2D: a learning based network protocol for the cloud

April 30, 2020
2:00
Zoom
Speaker: Abdullah Bin Faisal
Host: Fahad Dogar

Abstract

Popular cloud applications (e.g., Facebook) have high performance requirements (e.g., low latency). Network protocols designed to allocate bandwidth for the internet (e.g., TCP) often fall short of meeting these requirements. Consequently, there is a need to radically redesign network protocols -- based on application requirements -- that run inside the cloud.

In this talk, I will present 2D, a learning based protocol for scheduling network traffic inside clouds. 2D offers (1) tail-optimal performance across different applications while (2) maintaining good average performance. To meet its goals, 2D learns application workloads and combines insights from basic scheduling policies (processor sharing and FIFO) in a principled way.

To realize 2D for network traffic scheduling at the cloud scale, we break-up the scheduling decision into two parts: coarse time-scale decisions are offloaded to a centralized controller while critical per-flow decisions are made locally at end-hosts. Our experiments on a 20 server testbed show that, for realistic application workloads (e.g., web search, data mining), 2D achieves significant performance improvement at tail and average latency, over existing techniques.

Zoom details: Link: https://tufts.zoom.us/j/92873494819

Meeting ID: 928 7349 4819

Please see colloquia email for password.