Unleashing the Potential of Approximate Computing Systems

Today, we are seeing an explosion in emerging compute/data-intensive applications, such as AI and data analysis services, the rollout of next-generation networks, and the growth of smart edge devices. However, the transition to a post-Moore era raises significant concerns ranging from compute and storage efficiency to carbon footprint/energy consumption. An underexplored but promising opportunity to improve the cost-performance-sustainability tradeoffs of existing computing systems is the use of approximation. Data systems may not need to calculate results with 100% precision to maintain operational reliability. In this talk, I will present my research on scaling data systems with approximation techniques for various analytical tasks across the computing stack, such as dynamic connected data processing and network traffic analysis. First, I will discuss efficient algorithms and system optimizations that enable mining complex structures in large-scale graph data. Second, I will describe how bridging theory and practice with sketching and sampling techniques can significantly speed up network analytics under tight resource budgets. The systems I have built are backed with rigorous theoretical guarantees and achieve several orders of magnitude improvements with small accuracy losses. The developed approximate graph systems are under evaluation in the industry (e.g., a fintech company and a cloud provider), and the network analytics solutions are deployed in popular open-source products (e.g., Data Plane Development Kit). Finally, I will chart paths to designing future approximate computing systems with heterogeneous hardware that balance performance, reliability, and sustainability.