CMSC 34702-1 Topics in Networks: Cloud Computing

CMSC 34702-1 Topics in Networks: Cloud Computing

CMSC 34702 Topics in Networks: Machine-Learning for Networking & Systems Junchen Jiang Fall 2019, TR 11:00-12:20, JCL 354 https://people.cs.uchicago.edu/~junchenj/34702-2019/ 1 Todays agenda Logistics (Paper discussion, Project)

Overview on ML for networking/systems 2 What will we discuss? Case studies of ML for optimizing networking/systems E.g., How ML can help design better congestion control? Architecture for data-driven approach to networking E.g., How to collect the data needed by ML?

3 But more importantly, Why ML might work for networking/systems How to do interdisciplinary research: Networks/Systems + X How systems/networking communities judge research 4

Class Format One topic per class meeting: 1-2 papers Research project: 1-2 students 50% 20% 15% 15% Class Participation

Paper Summary Paper Presentation Research Project 5 Class Format One topic per class meeting: 1-2 papers Research project: 1-2 students

50% 20% 15% 15% Class Participation Paper Summary Paper Presentation Research Project Pass/fail option (non-elective): Project is optional.

6 One topic per class meeting: 1-2 papers Before meeting We all read the papers ahead of time One classic approach paper + One ML approach paper During meeting Ask questions!

7 How to read a paper as a reader? Whats the context? Something is missing in prior work Whats the key delta? Key insights, original contributions, technical nuggets Strength & weakness Assumptions, experiment methodology, presentation, etc

Extensions Synergy with other problems 8 Tips for reading a paper Keshevs How to Read a Paper 1st Pass: Abstract, Intro & Conclusion 2nd Pass: Key technical ideas 3rd Pass: Details in depth

9 How to read a paper as a reviewer? How networking/systems program committees (PC) work How does this paper work? vs. Why should it be accepted? Crucial to have a crisp one-minute pitch! 10 Paper presentation

This is a seminar-style course Each student presents 2-3 papers in total 1-2 from the syllabus, one picked by yourself (anything related ) 11 Paper presentation Prepare a 5-min talk: Still need to read the paper carefully! Image you are championing (or arguing against) the paper Other students: PARTICIPATE

12 Research Project Group of 1-2 students Topic Applying ML to systems (defined broadly) Your own idea or talk to instructor Timeline Proposal meeting with the instructor Oct 15

Presentation Nov 26 13 Overview of ML for Networking/Systems 14 Why ML for networking/systems? A disruptive innovation in networking/systems communities Unprecedented improvement on many classic problems

Massive investments from industry Bringing expertise from both ML and systems 15 Why it might work? Netwo rked sy stems h many c

ave omplex problem s t n e m e r u

s a e m e v i s s le b

a Ma l i a v a a t a d

ML/data-driven approaches to networking/systems 16 What problems might be a good fit? Problems that are incredibly messy 1. Complex operational environments 2. Complex internal mechanisms

17 Example #1: Learned Data Structures B-Tree Why is this a good idea? Why is this not a good idea? The Case for Learned Index Structures Tim Kraska, et al. SIGMOD 2018 18

Key Innovation: Casting it as a learning problem First attempt (Single NN+Tensorflow) B-Tree outperforms NN! Why? The Case for Learned Index Structures Tim Kraska, et al. SIGMOD 2018 19 Non-trivial to make ML work in

practice Idea #1: Use TensorFlow for training, but use C++ for inference Idea #2: A hierarchy of models NN on top and linear regression at bottom Idea #3: Hybrid model Back off to B-tree for very complex structures The Case for Learned Index Structures Tim Kraska, et al. SIGMOD 2018 20

Does it work? The Case for Learned Index Structures Tim Kraska, et al. SIGMOD 2018 21 Example #2: Learning User Experience Quality of Experience (QoE) $$$

22 Developing a Predictive Model of Quality of Experience for Internet Video, Athula Balachandran, SIGCOMM13 Casting it as a ML problem Handcrafted model by experts Machine Learning Model

23 Developing a Predictive Model of Quality of Experience for Internet Video, Athula Balachandran, SIGCOMM13 QoE model can be complex Interdependencies between metrics Ideal scenario

Confounding factors Complex QoE-to-metric relationship 24 Developing a Predictive Model of Quality of Experience for Internet Video, Athula Balachandran, SIGCOMM13 A customized Decision Tree-based solution

Original formulation 25 Developing a Predictive Model of Quality of Experience for Internet Video, Athula Balachandran, SIGCOMM13 Does it work? 26 Developing a Predictive Model of Quality of Experience for Internet Video, Athula Balachandran, SIGCOMM13

Reminders Send the instructor ([email protected]) your paper presentation preference by Oct 8 Start thinking about projects 27 Questions? 28

Recently Viewed Presentations

  • Modeling the Bombardment of Saturns Rings and Fit

    Modeling the Bombardment of Saturns Rings and Fit

    Using these values, which define a power law size distribution, we calculate the . mass weighted . average particle size for both the B and C rings. Personal conversations with Jeff and Paul indicate that mass flux may be higher...
  • Seven Ways to Size Up Your Servings

    Seven Ways to Size Up Your Servings

    Measure food portions to know exactly how much you're eating. ... 6-11 servings from the Grain Group for B vitamins and fiber ... Group for vitamin C 2-3 servings from the Meat Group for iron Adapted from a National Dairy...
  • Chapter 11 Learning Objectives Autonomic Nervous System Outline

    Chapter 11 Learning Objectives Autonomic Nervous System Outline

    Define a sensation and describe the conditions necessary for a sensation to occur. List and describe the somatic sensations. Define proprioception and describe the structure of proprioceptive receptors. Describe the receptors for olfaction and the olfactory pathway to the brain.
  • Comparing the Economics of Canada, Cuba, and Brazil

    Comparing the Economics of Canada, Cuba, and Brazil

    Laws and a good court system protect people's property rights. Canada. How difficult is it to start your own business? Very easy, in a few days. Cuba. Who owns businesses and farms? ... Comparing the Economics of Canada, Cuba, and...
  • The Usability of Electronic Finding Aids during Searches

    The Usability of Electronic Finding Aids during Searches

    The Usability of Electronic Finding Aids during Searches for Known Items Society of American Archivists Annual Meeting August 22, 2003 Los Angeles, California
  • 超時空熱浴マルコフ - 京都大学

    超時空熱浴マルコフ - 京都大学

    Quantum Ratchet system Classical distribution (10 hierarchy) Wigner distribution (219 hierarchy ) Under damping Quantum Ratchet system Current across the barrier Classical result Quantum result Brownian distribution (nonOhmic) Multi-level system with the BO distribution hierarchy equations for nonOhmic noise YT...
  • Financial Goals and the Investment Process

    Financial Goals and the Investment Process

    6. Invest long-term. Avoid short-term and day trading. Its expensive and generates transactions costs and taxes. Invest wisely. There are no get-rich-quick schemes that work. Stay at least partly in the market . Taking money out of the market or...
  • The Bright and Dark Sides of Leader Traits

    The Bright and Dark Sides of Leader Traits

    The Bright and Dark Sides of Leader Traits:A review and theoretical extension of the leader trait paradigm Ronald F. Piccolo. Rollins College. Introductory slide, can be displayed before talk begins.