1. DOE workshop on “COmputational Modeling of BIg NEtworks” (COMBINE) Constantine Dovrolis Georgia Tech constantine@gatech.edu

2. Outline Why this workshop? –Why modeling? –Why computational modeling? –Why “big” (as in “complex”) networks? Vertical complexity Horizontal complexity This workshop: –3 distinct parts Acknowledgments

3. Do we need modeling in the era of “big data”? My opinion: this is a silly question.. Humans understand and design complex systems based on models –It is how our brain works.. “Big data” (measurement, data mining, machine learning) are great tools to discover patterns and relations that we have not suspected to exist But “big data” cannot explain these patterns or relations

4. Why computational modeling? Traditionally, modeling has been thought of as synonymous to mathematical modeling However, useful mathematical results can be typically derived only for very simplified, “toy models” Computational modeling and simulation allow us to experiment with models of significant complexity and scale But there are many who doubt the generality and robustness of simulation results –Plus, complex simulations are hard to parameterize and validate –Instead of ignoring these doubts, we need rigorous methods to address them

5. What do you mean by “big networks”? Packet-level simulations with a large number of links and queues?? –This has been investigated in the early 2000s by several people in this room –It is only part of the story Together with scale, “big networks” refers to models that capture the complexity of modern communication networks –Vertical complexity: across the protocol stack –Horizontal complexity: along the network path

6. Vertical complexity What happens when you watch a Netflix movie online? UserMay abort, switch to other source Netflix player (HTTP)Adaptive rate selection per fragment TCPFlow and congestion control IPPacket-level effects, queueing, RED e.g., 802.11 at end-hostRate adaptation, retransmissions, etc Physical layerNoise, interference, fading, etc How many control-loops do you count in this picture? Modeling just two of them (the player and TCP) has allowed us to understand an oscillatory phenomenon that emerges as a result of the interactions between these two layers See NOSSDAV’12 paper by Akhshabi et al.

7. Horizontal complexity Typically, a network path is modeled as a sequence of links and queues. How close is this to reality? Think of a typical path between Netflix and your smart phone or laptop –Different video fragments may arrive from different CDNs –Each CDN may service that fragment request from a number of different locations and caches –The e2e transfer may cross 3-4 ASes –Each AS may deploy its own layer-2 technology and/or different QoS/DiffServ/AQM mechanisms –Private or public peering links between ASes –Broaband link: downstream vs upstream protocols –WiFi network at your home

8. Workshop’s structure Part-1: (after keynote, till 3pm today) –18 talks, 10 minutes each, 2 min for basic clarification questions –These short presentations will give us different perspectives on research challenges, state of the art, and potential breakthroughs in this area Part-2: (3:30-5:30 today) –Reflect on what was presented in Part-1, and identify specific themes, questions or research topics to discuss during Part-3. Part-3: (8:30-2:30 tomorrow) –A sequence of three discussion sessions, with each session focusing on a specific topic or question identified in Part-2. –Each discussion session will have two discussion leaders. Part-4: (2:30-3 tomorrow) –Summarize and decide next steps

9. Acknowledgments DOE and Rich Carlson Organization committee –k. Claffy, CAIDA –David Clark, MIT –Constantine Dovrolis, Georgia Tech (Chair) –John Heidemann, USC-ISI –Richard Fujimoto, Georgia Tech –Srinivisan Keshan, University of Waterloo –Don Towsley, Univ-Massachusetts, Amherst –Zhi-Li Zhang, University of Minnesota Keynote speaker –Ian Foster, ANL and Univ-Chicago