BROAD: bold and reliable online approximate computing framework for diverse applications

Approximate computing is an emerging computing paradigm that leverages the inherent resilience of applications while designing energy-efficient computing systems. Approximate computing systems must satisfy user-provided requirements for quality of service (QoS), a quantitative criterion imposed on the output of an application such that the output is qualitatively useful. Previous software frameworks for approximate computing rely on the assumptions that approximation errors do not propagate through applications and that occasional QoS violations are acceptable. In this thesis, we explore the application of software approximations to applications for which these assumptions do not hold. We also observe that to avoid unacceptable QoS degradation (i.e., degradation beyond the QoS requirement), previous frameworks had to include a static approximation level guardband, which reduces the benefits to energy-efficiency. We propose BROAD, a Bold and Reliable Online Approximate Computing Framework for Diverse Applications. BROAD explicitly provides a checkpoint/rollback mechanism to allow applications to recover from QoS violations and error accumulation. The checkpoint/rollback mechanism further obviates BROAD from having a static approximation level guardband by allowing BROAD to operate near the QoS requirement without concern for permanent QoS degradation