Aucun titre de diapositive

1
Mapping Logical Network Routing to Physical
Network Routing
Valentin Mesaros (UCL), Luc Onana (KTH) Peter
Van Roy (UCL), Seif Haridi (KTH)

Most ideas taken from article NetProber A
component for enhancing efficiency of overlay
networks in P2P systems, to appear in IEEE
P2P02.
2nd PEPITO workshop, Stockholm Jun. 2002
2
Contents

• The mapping problem
• Principles of our solution NetProber
• Some simulation results
• Related work
• Conclusions
• How far can we go?
• Further work

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
3
The Mapping Problem

• general facts (the current status)
• - the logical nodes do not (want to) have
  knowledge about underlying network
• - the logical network is usually randomly
  constructed
• problems
• - mismatching - depending on a QoS param
  (e.g., hops, latency, bandwidth)
• consequence
• - inefficient use of the underlying network
  Ripeanu01
• many more physical hops needed for
  logical routing than for physical routing
• physical links may be employed more
  times than necessary (i.e., stressed links)
• - increase of latency
• awkward randomly connections lead to
  increasing the communication latency

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
4
Logical Network over Physical Network graphical
notation
BL
AL
GL
FL
B
D
G
E
A
F
C
H
B
A
G
2
F
3
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
5
The Mapping Problem example

• - consider the Gnutella overlay network
  A,B,C,D,E
• a broadcast msg. (TTL4) from A passes 4 times
  link (B,D) (stressed link)
• - there are much more physical messages generated
  in case (a) than in (b)

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
6
Our Solution principles

• use information about the underlying network
• - of hops between two nodes
• - latency/bandwidth between nodes
• peers try to find the physically closest
  neighbors (proximity selection)
• a complementary component (NetProber) is
  associated with each peer
• - NetProber tries to collect/process info
  about the underlying network

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
7
Our Solution algorithm

• core algorithm
• - if the neighbors neighbor is closer to me
  than my neighbor, connect to it
• and disconnect from my neighbor
• - can be run with different depths
• location where to run the algorithm
• - each peer runs the algorithm with respect to
  each of its neighbors
• control when to run the algorithm
• - run the algorithm when a certain event
  occurs (e.g., new connection, latency
• change, periodically)

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
8
Overlay Goodness

• d-good node
• - a node N such that in every path of length d
  from N,
• the closest node is already a neighbor of N
  
• - it is computed with respect to a certain
  depth d
• overlay network goodness Gd
• - the ratio of the number of good nodes over
  the number of nodes
• in the overlay network

C is a good node G2 0.2
E
A
D
2
B
C
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
9
NetProber

• component for enhancing efficiency of overlay
  networks and
• increasing its goodness
• component to be attached to peers
• responsible for finding a closer neighbor
• make use of the underlying network (IP)

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
10
NetProber how it works
G
F
npG
npF
A
C
2
npC
npA
B
npB
D
npD
Msgs exchanged

• A tries to find a closer neighbor w.r.t. C,
• run an algorithm instance (depth 2),
• npA suggests A to switch from C to B

A -gt npA FindBetterNeighbor(C)
npA -gt npC NetProbeRequest
npC -gt npA NetProbeReply(Neighbors, Param)
npA -gt npB,npF NetProbeRequest
npB,npF -gt npC NetProbeReply(Param)
npA -gt A Suggestion(C,B)
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
11
The Mismatching Factor
- Let R(i,j) be


- Let NL(k) be the set of nodes reached from node
k doing a broadcast level L
NL(k) j i0, i1, i2, , im logical
nodes such that i0k, m?L, (ir, ir1) logical
edge
- Let MFL(k) be ? R(k,j) ,
where j ? NL(k)
- The ideal value for MFL(k) is 1
F
G
Example (MF of node A)
A
R(A,B) 5
C
2
R(A,D) 3
B
D
MF3(A) 2.2
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
12
Simulations

• the simulator
• - simulate the physical and logical networks,
  and NetProber in Mozart
• - physical nodes are connected via Oz ports
• - the logical nodes are attached to physical
  nodes
• the physical network
• - generated with Inet-3.0 (from real Internet
  data)
• - nodes 3100, edges 4904, graph
  diameter 9
• the logical networks
• - randomly generated a new node connects to a
  node already in the network
• - nodes 50, 150, 250, 500, 1000

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
13
Related Work

• entry optimization in Tapestry Zhao01
• - switch from primary to closer (in latency)
  secondary neighbors
• - procedure run at joining, and periodically
  afterwards
• - costly in time
• binning applied in CAN Ratnasamy02
• - peers organize in bins with respect to their
  distance (in latency) to
• certain well known landmarks
• - there is no optimization inside a bin
• overlay efficiency optimization in Narada
  Chu00
• - each member periodically probes (in latency)
  every other
• - switch to a new neighbor for better
  performances
• - it performs badly for large groups ( gt 300 )

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
14
Different distance metrics
advantages
disadvantages
of hops

• - limited access
• - rapidly-oscillating
• do not lead to accurate results
• hard to provide
• difficult to measure

• - one-shot measurement
• accurate
• easy to process
• - easy to access
• practical
• - practical

latency
bandwidth
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
15
Conclusions

• randomly constructed P2P networks can make
  efficient use of the underlying network
• possible mapping measures
• - overlay goodness
• - mismatching factor
• complementary adaptive component for peers
  NetProber
• the use of of hops as a metric gives good
  results in simulations
• - the overlay goodness improves
• - less physical msgs. are generated when
  addressing nodes in the system
• optimality
• - running the algorithm with depths 2, 3 gives
  significant improvements

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
16
How Far Can We Go

• optimality
• - for tree physical networks, NetProber with
  depth 2 gives the global optimum
• (best with given number of logical edges)
• - for physical networks with redundant paths,
  depth 3 and higher gives significant
• improvements

for depth 2, A does not see that E is close
A
D
E
w1
w2
B
C
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
17
How Far Can We Go (II)

• leaf effect
• - unless the logical nodes are located on the
  roots of the sub-trees of
• the physical network, MF can not reach 1

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
18
Further Work and Open Questions

• Adapting NetProber for faults (site problems,
  network problems)
• What is the correct goodness measure(s)?
• how to better combine different distance
  metrics?
• What should be put in the DSS?

Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002
19
Bibliography
Y.Chu, S.Rao, and H.Zhang. A case for end system
multicast. In Proc. of ACM SIMETRICS, June 2000
Chu00
S.Ratnasamy et al. Topologically-aware overlay
construction and server selection. In Proc. of
IEEE INFOCOM, June 2002.
Ratnasamy02
M. Ripeanu. Peer-to-Peer architecture case study
Gnutella Network. Tech. Report, University of
Chicago, July 2001.
Ripeanu01
B.Zhao, J.Kubiatowicz, and A.Joseph. Tapestry An
infrastructure for fault-tolerant wide-area
location and routing. Tech. Report
UCB/CSD-01-11041, U.C. Berkeley, April 2000.
Zhao01
Mapping logical routing to physical routing - 2nd
PEPITO workshop, Stockholm Jun. 2002