Ao-Jan Su, David R. Choffnes, Fabi

1
Relative Network Positioning via CDN Redirections

• Ao-Jan Su, David R. Choffnes,Fabián E.
  Bustamante and
• Aleksandar Kuzmanovic
• Department of EECS
• Northwestern University

IEEE ICDCS 2008
2
Network Positioning

• Why do we need network positioning systems?
• Emerging large scale distributed systems can
  benefit from selecting among alternative nodes
• Example select an on-line gaming server
• All-to-all measurements are not scalable
• Current approaches
• Provide network positioning services in a
  scalable way (e.g. landmark based)
• Clear tradeoffs
• Precision vs. overhead
• Precision vs. deployment
• And others

3
Observations

• Content distribution networks (e.g., Akamai)
  improve web performance by
• Performing extensive network measurements
• Redirecting clients to their closest replica
  servers
• Publishing the results through DNS

• Can we reuse those measurements collected by CDNs
  to build a network positioning system?

4
CDNs Basics

• Web clients request redirected to close by
  server
• Client gets web sites DNS CNAME entry with
  domain name in CDN network
• Hierarchy of CDNs DNS servers direct client to
  2 nearby servers

Hierarchy of CDN DNS servers
Internet
Customer DNS servers
Multiple redirections to find nearby edge servers
Web replica servers
(3)
(4)
Client is given 2 nearby web replica servers
(fault tolerance)
(5)
(2)
Client gets CNAME entry with domain name in Akamai
(6)
Client requests translation for yahoo
LDNS
(1)
Web client
5
Our approach

• CDN-based Relative Network Positioning (CRP)
• Clients are redirected to currently closest
  replica servers in general
• CDNs redirections are primarily driven by
  network conditions (latency) Su et al. 2006
• Inferring relative network distance by
  overlapping CDN replica servers

R1
R2
A
B
C
6
Uses of CRP

• Closest node selection
• Select the closest node (shortest latency) from a
  group of candidates (e.g. select the closest
  on-line gaming server)
• Methodology
• Encode redirection frequency from a node to its
  redirected replica servers by a vector
• Compare similarity (cosine similarity) of nodes
  redirection vectors to estimate proximity

R1
Server A
Replica servers
0.8
0.2
R2
Server B
R3
Client
7
Uses of CRP (Cont.)

• Clustering
• Select a set of nodes that are close to each
  other (e.g. replicate content to a group of
  nodes)
• Methodology
• Select cluster centers
• Assign strong mapping peers to the cluster
  centers

8
Evaluation Goals

• Comparing the performance of CRPs closest node
  selection to
• Ground truth active measurements
• A state of the art network positioning system
  Meridian Wong 2005
• With respect to
• Accuracy
• Scalability
• Deployment
• Overhead

9
Experiment Setup

• Globally distributed nodes
• 1000 DNS servers as clients
• 240 Planet Lab nodes as candidate servers (on the
  same nodes as our reference system Meridian)
• Concurrent data collection
• Monitoring CDN redirections by recursive DNS
  queries for CRP
• Querying Meridian via its interface
• Measuring end-to-end latencies by pings as the
  ground truth

10
Selecting the closest node

• Clients are not close to any servers due to
• Limited Planet Lab nodes coverage (Meridian)
• Located in areas not well served by CDNs (CRP)

CRP outperforms Meridian by 25 of the nodes due
to larger deployment of CDN replica servers
CRPs accuracy is comparable to its alternative
without active measurements and dedicated
infrastructure
CRPs recommendations for 65 of nodes differ
from Meridian by lt 7ms
11
Selecting the closest node (Cont.)

• Relative Error estimated latency ground truth

CRPs is quite accurate comparing to ground
truth, with virtually no measurement overhead
80 of CRP nodes have relative error lt 50ms
12
Load on CDNs DNS System

• Rank rank 0 is the closest server

• Low probe frequency
• Smaller overhead
• Less accurate
• Miss overlapping replica servers

• 100 mins probe frequency
• Appropriate for 95 of nodes
• Much less than CDNs DNS TTL (20 secs)
• Overhead is too small to impact CDNs operations

• High probe frequency
• Can Improve accuracy
• Larger overhead

13
Load on CRP Clients

• Small history
• Sufficient for CRP
• Small overhead
• Capture network dynamics

• Large history
• More refined results
• Larger computation overhead

14
You May Be Wondering

• Will CDNs be unhappy because of CRP?
• CRP nodes behaves as regular web clients
• CRPs overhead does not impact CDNs daily
  operations
• Could be an additional service provided by CDNs
• What if CDNs change their redirection policy?
• CRPs goal aligns with CDNs
• Our approach is not restricted to a specific CDN,
  CRP can reuse results from other measurement
  infrastructures

15
Summary

• CRP discovers relative positions of end hosts
• Closest node selection
• Clustering
• Key features of CRP
• Accurate
• Light-weight
• Reuse CDNs network measurements
• Scalable
• No dedicated infrastructure is required

16
Cosine Similarity