Channel Aware Distributed Scheduling For Exploiting MultiReceiver Diversity and Multiuser Diversity

1
Channel Aware Distributed Scheduling For
Exploiting Multi-Receiver Diversity andMultiuser
Diversity in Ad-Hoc Networks A Unified PHY/MAC
Design

• Dong Zheng, Min Cao, Junshan Zhang and P.R. Kumar

INFOCOM 2008 Phoenix, AZ

2
Unified PHY/MAC Design

• Unique challenges in wireless communications
• 1) co-channel interference and 2) fading
• Traditional wisdom treats link losses due to
  fading separately from those incurred by
  interference
• MAC layer scheduling, contention resolution and
  avoidance.
• PHY layer coding/modulation, diversity schemes.
• However, fading can often adversely affect MAC
  layer!
• Indeed, time scales of channel variation and MAC
  transmission are of the same order.
• This calls for channel-aware scheduling!

3
Related Work

• Centralized opportunistic scheduling
• Assumption BS knows instantaneous channel
  conditions of all users.
• Opportunistic picks the user with good channel
  conditions at each slot seeTse00, Borst01,
  Liu-Chong-Shroff01, Viswanath-Tse-Laroia02,
  Andrews01,
• Channel-aware Aloha
• Many-to-one network model contention
  probability is a function of its own channel
  condition Adireddy-Tong05Qin-Berry03.

4
Motivation

• There exist rich diversities in wireless
  communications
• Spatial time frequency multi-user ...
• Open question How to exploit rich diversities
  for ad-hoc communications?
• Challenges in devising channel-aware scheduling
  for ad-hoc communications
• Links have no knowledge of others channel
  conditions even their own channel conditions are
  unknown before contention.
• RBAR Holland-Vaidya-Bahl01, OAR
  Sadeghi-Kanodia-Sabharwal-Knightly 02 are
  perhaps among the first few that exploit channel
  condition for rate-adaptive MAC.
• Adapts the rate based on current channel
  condition
• Our solution Distributed Opportunistic
  Scheduling (DOS) through joint optimal channel
  probing and transmission.

5
DOS for Single Receiver Case

• Consider a single-hop random access network where
  every transmitter has only one receiver.
• A successful channel probing is performed after a
  successful channel contention.
• Suppose after one probing, channel condition
  turns out to be poor. Two options available
• Continue data transmission
• Or, alternatively, let this link give up this
  opportunity, and let all links re-contend.
• Key observation At additional time cost, further
  channel probing can lead to data transmission
  with better channel conditions -gt tradeoff
  between high data rate and probing cost -gt
  optimal stopping rule for channel probing

D
A
E
B
C
F

6
DOS for Multi-receiver Systems

• In multi-receiver/multi-channel systems, we have
  another degrees of freedom multi-receiver/multi-c
  hannel diversity.
• Existing work on exploiting multi-receiver/multi-c
  hannel diversity is for point-to-point
  communications only.
• Systematic approaches to leverage multi-receiver
  diversity in optimizing network performance are
  needed!

7
Probing Phase

• Basic setting a single-hop network with M
  transmitters, each with L receivers.
• Probing takes places in two phases
• In Phase I, random contention is used to acquire
  the channel, and a successful contention ? a
  successful probing to one of the receivers the
  probing cost is a random duration of K .
• In Phase II, specific probing mechanisms are
  carried out to probe the channels to different
  receivers each probing costs a constant time .

8
Assumptions and Objective

• Each transmitter m (1ltmltM) contends with
  probability Pm.
• Let t(n) denote the successful transmitter in the
  n-th round of channel contention, and
  denote the corresponding rate for receiver j,
  j0,1,, L-1.
• WOLG, we impose the following assumption

Objective to maximize the average network
throughput!
9
DOS for Unicast Traffic

• In the unicast case, the transmitter node
  transmits to only one receiver each time.
• Recall that channel probing takes place in two
  phases
• In phase 1, the initial channel probing occurs
  when a transmitter node has a successful channel
  contention
• In phase 2, subsequent probings are performed
  according to specific probing strategies.
• We will see that multiuser/time diversity is
  achieved in Phase 1, and multi-receiver diversity
  is achieved in Phase 2.
• We consider four different schemes of utilizing
  multi-receiver diversity.
• Random selection (same as the single-receiver
  case)
• Exhaustive Sequential Probing With Recall
• Sequential Probing Without Recall
• Sequential Probing With Recall.
• Different strategies lead to different forms of
  the transmission rate and the system time, and
  thus different optimal scheduling policy.

10
Exhaustive Sequential Probing With Recall (ESPWR)

• After a success contention, the transmitter
  probes all the receivers sequentially, and picks
  the best one for data transmission.
• Define

11
Result for ESPWR
12
Sequential Probing Without Recall (SPWOR)

• The transmitter probes its receivers
  sequentially, and stops the probing process once
  it probes a good'' channel, followed by data
  transmission.
• Note that the transmitter is only allowed to
  transmit to the current receiver that is being
  probed.

13
Result for SPWOR
14
Result for SPWOR (Contd)
15
Sequential Probing With Recall (SPWR)

• In SPWR, the transmitter can tx to any probed
  receivers.

16
Result for SPWR (Contd)
17
Result for SPWR (Contd)

• Intuitions
• The monotonic increasing of the initial L-1
  thresholds in SPWR is due to the fact that SPWR
  can recall the previous probed receivers.
• Since channel contention (the channel probing in
  Phase I) costs much more time resources than the
  channel probing in Phase II, thus the threshold
  for further channel probing in Phase I should be
  smaller than the thresholds in Phase II.

18
DOS for Multicast Traffic

• Every receiver requires the same info. from the
  transmitter, and the transmitter can transmit to
  multiple receivers each time.
• The transmitter probes all the receivers.
• Depending on how the reward is defined, we may
  have different multicast scenarios. For example,
• in the first scenario, the reward is defined to
  be the number of ready receivers based on some
  threshold , i.e.,
• in the second scenario, the reward is the sum of
  the rates, i.e.,

• For both scenarios, the optimal scheduling policy
  is same as the single-receiver case with
  different distribution .

19
Iterative Numerical Algorithms

• Observe that the optimal stopping rule are
  multi-threshold policies.
• For every threshold
  , the throughput usually takes the
  form of
• Define and
  .
• We propose the following iterative algorithm
• It can be shown that for any positive , the
  iterates
  generated by the above
  algorithm converge to the optimal thresholds in a
  quadratic time.

20
Example Iterative Alg. For SPWOR
21
Numerical Examples

• We consider continuous rate case based on Shannon
  capacity, i.e.,
• Set
• First, we examine the convergence speed of
  iterative alg. for SPWOR

22
Numerical Examples (Contd)

• Performance comparison between ESPWR and SPWOR

• Some observations
• SPWOR gt ESPWR
• Gain diminishes for SPWOR
• Throughput loss for ESPWR when probing cost
  dominates.

23
Numerical Examples (Contd)

• Performance comparison between SPWR and SPWOR

• Observations
• SPWR gt SPWOR, but difference is very little
• Note that complexity-wise, SPWR gtgt SPWOR.
• Conclusions SPWOR is preferred than SPWR.

24
Numerical Examples (Contd)

• Performance gain of SPWOR over RS

• Observation
• For fixed rho, gain increases and then decreases
  as delta decreases.
• Intuition the difference between random probing
  cost in Phase 1 and constant probing cost in
  Phase 2 diminishes as delta becomes sufficiently
  smaller. As a result, the multiuser diversity
  gain dominates.

25
Conclusions

• Multiuser diversity and multireceiver/multichanne
  l diversity could be jointly utilized by a
  smart distributed probingscheduling algorithm.
• We studied four different probing mechanisms,
  namely,
• 1) random selection, 2) exhaustive
  sequential probing withrecall, 3) sequential
  probing without recall, and 4) sequential probing
  with recall.
• Under the stochastic decision framework, we show
  that the corresponding optimal scheduling
  policies exhibit threshold structures.
• The optimal thresholds could be obtained via
  simple iterative algorithms with quadratic
  convergence speed.
• SPWOR has the best performance in terms of
  throughput and complexity

26

• Thank You!