Title: On the Benefit of Random Linear Coding for Unicast Applications in Disruption Tolerant Networks
1On the Benefit of Random Linear Coding for
Unicast Applications in Disruption Tolerant
Networks
- Xiaolan Zhang, Giovanni Neglia
- Jim Kurose, Don Towsley
2Disruption/Delay Tolerant Network (DTN)
- Intermittent connectivity
- Limited/no infrastructure gt ad hoc network
- Mobility and sparse settings gt frequent
- partition
- Examples
- Vehicular network DakNet, UMassDieselNet
- Sparse mobile sensor networks ZebraNet, under
water sensor networks - Disaster relief team/military ad hoc network
- Resource constrained power, bandwidth, storage
- We focus on DTN with random contact
3Routing unicast packets in DTN
- Trade-off delivery delay versus num. of copies
made for each pkt - Why limit copies ?
- Transmission power affect network lifetime
- Bandwidth consumption affect network throughput
src/dest transmission
Probabilistic forwarding
Average Delay
2-hop forwarding
epidemic routing
Average Copies Sent
4Network Coding Benefit
- Previous work studied network coding benefit in
wireless network (including DTN) - Broadcast/multicast network coding increases
energy efficiency Fragouli06, Widmer05, lun04,
Wu04 - Unicast network coding increases throughput by
leveraging broadcast nature of wireless medium
katti et al 05, Wu et al 04 - Our contribution network coding (RLC) improves
delay vs. energy efficiency tradeoff for unicast
application in DTN.
5Outline
- Background
- Random Linear Coding, Network Setting
- Benefit under single generation case
- Summary and Future Work
6Random Linear Coding Ho et al 03
- Packet a vector of symbols in Fq a finite field
of size q - Pkts form generations of size k m1,m2,,mk
- Random linear combination of pkt m1,m2,,mk
- Network nodes
- Collect and store encoded pkts
- Forward random combinations of currently stored
encoded pkts, together with coefficients - Decodable on collecting k independent
combinations
encoded pkt
7Network Setting
- A DTN of N mobile nodes running unicast app.
- Pair-wise Poisson meeting process with rate ß
groenevlt05 - Bandwidth constraint b pkts/per contact in each
direction - Buffer constraint node can store limited num. of
relay pkts
8Forwarding/Coding schemes compared
- Schemes
- random uniformly randomly selects pkt to forward
(and drop) - RR_random src selects pkt to send in round robin
manner relay nodes perform random scheduling - RLC scheme apply RLC to a block of K pkt
- Common delivery notification mechanism
- On first delivery of pkts/generations,
pkt/gen-delivered info generated, propagated - Performance metrics
- Block delivery delay (Dblock) time to deliver
the last pkt (other metrics avg. pkt delay,
in-order delay) - Num. of transmissions made
9Outline
- Background
- Random Linear Coding, Network Setting
- Benefit under single generation case
- Summary and Future Work
10Single Generation bandwidth- constrained case
- Single block of K pkts
- Generated at same time
- Sent from a source to a dest
- Contact graph
- Vertex network nodes
- Edge contact between nodes, labeled with contact
time b directed edges for bandwidth b contact - Time-respective path 1-gt3-gt4, 1-gt2-gt4
N4, b1 pkt/contact/dir
1
t7
t3.5
t1.2
3
2
t10.2
t23
4
Minimum block delivery delay time to have K
edge-disjoint time-respective paths from src to
dest
11Single Generation Case (contd)
- Can this minimum delay be achieved ? For this
example, - RR_random scheme achieves min. delay with prob.
0.5 - RLC achieves min. delay with prob 1-1/q (q size
of the finite field) - RR_random chooses from K packets RLC chooses
from qK-1 different combs - Similar in spirit to algebraic gossip Deb and
Medard 04
m1
m2
m2
m2
m1
m2
1
c1
t7
c2
c3
t3.5
t1.2
3
2
m2
c3
c12
t10.2
t23
Nothing to send !
4
With prob. 1-1/q, c12 and c3 independent !
12Different delay metrics
- Simulation result
- N101,K10,q701, ß0.0049
RR_random avg. delay
RR_randomblock delay
RR_randomin-order
RLC
RLC tradeoff average pkt delay for block delivery
delay
13Improvement in block delivery delay
- Larger coding benefit under smaller bandwidth
N101, K10, q701, ß0.0049
14Num. of transmissions made
- RLC makes more transmissions
- More often to have info. to exchange when nodes
meet - Delivery notification starts later
- Can RLC decrease block delay without increasing
num. of transmissions ?
N101, K10, q701, ß0.0049
15Limiting Copies token scheme
- Token scheme src assigns fixed num. of tokens to
pkt/generation - Decrease tokens after each transmission
- Reallocate tokens when nodes meet (RLC in
proportion to ranks) - No copy when no token left
- Spray and wait spyropoulos05, small05
N101, K10, q701, ß0.0049
RLC achieves smaller block delivery delay than
non-coding scheme with similar transmission num
16Buffer-constrained Case
N101,K10,bw1,q701
- When relay buffer size decreases
- RR_random/random delay increases substantially
- RLC delay increases slightly
- More copies made under RLC
- under token scheme, RLC achieves better delay vs.
transmission tradeoff than non-coding scheme
Buffer2
17Outline
- Background Motivation
- Random Linear Coding, Network Setting
- Benefit under single generation case
- Summary and Future Work
18Summary
- Benefit of RLC for unicast app. in DTN
- Achieves smaller block delay for given num. of
copies, especially with limited buffers - Other results
- Applied RLC to a single block of pkts
- From diff. src to same dest
- From diff src to diff dest
- Multiple generation cases up to 22.5 reduction
in block delivery delay - Insight benefit due to increased randomness of
RLC (coupon collector problem)
19Future Work
- Practical issues
- RLC overhead
- Small application message size
- Quantify coding benefit analytically
- RLC benefit under real mobility traces
- Can RLC increase network throughput for this
network setting ?
20Questions ? Comments ?
21Benefit of RLC
- Num. of ways of choosing K linearly independent
combination of K pkts in a finite field is (deb
and Medrad04) - Lemma 2.1
22Multiple Generation Case
- Uniform traffic load
- N flows with each node being src of a flow and
dest of one other flow - Each flow generates blocks of K10 pkts according
to Poisson process with rate ? - Randomized scheduling
- non-coding randomly select pkts the other node
does not have - RLC randomly select gen. that has useful info.
to other node - Randomized buffer management
- Non-coding randomly select pkts to drop
- RLC randomly select gen. to compress
23Without limiting transmissions
- Under low traffic rate (?)
- RLC achieves smaller block delay than non-coding
- Under higher traffic rate (?), RLC incurs larger
block delay than non-coding - More pkts to choose from for non-coding
- Higher contention for bandwidth under RLC
randomized scheduling does not favor new gen.
N101,K10, ?0.00045,bw1
24The need to limit copies
N101,K10, ?0.00045,bw1
- One solution limit copies when network is loaded
- Optimal per-packet token num. under certain
traffic rate - Too small gt some contacts not utilized
- Too big gt contention
- With higher traffic rate, non-coding benefits
from limiting copies too - Up to 22.5 reduction of block delivery delay
Buffer5