# CS 176C: Self-Driving Networks [Slides](https://hackmd.io/@arpitgupta/HkLlI2L28) --- ## Learning Objectives - Self-Driving Networks - What's the history, what's new, and what are the key challenges? - How can we realistically realize the goal of developing self-driving networks? - Course Summary - What did we learn this quarter? - Logistics and next steps --- # Self-Driving Networks --- ## Long Road to Self-Driving Networks - Definition: - Networks that run with minimal to no human intervention—able to configure, monitor, and maintain themselves independently - Past efforts: - Active networking (late 90s) - Self-* networking (early 2000s) - Autonomous networking (~2010) --- ## What's Different Today? (1) ### Market Pull - Zero-touch networks (Google, ~2016) - Self-driving networks (Juniper, ~2017) - Adaptive networks (Cienna, ~2018) --- ## What's Different Today? (2) ### Technology Push - Fully programmable, protocol-indepdentent data planes - with language to program (e.g., P4) - what methods for formal verification - AI and ML techniques - AI/ML for networking - Networking for AI/ML --- ## AI/ML and Self-Driving Networks - AI/ML techniques have been potent for areas where - (labeled) data is abundant - trust/privacy are not key requirements - large-scale testing in productions settings was feasible Is Networking such an area? --- ## (Labelled) Data Problem - What is the right type of data (scale)? - How to obtain the data (privacy)? - How to address data quality issues (scale)? --- ## Self-driving Car Analogy - DARPA Grand Challenge - 2004: **None** of the finalists finishes the 150-mile route - 2005: **5 of the 23** finalists completed the 132-mile course - 2007: **6 of the 11** finalists completed the 55-mile urban course - What changed after 2004? - [LIDAR](https://en.wikipedia.org/wiki/Lidar) --- ## LIDAR  LIDAR measurements have been viewed as *game changer* for self-driving cars --- ## Why LIDARs made a difference? - Features - Millions of beams of laser light per second - 3600 visibility, for distances up to ~200m - +/- 2cm depth perception accuracy --- ## Why LIDARs made a difference? - Features - Millions of beams of laser light per second - 3600 visibility, for distances up to ~200m - +/- 2cm depth perception accuracy - Criticism - Cost (was $80K in 2010, now ~$8K) - Overkill (too much data) --- ## Today's Self-Driving Cars  LIDAR contribute significantly to total data usage --- ## LIDAR for Self-Driving Networks? (1) - Network data monitoring/collection today … - Piecemeal, problem-specific, ad-hoc, an after-thought, ... - The curse of complexity, scale and dimensionality --- ## LIDAR for Self-Driving Networks? (2) - Network data monitoring/collection tomorrow … - Holistic, network-wide, all the time, no sampling, - Augmented with auxiliary data (e.g., config files, routing info) - Early criticism: - Too expensive! - Overkill! --- ## AI/ML Trust Problem - A clash of two cultures - **Black-box** nature of conventional AI/ML - Opaque, unintuitive, unintelligible - Network operators desire for **white-box** solutions - Want intuition - Look for understanding - Seek to gain insights Such clash impedes deployment of self-driving networks in conventional settings --- ## Summary of Challenges - Data problem - Collect network data at **scale** in a **privacy-preserving** problem - Trust problem - Make inferences and actions more **explainable/interpretable/verifiable** - Ownership - Democratize ownership of data and learning algorithms/models --- ## Approach  Scalable and privacy-preserving data collection/analysis pipeline --- ## Privacy-preserving Model/Data Sharing  Use two non-colluding cloud providers, homomorphic encryption, and Yao's garbled circuits to train learning models in a privacy-preserving manner --- # Course Summary --- ## Last-Mile Networks - WLAN Networks - Cable Networks - Cellular Networks --- ## WLAN Networks - Basics - key components, different standards - Wireless medium access - hidden and exposed terminal problems - Mobility - client's association/disassociation - 802.11 n - what enables high data rates --- ## Cable Networks - Basics: - key components (HFC networks) - DOCSIS: - how it evolved over time? - DOCSIS 3.1 deep dive --- ## Cellular Networks - Basics: - Key components, different generations - Mobility: - How cellular networks handle mobility? - 5G Networks: - key components, architecture, research directions --- ## Multimedia Networking - Video streaming - requirements, client-side buffering, transport protocols - VoIP - requirements, playout delays, operational challenges - Real-time protocols - RTP headers, RTCP protocol --- ## Network Support - Scheduling/Policing Mechanisms - packet marking, policing, traffic shaping - Bufferbloats - Active Queue Management --- ## Network Measurements Systems - Last-mile Networks - BISMark - CableMon - CDN Measurement Systems - Odin --- ## Networking Streaming Telemetry - System: - Sonata - Enabler: - RMT Architecture --- ## Future directions - Digital divide - Self-driving networks