CS 294-4: Network Support for Reliable Adaptive Distributed Systems (NetRADS)

MW 1-2:30, Room 310 Soda

3 Units

Prof. Randy H. Katz, Instructor


Course Overview

Distributed systems are known to be fragile, easily broken, and exhibit poor reliability. From a network perspective, reliability is affected by two key factors. In the wide-area, routes occasionally fail and the network can take too long to find an alternative, thus partitioning distributed components. In the local-area and in enterprise networks, unplanned for traffic surges (e.g., worm outbreaks, p2p file sharing, flash crowds) can starve applications for network bandwidth while simultaneously rendering the network difficult to manage in order to respond to the surge.


In this seminar, we will examine new network-layer approaches for improving the dependability of distributed systems. A key emergent technology of interest is programmable network elements (PNEs), providing protocol-aware packet and stream classification at line speeds, invocation of specific packet and flow-level actions, and policy-based forwarding. Packet contents can be modified, delayed, filtered or redirected, based on preferences, context, or policies specified by network administrators. For example, a strategically place PNE could detect traffic surges and shape traffic to reserve bandwidth for control messages when the network is under stress.


In addition, PNEs provide pervasive monitoring at the network layer, as a possible foundation for a network "flight data recorder." This could enable analysis and replay of network failures after the fact. Given the ability to observe and act, PNEs can coordinate with algorithms at a higher level to drive predictions of emerging attacks or impeding failures. Successful coordination of observe-analyze-act across the network could dramatically enhance the reliability and dependability of distributed systems.


The course will be organized around reading and discussing key papers in the field, drawn from the literature of network virtualization, routing and health monitoring, and active networks and middlebox functionality. These readings will be augmented with invited presentations by leading experts in the field. Students will propose a significant research project and see it through to completion and presentation by the end of the semester. PNEs are available for student evaluation and prototyping in support of their projects.

Expected Background

Graduate level courses in networking (e.g., CS 268) or operating systems (e.g., CS 262) or familiarity with recent research literature in these areas

Evolving Seminar Agenda

See link.

Grading

50% -- Seminar Participation including leading paper discussions and preparing tutorial presentations;
50% -- Course Research Project, including proposal and project presentations, and end of semester poster session;
No Exams!

Last updated: 30 December 2004, randy@cs.Berkeley.edu