1. Towards efficient content
dissemination over disruption
tolerant networks
PhD Thesis
Candidate: Amir Krifa, INRIA
Supevisor: Chadi Barakat, INRIA
Monday, April 23 2012
2. Mobile Networking Traffic Growth
Shift
Access to novel applications (social
networks, blogs, music …)
Generation of unprecedented
amounts of mobile data
Complementary architecture ?
Second class
customer
2/39
First class
customer
3. The DTN concept
Take advantage of increasing mobile nodes resources
Rely on nodes mobility to route messages through
disconnected networks
A node can be a human carrying a laptop or SmartPhone, a bus, a car, etc
At the opposite of existing networks, no end-to-end path is
required during the communication
Hop-by-Hop networking
Message replication
3/39
4. DTNs: Not as futuristic as it
sounds !
World’s First Flying FileSharing Drones in Action @
GLOW Festival 2011
Netherlands
Wildlife tracking
systems:
ZebraNet, Env.
Monitoring, etc
4/39
5. Challenges
Challenges:
Disruption and dynamic environment
-> long-term storage +
replication (Routing Algorithms: Global Optimal, Epidemic, Spry and
Wait, etc …)
RAPID
By Levine et al.
Long-term storage + replication
– -> buffers congestion (Drop
Policies: Drop Oldest, Drop Last …)
– -> lack of Bandwidth (Scheduling:
FIFO …)
Mobile devices controlled by rational people
5/39
-> selfishness
6. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
6/39
7. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
7/39
8. Methodology
Suppose first global knowledge
Take a global routing metric as the delay or delivery rate
Find what is the best policy to drop and schedule
Which message should be dropped/scheduled first and that leads to
the best gain in the considered global metric,
Model this gain as a per-message utility function.
Try to estimate the global knowledge using global
information BUT on old messages …
8/39
9. Case of delivery rate
Message has the same limited lifetime (TTL)
Suppose global knowledge on m and n
Assumption: meeting times have an exponential tail
In case of congestion, the global delivery rate is :
At least one copy of message i
Will be delivered
Message i is not
delivered yet
m (T )
K(t) K(t) m (T )
i i * 1 exp λn (T )R i i
DR P 1
i i1 L 1
i i i L 1
i1
Message i will
be delivered
9/39
Message i has been
already delivered
10. Case of delivery rate
We differentiate:
k(t) P
i * Δ(n (T ))
Δ(DR)
i i
i1 ni(Ti)
Δ(n (T ))
i i
-1 : drop
0 : no action
+1 : replication
GBSD (DR):
The best message to drop is the one having the minimum
partial derivative:
m (T )
1 i i λR exp λn (T )R
i i i
L 1 i
And the message to schedule first is the one maximizing
it
10/39
11. Case of delivery delay
GBSD (DD):
The best message to drop is the one having the minimum
partial derivative:
1 1 mi(Ti)
n2(T )λ L 1
i i
And the message to schedule first is the one maximizing
it
For more details:
Amir Krifa, Chadi Barakat, Thrasyvoulos Spyropoulos, “Message Drop and Scheduling in DTNs: Theory and
Practice”, in IEEE Transactions on Mobile Computing (TMC).
11/39
12. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
12/39
13. Distributed version:
How to calculate n and m ?
n = number of copies of a message
m = number of nodes that have seen the message
Flood information on messages (like in RAPID by UMASS)
takes long time to converge
The information is stale by the time it reaches everyone
Our solution:
Still flood information on messages
BUT, Estimate n and m at a given elapsed time from what has
happened to old messages at the same elapsed time
13/39
14. Distributed version (DR)
Suppose m and n follow two random variables M and N
Estimated delivery rate = Mean delivery rate
ˆ
ˆ
m(T) 1 exp λn(T)R m(T) E1 M(T) 1 exp λN(T)R M(T)
ˆ
1
i L 1
i L 1
L 1
L 1
We set the estimator of m to its expectation (justified by a Gaussion
distribution)
ˆ
m(T) m (T) EM T
14/39
15. Distributed version:
Message utility expressions
(H
BS
D)
For the delivery rate:
M(T) exp λR N(T)
λR E 1
i
i
L 1
ist
2
E L 1M(T)
N(T)
λ L 1 L 1 m (T)
H
For the delivery delay:
or
y
Ba
se
d
SD
Expectation calculated by summing over old messages
15/39
16. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
16/39
17. Validation Setup
DTN architecture added to the NS-2 simulator
Random Waypoint and KAIST real mobility trace
Wireless Range=100m,
CBR sources, random sources and destinations,
Each node maintains a buffer with a capacity of 20 messages
Mobility model
KAIST
Random Waypoint
Simulation duration (h):
Simulated Surface (km2):
24
-
7
3*3
Number of nodes:
50
70
Average speed (m/s) :
2
-
TTL (h) :
4
1
1440
360
Interval CBR (s) (10/TTL):
17/39
18. Delivery Rate
HBSD outperforms existing protocols (RAPID and Epidemic based on
FIFO/drop-tail) and performs close to the optimal GBSD
Random Waypoint
Almost 60% gain over
RAPID
18/39
KAIST Traces
Close by 14%
19. How HBSD utilities look like ?
Reduce the number of sources to 15 and decrease the CBR rate of
sources from 10 to 2 messages/TTL (Low congestion regime)
Schedule Youngest First
Drop Oldest
For a lightly loaded network, things are easier and simple policies
can be applied.
19/39
20. How HBSD utilities look like ?
We fix the number of sources to 50 (high congestion regime)
help the message
over younger ones
penalize – help - penalize
prefer younger ones
For a highly loaded network (complex function)
20/39
21. Implementation / Web page
And is also available for the DTN2 architecture as an
external router (in C++)
Code has been recently tested in the Scorpion testbed
at the University of California Santa Cruz
Code, papers, presentations are available at:
http://planete.inria.fr/HBSD_DTN2/
21/39
22. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
22/39
24. New context: Content centric
Node Centric vs Content Centric communications
Madonna M. Album
Muse +
Madonna M.
Album
Madonna +
Muse M.
Madonna M.
Album
Album
Track – 1
Track – 2
…
Track – 1
Track – 2
…
Madonna M.
Album
Muse M.
Album
Track – 3
Track – 4
…
Muse Madonna
Muse + M.
Album
M. Album
Track – 1
Track – 2
…
Track – 1
Track – 2
…
Selfish
user !
[Question]: which channels and how
Store Local and a node Channels
much of each shouldForeign carry in !its!
Make Everybody happy ? -> Store Local and Foreign Channels
buffer, Selfishto maximize its future
Block so as users ! (TFT)
reward ??
24/39
25. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
25/39
26. MobiTrade
MobiTrade turns each node into a merchant fetching the content that
has the highest chance to be sold to its good clients
MobiTrade calculates one utility per channel that defines:
The optimal amount of storage to allocate / channel Drop policy +
Scheduling policy
Node Storage
Xi
αi B
B
Xj
*
j
Channel 1
Channel 2
Channel 3
Channel 4
MobiTrade approximates the Optimal U. based on the amount of
exchanged content per channel @ each meeting while ensuring that
selfish users are blocked
For more details:
Amir Krifa, Chadi Barakat, Thrasyvoulos Spyropoulos, “MobiTrade:
Trading Content in Disruption Tolerant Networks”, in proceedings of ACM CHANTS, Las Vegas,
September 2011.
26/39
27. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
27/39
28. Collaborative experimental scenario
MobiTrade architecture added to the NS-3 simulator
Synthetic mobility model HCMM
50 users distributed into 5 groups. The simulation area is divided into a 10*10
grid of cells (5000 meters wide).
Wireless Range = 60m .
Simulation Scenario
Nbr. Of Users:
Requested CH(s)/User:
50
2
Size of CH(s):
Compare to Podcasting
(PodNet project)
20
Average Delivery Rate (DR): amount of content received for channels a node
requested / total amount of content generated for these channels
28/39
29. How MobiTrade performs in a
Collaborative scenario ?
MobiTrade efficiently outperforms the two versions of Podcasting
TFT causes a drop in performance among CU
Drop of 6%
Importance of
FC(s)
Almost 2x gain
29/39
30. Experimental scenarios including
selfish users
Scenarios
SS1
Nbr. Of Users:
40 CU + 10 SU
40 CU + 10SU
CU: 2/20 – SU: 2/10 (SU
and CU channels differ)
CU, SU: 2/20 (among
same channels)
CU: 20 – SU: 40
Requested CH(s):
Size of CH(s):
SS2
CU, SU: 20
We deem such scenarios as the norm rather
than the exception in the real world
30/39
31. Does MobiTrade keep the system
resources safe ?
Enabling the TFT mechanism blocks selfish users and makes MobiTrade redispatch/reuse the saved resources among the channels shared by
collaborative users
SS1: CU ask for
2/20 channels and
SU ask for 2/10
different channels
Impact on
selfish users
Impact on collaborative users
31/39
32. Does MobiTrade keep the system
resources safe ?
When TFT is used, the performance of collaborative users is not harmed,
while the one of selfish users drops severely, by up to 2x for a storage of 110
contents.
SS2: each user ask
for 2/20 channels
Impact on SU: Drop by up
to 2x
No Impact on CU
32/39
33. Implementation / Web page
MobiTrade available for the Android platform
Code, papers, presentations are available at:
http://planete.inria.fr/MobiTrade/
App Screenshots:
33/39
34. Outline of the talk
Point-to-point (Node Centric) communications
Optimal solution that requires global knowledge (GBSD)
Distributed version that works in practice (HBSD)
Validation results
Point-to-multi-points (Content Centric) communications
The content centric context
MobiTrade: optimal resources management solution
Validation results
Conclusion and Perspectives
34/39
35. Conclusion
A deep study of content sharing in DTN(s) for both:
Point-to-point communication model
Point-to-multipoint communication model
New resources management policies in two versions:
Optimal one that is based on global knowledge
Practical one that efficient approximate the optimal policy
GBSD/HBSD
MobiTrade
Validation via simulations based on synthetic mobility models and
real mobility traces
Implementation on real word environments (DTN2 and Android)
35/39
37. Perspectives
With respect to GBSD/HBSD:
Tune the utilities of our resources management policies in order to
take into account different messages sizes ...
Study and design a congestion level detection mechanism to be able
to switch efficiently between resources management policies …
With respect to MobiTrade:
Implementing the MobiTrade protocol for other types of devices and
experiment with real large scale communities of users...
Consider more complex content structures …
Study of the needed mechanisms to control possible advanced
malicious attacks and behaviours that could impair MobiTrade
content sharing sessions ...
37/39
38. References
Amir Krifa, Chadi Barakat, Thrasyvoulos Spyropoulos, “MobiTrade: Trading Content in
Disruption Tolerant Networks”, in proceedings of ACM Mobicom Workshop on Challenged
Networks (CHANTS), Las Vegas, September 2011.
Amir Krifa, Chadi Barakat, Thrasyvoulos Spyropoulos, “Message Drop and Scheduling in
DTNs: Theory and Practice”, in IEEE Transactions on Mobile Computing.
Amir Krifa, Chadi Barakat, Thrasyvoulos Spyropoulos, "Optimal Buffer Management Policy
for Delay Tolerant Networks", in proceedings of the 5th IEEE Conference on Sensor,
Mesh and Ad Hoc Communications and Networks (SECON 2008), San Francisco, June 2008.
(CA), June 2008. ---- BEST PAPER AWARD
Amir Krifa, Chadi Barakat, Thrasyvoulos Spyropoulos, "An Optimal Joint Scheduling and
Drop Policy for Delay Tolerant Networks”, in proceedings of the WoWMoM Workshop on
Autonomic and Opportunistic Communications, Newport Beach (CA), June 2008.
38/39