Introduction to the Special Issue on Approximate Systems

Abstract

Resource efficiency is becoming an increasingly important challenge for many important applications that at the same time have nondeterministic specifications or are robust to noise in their execution. While trading correctness for efficiency has been part of computing since the early days, it has seen renewed interest in the past decade under the name Approximate Computing. A variety of techniques have been developed for applying and controlling approximations and the errors they introduce at different levels of the compute stack, from circuit to architectures and applications. However, most of these techniques have been applied in isolation at one level of the stack, making simplified assumptions about the other levels. This special issue on Approximate Systems focuses on concepts and methods for applying approximate computing principles end-toend across the compute stack. The idea for this special issue originated at a workshop on “Theory and Practice for ErrorEfficient Computing Systems” held in 2017 as well as a recent followup Dagstuhl seminar on Approximate Systems held in 2021. In response to our call for papers released in early 2021, we received 21 submissions, of which 16 were selected for an accelerated review and revision process. This special issue collects the final 7 accepted articles covering a wide range of topics at all levels of the computing stack ranging from applicationand algorithm-level approximations and adaptive application frameworks to approximation-aware hardware synthesis and custom hardware and memory system design all the way to approximations in optical interconnect. The articles presented in this special issue are aimed at providing a broad systems perspective beyond a single isolated domain to stimulate discussion and development of novel cross-layer approaches for end-to-end approximate system design. The first article, “Towards Fine-grained Online Adaptive Approximation Control for Dense SLAM on Embedded GPUs,” exploits the fact that simultaneous localization and mapping (SLAM) algorithms often have an internal probe to measure how good they are estimating the location and the map of the surroundings. This internal probe is subsequently used in a feedback loop to adaptively tune the approximation knobs and save energy without compromising the accuracy of SLAM. Next, “ParTBC: Faster Estimation of Top-k Betweenness Centrality Vertices on GPU” shows how to use controlled approximation to identify the k most important vertices in a graph faster and with small inaccuracy, leveraging both algorithm insights and executions targeting GPUs. “An Adaptive Application Framework with Customizable Quality Metrics” proposes a novel graph representation to allow users to define higher-level customized notions of quality that are used at runtime to select a configuration with maximal quality while respecting a resource budget.