Performance of group communications over adhoc networks

1
Performance of group communications over ad-hoc
networks

• Marc Mosko and J.J. Garcia-Luna-Aceves
• Department of Computer Engineering
• U.C. Santa Cruz
• ISCC 2002
• Sicily, July 1-4 2002

2
Presentation Outline

• Goals motivation
• Overview of model
• Comparison to simulation
• Further research
• Conclusion

3
Goals Motivation

• Goals
• Steady-state model of broadcasting in
  contention-based (ALOHA) ad-hoc networks that
  accounts for correlated packet loss due to
  hidden-terminal collisions for reliable (NAKs)
  and unreliable broadcast.
• Use model to predict performance of ad-hoc
  network using FEC coding for inter-packet erasure
  correction for broadcast/multicast distribution
  (not in present work).

4
Model Introduction

• Radio Link Layer
• Common channel radio, no capture, no loss from
  noise
• May not transmit and receive at same time
• 802.11 / ALOHA broadcasts (i.e. unreliable)
• Topology cluster tree formation
• Nodes grouped in fully connected components
• Components connected in tree
• Each component has a single cluster head that is
  in hearing range of parent, repeats for other
  cluster members
• Subset of other cluster members repeat for other
  components

5
Binary cluster tree
Cluster tree allows Repairs through
siblings. Only allows 1 repair attempt even if
more siblings.
6
Rate-based fluid model

• Infinitesimal bit rate
• Model based on rates, not information. No
  queuing.
• Assume all rates are Poisson (or modulated P.)
• Total in-bound rate l l1 lk
• Prob zero arrivals (1-l) Dt
• Prob exactly one arrival l Dt
• Out-bound rate in next Dt will be collision-free
  part of in-bound rate in current Dt.

7
Filtering process

• mi out-bound rate from i
• ei,z, info density i to z
• Noise part of m that add no new information
• needs non-local terms
• pi,z , prob z hears i
• pi,z 1 S mk,
• k ? N(z)z-i
• li,z, useful rate i to z
• li,z mi ei,z, pi,z
• lz S li,z

8
Unreliable Broadcast

• Source m0 driving rate
• Difference eq.
• mZ(tDt)- mZ(t)lZ(t)-mZ(t)Dt
• Differential eq. (i ? N(z))
• m?Z(t) -mi(t) S mi ei,z, pi,z
• Initial conditions anything reasonable for
  mZ(0), such as zero or m0.
• Only m?Z is differential.

• mi out-bound rate from i
• ei,z, info density i to z
• pi,z , prob z hears i
• pi,z 1 S mk,
• k ? N(z)z-i
• li,z, useful rate i to z
• li,z mi ei,z, pi,z
• lz S li,z

9
Reliable broadcast

• NAK-based, 1 NAK per missed info quanta
• ??Z(t) lp (t) lz (t) , p parent of z
• Retransmission rate
• z?Z(t) -zZ(t) S ?i (t)pi,z (t) , i ? C(z)
• Problem Should use non-local terms
• ??Z(t) lp (t-Dt) lz (t)
• z?Z(t) -zZ(t) S ?i (t -Dt)pi,z (t)
• l?z(t) also needs them

10
Reliable broadcast equations

• At source node with driving rate m0
• m?0 m0 mo z0
• At other nodes
• m?z lz mz ?z zz
• All terms are differentials to be solved each Dt.

• mi out-bound rate from i
• ei,z, info density i to z
• pi,z , prob z hears i
• li,z, useful rate i to z
• ?z NAK rate at Z
• zZ Retrans. rate at Z

11
Comparison to simulation
Unreliable broadcast
Reliable broadcast
12
End-to-end seqno NAKs

• If you forward out-of-order packets immediately
  (min. delay), you may create nak storms.
• Snooping limited to 1-hop radius
• Delays do not work well and if based on hop count
  from source may cause long latency
• We used per-hop seq. no. in paper, solved
  problem. Better solution would be to use
  forward NAK notification bit in packet to
  indicate it was already NAKed.

13
Future work

• Consider non-local terms.
• Validate against other network topologies.
• Allow multiple repair attempts in cluster, this
  appears to be main problem in larger clusters.
• Use results in FEC code calculations similar to
  ICNP 2000 work.
• What rate codes make sense in this environment?
• Packet lengths packet pacing.

14
Conclusion

• Classifications terms
• Definitions of symbols seem to make sense, rate
  equations may be easily read.
• For binary cluster tree, results are not too bad,
  even without non-local terms.
• Exposed issue of using end-to-end sequence
  numbers for NAK generation.