Approximate Recovery (ASAR). ASAR provides 3.8%–29.9% speedup relative to rerun for six approximate-computing applications. Furthermore, the work proposes a hybrid recovery mechanism which allows a user to set desired output quality and exploit the performance-quality tradeoff curve at a finer-granularity. Hybrid recovery uses a mixture of ASAR and rerun-based recovery to demonstrate 1.5%–11.6% speedup compared to only rerun, while maintaining user-specified output quality.