SAMAN: Simulation Augmented by Measurement and Analysis for Networks John Heidemann 28 September 2000 PIs: Heidemann, Deborah Estrin, Ramesh Govindan, Ashish Goel Students: Kun-chan Lan, Xuan Chen, Debojyoti Dutta USC/ISI and UCLA NMS Albuquerque PI Meeting / Sep. 2000 1 SAMAN Challenge • Network robustness is a key challenge facing the Internet: – Understanding, predicting and avoiding failures – Understanding, predicting and avoiding cascading failures – Planning failure recovery strategies • SAMAN will apply network simulation to address these problems NMS Albuquerque PI Meeting / Sep. 2000 2 “the Internet” Example Scenario 1 C1 C2 Clients Network Provider • What if the blue link becomes overloaded? – Today: discover the symptom (high loss found through manual monitoring) • SAMAN will help identify the cause: – Change in C2 traffic mix? – Interactions between C1 and C2 traffic? Need good traffic models NMS Albuquerque PI Meeting / Sep. 2000 3 “the Internet” Example Scenario 2 C1 C2 Clients Network Provider • What if the green router goes down? (DDoS?) • May produce cascading failure (blue link) • SAMAN will support prediction, understanding, and avoidance of cascading failures Need to explore correct part of large space of simulations NMS Albuquerque PI Meeting / Sep. 2000 4 Specific Failure Conditions • Fail-stop failures due to external events – accidental (backhoes) or intentional • Traffic overload – Loss rates higher than p • Good ISPs consider p>1% serious – Loss rates map non-linearly into performance degradation and load – Benign (simple overload), unexpected (traffic shift), or malicious (DDoS) • Current challenge: failure propagation (cascades, delayed convergence, etc.) [Shaikh00a,Labovitz00a] NMS Albuquerque PI Meeting / Sep. 2000 5 goodput Why Simulation? load Answer “what if?” For protocols, scales, scenarios outside experimentation. (But depends on good models in interesting part of space.) NMS Albuquerque PI Meeting / Sep. 2000 6 Agenda • Challenges • SAMAN in NMS – Applications – Technologies • Early results • Potential collaborations NMS Albuquerque PI Meeting / Sep. 2000 7 SAMAN Applications • Failure prediction: – Understanding and reproducing protocol behavior under extreme conditions – Network early warning system – Tools to automatically generate models NMS Albuquerque PI Meeting / Sep. 2000 8 Protocol Robustness • Reliable networks demand reliable protocols • How do individual protocols behave near the edge of their operating limits: – What conditions are important to study? – Are simple protocol improvements possible? • How do protocols interact in extreme conditions: – How do individual and aggregate behavior relate? – When does individual failure trigger cascading failure? NMS Albuquerque PI Meeting / Sep. 2000 9 Network Early-Warning Systems • Tools to predict imminent network failures – Trigger preventive or corrective actions • Clear mappings from tools to specific failures – Many current tools do local measurements – Are measurements topologically or temporally related? • Minimize control loop – Performance, understandability, deployability… NMS Albuquerque PI Meeting / Sep. 2000 10 Model Generation Tools • Tools to automatically configure simulation models from network measurements – Integrate data from multiple network points – Serve as input to other portions of work – Validated across multiple time-scales • Build on library of validated simulation models NMS Albuquerque PI Meeting / Sep. 2000 11 Agenda • Challenges • SAMAN in NMS – Applications – Technologies • Early results • Potential collaborations NMS Albuquerque PI Meeting / Sep. 2000 12 SAMAN Technologies • Just-in-time model generation – Accurate traffic models • Analysis-informed simulation – Constrain parameter search space • In a robust simulation environment – Build on widely-used ns platform NMS Albuquerque PI Meeting / Sep. 2000 13 Model Generation • Application-driven (structural) models – Capture application-level dynamics (feedback, user behavior) – Validated, applicable across range of time-scales • Network measurements to parameterize models – Integrate data from multiple measurement points • Resulting in just-in-time models – Network admins can measure and parameterize models NMS Albuquerque PI Meeting / Sep. 2000 14 Analysis-Informed Simulation • Failure analysis spans huge parameter space – Most of space is uninteresting • Analysis-informed simulation – Rapid analytic pre-simulation pass categorizes scenario as uninteresting (clearly out of scope) or interesting – Focus detailed simulation on interesting scenarios NMS Albuquerque PI Meeting / Sep. 2000 15 Ns Simulation Environment • Builds on rich ns simulation environment – Wired and wireless (radio and satellite) – Robust protocol library: many TCP variants, multicast, … – Validation experience and test suite • 648 scenarios in 58 categories – Multiple levels of abstraction • packet-level and abstractions eliminating per-hop routing, multicast tree formation, mixed abstract/detailed sims, etc. – Emulation: mix real-world and virtual nodes – Broad community support and use • ns-users mailing list: >1000 hosts (~institutions), >8000 e-mail addresses (~users) NMS Albuquerque PI Meeting / Sep. 2000 16 Large Simulations • Evaluating scalability in single dimension very risky – many dimensions: nodes, users, multicast senders vs. recievers, protocol agents, traffic volume – understanding is often bottleneck • Parallelism – sometimes key…if one simulation has the answer – don’t ignore free parallelism if multiple simulations needed (ex. vary parameters, replicate results) • Abstraction is critical to large and fast network sim: – ns went from 100s to 1000s by tuning [on desktop hardware], but 1000s to 10000s with abstractions – many abstractions: • centralizing computations (unicast and multicast routing, etc.) • packet delivery abstraction (trains, end2end delivery, fluid flow) • protocols abstractions (FSA TCP, etc.) • mixed abstract/detailed sims NMS Albuquerque PI Meeting / Sep. 2000 17 Agenda • Challenges • SAMAN in NMS – Applications – Technologies • Early results • Potential collaborations NMS Albuquerque PI Meeting / Sep. 2000 18 Early Results • Current focuses: – Reproducing failure scenarios in simulation – Multi-scale, application-driven traffic models – Pre-simulation scenario filtering NMS Albuquerque PI Meeting / Sep. 2000 19 Early Results • Current focuses: – Reproducing failure scenarios in simulation – Multi-scale, application-driven traffic models – Pre-simulation scenario filtering NMS Albuquerque PI Meeting / Sep. 2000 20 Modeling Real-Audio Traffic • Why real audio? – Example of a streaming media protocol – Very different from TCP – Possibly representative of future streaming media (certainly more representative than TCP) • Why now? – Help develop tools for multi-scale models – Modeling protocol effects without source code NMS Albuquerque PI Meeting / Sep. 2000 21 Basic R-Audio Behavior • Constant bit-rate time-sequence plot of single flow • … or not? mean and quartiles of 1200 flows (mean is smooth, quartiles at multiples of 1.8s) NMS Albuquerque PI Meeting / Sep. 2000 22 R-Audio Under the Microscope • More complex internal structure • Demonstrates importance of studying protocols at multiple time-scales • Able to capture internal structure after iteration bursts NMS Albuquerque PI Meeting / Sep. 2000 1.8s inter-burst interval 23 R-Audio: Time-Variance Plot trace model noticeably less variance at key scales (1.8, 3.6, etc.) NMS Albuquerque PI Meeting / Sep. 2000 24 R-Audio: Scaling Plot trace NMS Albuquerque PI Meeting / Sep. 2000 model 25 R-Audio Experiences and Plans • Currently validating model – stats seem promising – validation against additional traces in progress • Next steps: – Rapid model parameterization – Apply tools to complex models (mixed traffic) – Apply models to NMS challenge problem NMS Albuquerque PI Meeting / Sep. 2000 26 Early Results • Current focuses: – Reproducing failure scenarios in simulation – Multi-scale, application-driven traffic models – Pre-simulation scenario filtering NMS Albuquerque PI Meeting / Sep. 2000 27 Agenda • Challenges • SAMAN in NMS – Applications – Technologies • Early results • Potential collaborations NMS Albuquerque PI Meeting / Sep. 2000 28 Potential Collaborations • NMS can use models (ex. real audio) – In public ns releases now – Could be ported to other simulators • Model parameterization could use NMS measurement tools • Collaborative addition of NMS work into ns – Traffic, topology models – Simulation optimizations and abstractions • Non-NMS projects (STRESS, etc.) • Other opportunities? NMS Albuquerque PI Meeting / Sep. 2000 29 More information • http://www.isi.edu/saman/ NMS Albuquerque PI Meeting / Sep. 2000 30