The agenda for the seminar this week is listed below:
Title: Tasvir: Versioned Shared Memory for Distributed State Management
Abstract: Despite providing a convenient communication abstraction, distributed shared memory systems (DSMs) are rarely used due to their poor performance at scale. I will present Tasvir, our work that seeks to elevate the application-network interface without impeding performance. Unlike canonical DSMs, Tasvir’s synchronization is passive, bulk, and explicitly-controlled, and its shared memory regions are user-defined, single-writer, and hierarchically organized. This lets applications retain memory-local performance and gracefully trade user-configurable write visibility delay with scale. It also enables a high degree of efficiency for snapshotting and replication that is otherwise hard to achieve. I will talk about Tasvir’s performance and our experience in using it to simplify state management for network services, key-value stores, and distributed optimization.
Bio: Amin is an independent researcher in London broadly interested in systems and networking. His current work focuses on high-performance distributed state management. Previously, he was a researcher at Intel Labs and a researcher-in-residence at UC Berkeley where he primarily worked on performance isolation for consolidated software packet processors. During his PhD, he split time between University of Toronto and UC Berkeley where he was advised by Yashar Ganjali and Scott Shenker and worked on the performance and scalability of early SDN systems.
Venue: Bancroft 1.08 (Bancroft Building)
Time and Date: 2 to 3 PM, Wednesday, 29th May 2019.
Please keep the dates on your calendar.
Any comments, advice would be appreciated.