Upload
mahdis
View
25
Download
0
Embed Size (px)
DESCRIPTION
Dept. of Computer Science & Engineering. University of California - Riverside. DNSR: D omain N ame S uffix-based R outing in Overlay Networks. CS202 – Advanced O.S Spring ’03. Demetrios Zeinalipour-Yazti. Introduction. Most overlay networks don’t match the underlying topology - PowerPoint PPT Presentation
Citation preview
DNSR:DNSR: DDomain omain NName ame SSuffix-based uffix-based RRouting in outing in Overlay NetworksOverlay Networks
Demetrios Zeinalipour-Yazti
Dept. of Computer Science & Engineering.
University of California - Riverside
CS202 – Advanced O.S Spring ’03
IntroductionIntroduction
AverageRTT=9ms4 Router Hops
AverageRTT=140ms19 Router Hops
AverageRTT=163ms19 Router Hops
pc-62-30-117-83-cr.blueyonder.co.uk
p237-165.yahoo.co.jp
roc-24-169-109-208.rochester.rr.com
12-224-0-236.client.attbi.com
AverageRTT=46ms13 Router Hops
AverageRTT=130ms22 Router Hops
66-215-0-xxx.oc-nod.charterpipeline.net
66-215-0-xxx.oc-nod.charterpipeline.net
London
AtlantaRochester
Tokyo
Riverside
• Most overlay networks don’t match the underlying topology
• Transcontinental connections are expensive.• It would be desirable to keep the bulk of the
P2P traffic within the same domain
• Analyzing the Gnutella Network [D. Zeinalipour & T. Folias, cs204 Course Project]
– We analyzed ~300,000 IP addresses.– 58.73% of Gnutella IPs belongs to only 20 ISPs.– Organizing Peers into domains rather than
loosely interconnecting them might be feasible
MotivationMotivation
• DNSR:Domain-Name Suffix-based Routing – Decentralized Routing Algorithm that attempts to keep
P2P traffic within the same domain.– DNSR defines three Level factors per peer:
• Sibling Factor (sfi)• Parent Factor (pfi)• Children Factor (cfi)
– DNSR also defines a similarity function -suffix
DNSR IdeaDNSR Idea
ni=cs6368146-17.austin.rr.com
nj=node-17.rr.com
sfi=0.8
pfi=cfj=0.1di=10
*.austin.rr.com
cfi=0.1
• Given that each node maintains the Level Factors we end up with a semi-hierarchical topology.
DNSR TopologyDNSR Topology
Level 1 Level 1
Level 2
n1.ucr.edu
n1.cs.ucr.edu
n2.cs.ucr.edu
n3.cs.ucr.edu
n2.ucr.edun6.ucsd.edu
n2.cs.ucr.edu
n2.cs.ucr.eduLevel 2
n2.cs.ucsd.edu
n3.cs.ucsd.edu
n2.cs.ucsd.edu
DNSR Topology Instancedi = 3--------------pfi = di / 3sfi = di / 3cfi = di / 3
JoiningJoining a DNSR Topology a DNSR Topology
Level 1 Level 1
Level 2
n1.ucr.edu
n1.cs.ucr.edu
n2.cs.ucr.edu
n2 = n3.cs.ucr.edu
n3 = n2.ucr.edun4 = n6.ucsd.edu
n2.cs.ucr.edu
n1 = n2.cs.ucr.eduLevel 2
n2.cs.ucsd.edu
n5 = n3.cs.ucsd.edu
n2.cs.ucsd.edu
Domain-Name Lookup in DNSR Topology
Lookup
OverlayConnection
nj = n5.cs.ucsd.edu
Lookup Response
• A node obtains a random list from an out-of-band mechanism (e.g. hostcache).
• It probes for “best” entry point with Lookup
• Searching can be done with a variety of techniques (BFS, Random BFS,….)
• The bottom-line with all techniques is that the bulk of the traffic remains within the same domain
SearchingSearching a DNSR Topology a DNSR Topology
Level 2
Level 1n1.ucr.edu
n1.cs.ucr.edu
n2.cs.ucr.edu
n2 = n3.cs.ucr.edu
n3 = n2.ucr.edu
n2.cs.ucr.edu
n1 = n2.cs.ucr.edu
Searching in DNSR Topology using BFS
QUERY
QUERYHIT
Scenario• We generate a DNSR topology and a Random
Topology of 1000 nodes with following distributions
• We deploy the 1000 real nodes on 25 machines• We run a number of queries and observe the
distribution of hosts contacted in each case.
Experimental SetupExperimental Setup
• Each node reads its settings from the filesystem
• All nodes are launched concurrently with ssh public/private keys making the bootstrapping easy.
Experimental EvaluationExperimental Evaluation
• We connect to a .rr.com host
• Submit 40 queries and count the number of hosts contacted C.
• In a random topology C matches the actual distribution of hosts
• In a DNSR topology most of the hosts contacted are .com hosts for level 0.
Experimental ResultsExperimental ResultsLevel 0
• For Level 1 we can see that in DNSR topology we will contact more .rr.com hosts than with a random topology.
• Therefore more traffic remains within same domain
• Sibling Factor was 66%. If it was larger then 24% would be larger
Experimental ResultsExperimental ResultsLevel 1
DNSR Demo Follows…DNSR Demo Follows…
Dept. of Computer Science & Engineering.
University of California - Riverside