1. Non-pipelined Relay Improves Throughput Performance of Wireless Ad-hoc Networks

2. Puzzle You have 2 glass marbles Building with 100 floors
Both marbles identical in properties
A marble will break at a certain floor k and beyond
In other words: for floors 1 to k-1, if a marble is dropped from that floor, it will not break; for floors k to 100, if a marble is dropped from that floor, it will break
What is the minimum number of “drops” required to find k? (worst case)

3. Ad-hoc Networks
Infrastructure-less wireless networks dynamically formed using only mobile hosts (no routers)
Mobile hosts themselves double up as routers
Multi-hop paths …
Highly resource constrained - limited wireless bandwidth and high interference
Network topology dynamic as all hosts are mobile

4. Outline Pipelined and Non-pipelined relay models
Performance comparisons
Idealized conditions
Practical conditions
Distributed Forwarding Protocol (DFP)
Summary

5. Pipelined Relay (PR)
Source node transmits packets into the network with the objective of keeping the pipe full between the destination and itself
Source pumps packets at a rate sustainable by the end-to-end path
If the hop-length of a flow fi is hi , then at steady state the flow will have hi mini-flows contending in the network

6. PR Illustration

7. Non-Pipelined Packet Relay (nPR)
Source does not perform pipelining of data packets
Source transmits a packet into the network only when the preceding packet arrives at the destination
Each flow has exactly one mini-flow in the network at any given time-instant

8. nPR Illustration

9. Comparison – Idealized Conditions
Throughput Capacity
Fairness
Network Capacity

10. [Model adopted from Gupta & Kumar, 2001]
Network topology : 3-D sphere of unit surface area
n nodes randomly distributed in the network
Every node is a source : total of n flows in the network
Nodes use constant transmission range to keep network minimally connected
[Model adopted from Gupta & Kumar, 2001]

11. Notations W : capacity of the channel (or contention region)
M : Total number of contention regions in the network
lav : Average hop-length of flows
ph(k) : number of flows with hop-length k

12. Throughput Capacity
Sum of the e2e throughputs of all flows in the network
nPR provides more throughput to shorter hop flows than longer hop ones
Every unit of throughput that a longer hop flow sacrifices increases the throughput of lav single hop flows
Thus the average throughput achieved by the flows in the network is increased

13. Throughput capacity (contd.)
Contention level : Number of mini-flows contending for a single channel capacity
CPR : Contention level using pipelined relay model

14. Throughput capacity (contd.)
: Throughput achieved by a single flow using pipelined relaying
TCPR : Throughput capacity using pipelined relaying

15. Throughput capacity (contd.)
nPR favors shorter hop flows at the cost of longer hop flows
Given the hop distribution of the flows to be ph(k), we can derive the throughput capacity,TCnPR, using non-pipelined relaying:

16. Throughput Capacity (contd.)
Ratio of the throughput capacities achieved using the two models :
nPR provides a O(log(n)) improvement

17. Throughput Capacity (contd.)

18. Fairness
Measure of the deviation among the throughputs of the flows in the network
nPR favors shorter hop flows to longer hop ones
But it still achieves proportional fairness since it utilizes the network capacity better
Conventional pipelined model achieves max-min fairness

19. Fairness (contd.)
Related work have proved that proportional rate allocation is to maximize the function,

20. Fairness (contd.)
Raw network resources are allocated in hop-length independent fashion to the flows in the network
Flows obtain end-to-end throughput inversely-proportional to their respective hop-length : hence nPR achieves proportional fairness

21. Fairness (contd.)
Normalized standard deviation of the throughputs of the flows -measure of unfairness
Reduction of number of mini-flows per contention region - decrease in the location dependent unfairness within a contention region

22. Transport Capacity
Number of bits transported in one second by the network
Using PR, the flows are constrained by the bottlenecked region through which they flow
If all the flows in a contention region are bottlenecked in other regions, then the contention region is not fully utilized
nPR achieves the best throughput for a mini-flow in every contention region
nPR performs better as the deviation in mini-channel contention increases.

23. Transport Capacity (contd.)
cmax : maximum contention level in any contention region
Consider a contention region with contention level c
Probability that all the flows are bottlenecked somewhere else is :
Probability that a contention region is under-utilized using PR is given by :

24. Transport Capacity (contd.)
As long as there is deviation in contention levels of the contention regions, nPR will achieve better transport capacity than the PR model
Because of the deviation in contention levels, flows in the PR model do not utilize the capacity of the contention regions fully
nPR, through its decoupling of the scheduling of the mini-flows belonging to each end-to-end flow, fully utilizes the available capacity

25. Transport Capacity (contd.)
Degradation of transport capacity as the mini-channel contention deviates
nPR capacity is independent of the deviation in mini-channel contention because of the decoupling between mini-flows of each end-to-end flow

26. Analysis – Practical Conditions
Impact on MAC utilization
Impact on Load balanced routing
Impact on Route errors

27. Impact on MAC Performance
Under practical conditions, MAC protocols are contention based protocols
Performance decreases with increased load (contending flows) due to the increase in contention
nPR decreases the effective load in the network by decreasing the number of mini-flows contending for network resources
Further there is temporal separation between the mini-flows

28. Impact on MAC (contd.)
Beyond the optimal region the utilization decreases because of the mentioned distributed inefficiencies
nPR operates in the optimal region of the utilization curve by reducing the number of contending mini-flows

29. Impact on Load Balanced Routing
nPR provides spatial separation of the mini-flows
Increases potential for performing load balanced routing
nPR provides better widest-shortest routes than PR because of the reduction in the number of contending flows

30. Impact on LBR (contd.) Performance of LBR when using PR & nPR
nPR achieves better improvements of LBR over normal hop-length based routing

31. Impact on Route Errors
nPR increases the throughput for the end-to-end flows
Reduces transfer time of data for backlogged sources
Hence decreases the vulnerabilities to route failures

32. Summary: nPR’s Performance
Idealized Conditions
Better throughput capacity
Proportional fairness
Better network capacity
Practical Conditions
Increases MAC utilization
Enables load balanced routing
Reduces # of route errors

33. Distributed Forwarding Protocol (DFP)
Realizes non-pipelined relay model under practical conditions
Sits atop the routing layer and beneath the transport layer in the protocol stack
Every packet forwarded by intermediate node passes through the DFP layer
Three key elements of the protocol
Proactive Acknowledgments
Proportional Rate Adaptation
Load Balanced Routing

34. Proactive Acknowledgments
nPR implements the non-pipelined strategy of one packet per flow
Ensure that one data unit is always in transit for every flow in the network
ACKs are sent back when the data packet reaches the temporal mid-point of the flow
Temporal mid-points are calculated by sending timestamps on the packets

35. Proportional Rate Adaptation
MAC protocol optimization curve has an under-utilization region
Ensure that nPR operates in the optimal region of the utilization curve
Monitor the utilization of the network
Increases the number of packets in transit if the network is underutilized

36. Load Balanced Routing
nPR provides temporal separation between the mini-flows
High potential for improvement using Load Balanced Routing (LBR)
Use both hop-length and contention level for routing decisions

37. Performance Evaluation
ns2 based simulations
100 nodes
CBR traffic
1000x1000 network grid
Random way mobility model
DSR used for PR routing
CSMA/CA used for MAC
Data points averaged over 20 simulations

38. Impact of Load
Peak of utilization (of both throughput and network-capacity) occurs at a higher load using nPR
Scalability due to the reduction in the number of mini-flows contributed by end-to-end flows

39. Impact of mobility
Reduction in channel contention, thereby the increase in throughput reduces lifetime of flows using nPR
Mobility has lesser impact on flows using nPR because of the shorter lifetime of flows

40. Summary
Non-pipelining brings in benefits both under idealized and practical conditions
Practical instantiation of the nPR strategy : Distributed Forwarding Protocol (DFP)
Performance evaluation of DFP to compare against conventional forwarding policies
For more information:

41. Non-conversational Transport Protocol (NCTP)
A novel transport protocol that utilizes the concept of nPR
Conventional transport protocols use end-to-end conversational approach
NCTP uses the nPR model where the conversation is hop-by-hop and at any given time instant there is only one data unit in transit for every flow

42. Puzzle Four integers: 1, 2, 3, 4 Three operators: +, -, *
What is the minimum positive integer that cannot be generated using an expression containing all (repetition not allowed) four of the above integers, and three (repetition allowed) operators?
1 * * 4: allowed
: allowed
1 * : not allowed
1 + 2 * 3 : not allowed
: not allowed