Title: Topics in Extreme System Design
Speaker: Zhihong Luo
Advisor: Scott Shenker, Sylvia Ratnasamy

Date: Thursday, May 8, 2025
Time: 12:00pm – 1:00pm PDT

Location: 420 Soda

Abstract: In this talk, I will present a series of systems that push the boundaries of what is considered feasible and widely accepted in the design of datacenter and cellular infrastructures. First, I explore how finer-grained software management can improve datacenter efficiency. I introduce MSH, a system that harvests memory-bound CPU stall cycles, and Fava, which enables object-level tiered memory management. Second, I demonstrate how cellular network functionality can be expanded through clean-slate architectural redesigns. I present CellBricks, a cellular architecture that dramatically lowers the barrier to entry for small-scale operators, and LOCA, a system that protects user location privacy while preserving network utility.

Title: System Abstractions for the Internet of Things
Speaker: Silvery Fu
Date: December 12, 2024
Time: 12:30 – 1:30 pm

Abstract:

Three decades ago, Mark Weiser envisioned a future where technology seamlessly integrates into our daily lives. Today, our living spaces – homes, offices, and retail locations – are being transformed by the proliferation of Internet of Things (IoT) devices. Yet, we still face challenges in fully realizing Weiser’s vision. Current smart space apps remain confined to siloed domains and narrow use cases, while big data systems lack effective methods for handling the new class of data sources presented by IoT.

In this talk, I will discuss the design of a system framework for building data-intensive applications that harness these ubiquitous data sources. First, I will introduce the digi interface, an intuitive and powerful abstraction for expressing and composing data processing logic over physical-world data. Second, I will present Jut, a mechanism that enables just-in-time data integration between digis, and Zed, a self-describing data model for efficient representation and analytics of heterogeneous digi data. Third, I will describe how these abstractions can be implemented with Knactor, a microservice-based, Kubernetes-compatible runtime, allowing digis to easily deploy and scale on cloud and edge infrastructure. I will demonstrate how the framework supports a representative set of real-world data applications with significantly less development effort and how it makes these applications easier to manage and interoperate.