Rethinking Distributed State Management in Networks

Опубликовано 31 июля 2018, 16:40

A key challenge in scaling network control and data planes is to maintain access to shared state without comprimising performance. In this talk, I present Tasvir, a versioned distributed shared memory system, that achieves this goal by carrying out periodic bulk data-oriented synchornization and decoupling heavyweight synchronization from the fastpath. I describe Tasvir's design, performance benefits, and make a case for its general applicability to networks through a set of sample applications we built. I conclude by discussing our ongoing work on using Tasvir to handle distributed state management for key-value stores and distributed machine learning.

See more at microsoft.com/en-us/research/v...