NetProfiler: Profiling WideArea Networks Using Peer Cooperation

1
NetProfiler Profiling Wide-Area Networks Using
Peer Cooperation

• Venkata N. Padmanabhan MSR
• Sriram Ramabhadran UCSD
• Jitendra Padhye MSR

2
Motivation

• Operators have little direct visibility into the
  end users network experience.
• Although end users have direct visibility into
  their own network performance, they have little
  other information or recourse when they encounter
  problems.

3
NetProfiler

• End hosts monitor the network performance
• Share the information with other end hosts over a
  peer-to-peer network
• NetProfiler can detect network anomalies and
  localize their likely cause.
• This system allows users (and also ISPs) to learn
  about the network performance experienced by
  other hosts.

4
NetProfilers Characters

• An P2P application that leverages peers for
  network monitoring and diagnosis.
• Peer participation is critical in NetProfiler.
• NetProfiler considers the entire end-to-end
  Transaction.
• NetProfiler monitors, aggregates, and summarizes
  network performance data on a continuous basis.
• NetProfiler relies on passive observation of
  existing traffic.

5
NETPROFILER ARCHITECTURE AND ALGORITHMS

• Data Acquisition
• Data acquisition is performed by sensors.
• One sensor for each protocol or application
• Normalization
• Recording end host related information to avoid
  being misunderstood.
• Data Aggregation
• Based on a set of client attributes.
• Using SDIMS to aggregate data.

6
Continued

• Analysis and Diagnosis
• Distributed Blame Allocation
• Comparative Analysis
• Network Engineering Analysis
• Network Health Reporting

7
Experimental Environment

• 147 PlanetLab nodes, dialup hosts connected to 26
  PoPs on the MSN network, and 5 hosts on
  Microsofts worldwide corporate network.
• Repeatedly download content from a diverse set of
  70 web sites during a 4-week period (Oct 1-29,
  2004).

8
Experimental Results

• www.technion.ac.il www.hku.hk failured
• There are significant differences in the failure
  rate observed by clients that are seemingly
  equivalent.
• Sometimes a group of clients shares a certain
  network problem that is not affecting other
  clients.
• In other instances, the problem is unique to a
  specific client-server pair.

9
Discussion

• Deployment Models
• Coordinated
• Organic
• Bootstrapping
• Chicken-and-egg problem
• Allowing a limited amount of active probing
• Security
• Privacy
• Data integrity

10
Related Work

• Vs. ping, traceroute, tulip etc
• Vs. network tomography
• Vs. PlanetSeer
• Vs. NetFlow, Route Explorer
• Vs. SPAND
• Vs. Ganglia, Slicestat, IrisNet, PIER, Sophia,
  SDIMS, Astrolabe
• Vs. Knowledge Plane
• Vs. STRIDER, PeerPressure