David Choffnes, CDN-based Detouring, INFOCOM 2009

On the Effectiveness of Measurement Reuse
for Performance-Based Detouring
David Choffnes
Fabian Bustamante
Northwestern University
INFOCOM 2009
Detouring
FTP transfer between UCB and Dartmouth
UC Berkeley
Intel Research
300 KB/s
Dartmouth
400 KB/s
David Choffnes, CDN-based Detouring, INFOCOM 2009
2
Detouring
UC Berkeley
286 ms
Dartmouth
0.8 ms
Intel Research
78 ms
David Choffnes, CDN-based Detouring, INFOCOM 2009
3
Choosing Alternate Paths
David Choffnes, CDN-based Detouring, INFOCOM 2009
4
Locating Detour Paths
Guess randomly
– SOSR [Gummadi et al, NSDI 2005]
– Improves reliability, but not necessarily performance
Brute force
– RON [Andersen et al, SOSP 2001]
– All-to-all measurements (O(n2))
• 100 nodes: ~100 Kbps at each node
• 100,000 nodes: ~100 Mbps at each node!
– Fundamentally architecturally unsound
– Wasteful for every system to try to measure
Can we reuse someone else’s measurements?
David Choffnes, CDN-based Detouring, INFOCOM 2009
5
Sustainable Scalability
Reduce, Reuse, Recycle (3R)
– Long-running services already measuring the
network
– Sustainable yield by reducing wasted Internet
resources from distributed systems through strategic
reuse and recycling of network information
Long-running services as oracles
– Low-cost, open sources of information about the
network
– Examples
• Content distribution networks
• P2P file sharing
David Choffnes, CDN-based Detouring, INFOCOM 2009
6
CDNs as Potential Oracles
Content distribution networks (CDNs)
– Akamai, Limelight, Mirror Image
– Serve popular sites like Yahoo, ABC.com, MySpace
– Move load from content providers (e.g., Yahoo) to
edge of network (CDN servers)
– Provide interactive downloads for users by finding
replicas along good network paths to users
– Use extensive network measurements to inform
replica server selection
– Over 10,000 servers worldwide
David Choffnes, CDN-based Detouring, INFOCOM 2009
7
Road Map
Introduction
3R: Reduce, Reuse, Recycle
CDN-based detouring
– Using CDNs as oracles
– Locating high-performance paths
– Building a system to exploit them
Conclusion
David Choffnes, CDN-based Detouring, INFOCOM 2009
8
How CDNs Work
Web client’s request redirected to “nearby” server
– Client gets web site’s DNS CNAME entry with domain name in
CDN network
– Hierarchy of CDN’s 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 web replica
(2)
Client gets CNAME
entry (fault tolerance)
servers
with domain name in Akamai
Client requests
translation for Yahoo
LDNS
(5)
(6)
(1)
Web client
David Choffnes, CDN-based Detouring, INFOCOM 2009
9
Using Redirections as Hints
Good path
Good path
Replica server
… in terms of latency
[SIGCOMM ’06]
Replica server
Host B
Host A
Also good?
Replica server
David Choffnes, CDN-based Detouring, INFOCOM 2009
10
From Redirections to Detouring
• Use CDN redirections as hints to identify
candidate detour points
• Nodes experiencing similar redirection behavior
are good detouring points for each other
• SideStep detouring service
David Choffnes, CDN-based Detouring, INFOCOM 2009
11
SideStep Detouring Service
•
•
•
DNS lookups to determine
redirection behavior
Track ratio of time sent to
each replica server
Distribute these ratios so
others with similar
redirections can find each
other
NA: (r1=0.2, r2=0.8)
NB: (r1=0.3, r2=0.7)
Cosine Similarity: 0.91
DNS
Lookups
Ratio Manager
Distributed Storage
12
SideStep Detouring Service
•
•
•
•
Compare redirection
dynamics to determine
candidate detour points
Because hints are not facts,
validate them via data races
by sending data in parallel
Abandon path with slower
data rate
Export API to enable
applications to use detouring
Client
Stream Manager
Race Manager
DNS
Lookups
Detour
Manager
Ratio Manager
Distributed Storage
13
DraFTP
Java-based client/server FTP suite
Uses SideStep’s I/O stream interface
Changed only 37 lines of code to port an
existing FTP suite
Publicly available under GPL license
David Choffnes, CDN-based Detouring, INFOCOM 2009
14
Evaluation
Goals
– Demonstrate redirections can identify detour points
– Evaluate quality of detour points
Experimental setup
– Deployed on 170 machines worldwide (PlanetLab)
– Initiated transfers between pairs of hosts picked at
random
– Evaluated 13,700 paths
David Choffnes, CDN-based Detouring, INFOCOM 2009
15
Overall Results
Over half of evaluated paths improved
performance
Of those, 75% saw significant performance
improvement (10% higher throughput)
11% more than doubled TCP throughput
All of this is through without additional
network monitoring
David Choffnes, CDN-based Detouring, INFOCOM 2009
16
Wide-area Evaluation
Switch to
faster path
Direct path
Loses race
Race candidate
detour path
Race candidate
detour path
David Choffnes, CDN-based Detouring, INFOCOM 2009
17
Cumulative Results (Best cases)
If given only one shot,
SideStep is almost
always better
Near-optimal
performance
after four tries
David Choffnes, CDN-based Detouring, INFOCOM 2009
18
Effects of different CDNs (Avg performance)
On average,
random can pick
very bad paths
CDN-based paths
are rarely worse
on average
Different CDN names leads to
different performance
David Choffnes, CDN-based Detouring, INFOCOM 2009
19
Conclusion
Advantages to SideStep
–
–
–
–
No path probing to locate detouring points
Requires only periodic DNS lookups
Scalable (O(n) overhead instead of O(n2))
No new infrastructure required
Open Issues
–
–
–
–
Fairness in detouring
Forwarding policies
Incentives for measurement sharing
…
David Choffnes, CDN-based Detouring, INFOCOM 2009
20
Questions?
David Choffnes, CDN-based Detouring, INFOCOM 2009
21