Creating Data Repositories..

1
Creating Data Repositories..

• Sanjay Rao
• ECE Dept, Purdue University

2
Group Members

• Dave Maltz
• Rebecca Issacs
• Ratul Mahajan
• Yin Zhang
• Aditya Akella
• David Kotz
• Charles DiFatta
• ..

3
Motivation

• Network Management Research
• Barrier to entry is high
• Data/insights from operators/industry critical
• Examples
• Failure characterization of enterprise network
• VLAN characterization and use
• Configuration Management

4
What happens today..?

• End-user centric measurement studies
• Network black-box no operator involvement
• Real need white-box
• Campus Networks
• Difficulties in bootstrapping relationships with
  operators
• Enterprise/Operator Network
• Sprint or ATT (Microsoft with end-user)
• Limited pool of researchers
• Data across multiple enterprises??
• Trends over many years ??

5
Bottomline

• Need a data repository
• Contributors from operators, researchers,
  industry
• Accessible to all researchers
• Facilitate research much like Planetlab
• Vital to have critical mass of researchers on
  Network Management
• Research along high-impact real problems

6
Data Sharing what inhibits it?

• Sensitivity of data
• Security Issues (firewall policies, network
  structure)
• Privacy Issues (records of individual activity)
• Proprietary nature of data
• E.g. how many calls got, mobility models
• Possible to have others use it?
• Secret weapon for research
• Competition Vs. collaboration
• Inertia/ too much effort

7
Solutions

• Carrots/sticks to promote data sharing
• Must release data to publish
• IMC best paper award only to work releasing
  data.
• Technical ways to addressing concerns with
  sharing

8
Positive Example
Example HSARPA PREDICT make research on
network security possible. Firewalls and IDS
network security data
9
Research Anonymization

• Hiding provider, hiding individual information
• Need framework to reason about it
• What trade-offs do you make?
• What risks are posed?
• How to expose trade-offs in a way we can
  appreciate?
• Anonymization very domain specific
• E.g. configuration file Vs. packet trace
• Are there common themes?
• Other Models
• NDA-based
• Give me a question -gt return answer
• Exploratory nature of research

10
Community effort Cooperate on IRB

• Social Sciences
• Lots of experience with IRB
• Networking
• Lack of clear guidelines on IRB process
• Admins feel happier if IRB can sanction things
• As community
• Must appreciate need/process for IRB
• Develop guidelines for IRB process
• Share IRB documents

11
Creating shareable data

• 75 of time spent figuring how to use data
• Researcher needs vary
• Different forms of datum
• Historical Vs. Streaming
• Dated? Trending?
• Assumptions made/gaps in data
• timing info crucial at sub-RTT level?
• Sharing hard, many idiosyncrasies
• Data collection infrastructure, annotate

12
User Diagnostics

• One-on-one exact data provided
• Create shared repository(ies)
• What data do most users want?
• Is that 20 of stuff most critical to provide?
• Data Collection Tools
• Meta-data part of problem
• Create data in standard formats
• Observatory
• How to discover, describe, explain data
• Access policy, use policy

13
Other

• Streaming Data Online Vs Offline
• Scalable collection
• What to collect? Over how long?
• Compression techniques
• Fine-grained overhead, coarse-grained
  information loss
• What does it take to build this infrastructure?
• Get all types of data as painlessly as possible
• Massage, orchestrate data to fit researcher needs
• Simple APIs to get data out fast analysis tools
• Federated Access
• DataManagement - Lifecycle of data

14
Action Items

• Community-Wide Efforts
• Initiate efforts to create data repository
• How to manage? Who contributes? Who arbitrates
• How much storage? Lifecycle - How long to store
  data?
• Create IRB guidelines for networking data
• Research
• Anonymization
• Usage diagnostics -gt what to collect,release
  widely applicable
• Data Collection Tools, metadata information
• Industry,operators must be as actively involved
  as possible