Rumba

Abstract

Approximate computing can be employed for an emerging class of applications from various domains such as multimedia, machine learning and computer vision. The approximated output of such applications, even though not 100% numerically correct, is often either useful or the difference is unnoticeable to the end user. This opens up a new design dimension to trade off application performance and energy consumption with output correctness. However, a largely unaddressed challenge is quality control: how to ensure the user experience meets a prescribed level of quality. Current approaches either do not monitor output quality or use sampling approaches to check a small subset of the output assuming that it is representative. While these approaches have been shown to produce average errors that are acceptable, they often miss large errors without any means to take corrective actions. To overcome this challenge, we propose Rumba for online detection and correction of large approximation errors in an approximate accelerator-based computing environment. Rumba employs continuous lightweight checks in the accelerator to detect large approximation errors and then fixes these errors by exact re-computation on the host processor. Rumba employs computationally inexpensive output error prediction models for efficient detection. Computing patterns amenable for approximation (e.g., map and stencil) are usually data parallel in nature and Rumba exploits this property for selective correction. Overall, Rumba is able to achieve 2.1x reduction in output error for an unchecked approximation accelerator while maintaining the accelerator performance gains at the cost of reducing the energy savings from 3.2x to 2.2x for a set of applications from different approximate computing domains.

Dynamic Multi-Resolution Data Storage

“…Since these cells do not faithfully store raw data , systems that use them can neither support exact computation nor dynamically generate data in different resolutions; such approximate-storage systems simply sacrifice flexibility [21,28,43,73]. Quality control Without revisiting storage-interface design, existing quality-control mechanisms must request full-size, raw data from the storage device [25,37,41,49,68,72,81]. Most frameworks control output quality by comparing subsets of results for exact and approximate computation, missing the opportunity to capture low-quality input that failed the requirement before computation begins.…”

“…The resolution-reduction choices Autofocus makes are usually more conservative than those of a programmer, but Autofocus can nonetheless help applications adapt to datasets. To ensure the quality of the execution result, VS may leverage existing approximate frameworks [37,41,49,68,72,81].…”

“…The Autofocus and iFilter mechanisms provide quality control and reduce the amount of programmer effort required to adjust data resolutions. Autofocus and iFilter are inspired by two previously observed phenomena: (1) The quality of the input data affects the quality of the result in approximate computing [37,41,42]. (2) A small subset of input data is representative of the rest of the input data in approximate-computing applications that tolerate inaccuracies [41] .…”

“…To reduce the overhead of applying approximate hardware or software-based approximate-computing solutions, current research projects provide support and analysis through programming language extensions and compilers [5,7,13,15,36,37,41,[70][71][72]. Since VS simply exposes its features to applications through an API and proposes extensions in the I/O protocol and firmware programs, applications can adapt VS without programming language extensions or compilers.…”

“…In spite of the clear benefits of a storage device that can effectively implement data-resolution reduction, building such a storage device is challenging, as the design must consider all of the following: Performance The computations required to adjust data resolutions in the storage device need to be efficient enough to not exceed the latency of transferring the adjusted data and should not affect normal I/O workloads. Quality Reducing data resolutions lowers the latency in data transfer but also has the potential to degrade output quality [25,37,41,42]. If the input data leads to significantly inaccurate results, the application must recompute and/or iteratively retrieve the reduced data, both of which increase end-to-end latency.…”