Labor Marketplace Applications: Human Computer Interaction KAIST KSE Uichin Lee 9/26/2011.
Alexander Afanasyev Tutors: Seung-Hoon Lee, Uichin Lee Content Distribution in VANETs using Network...
Transcript of Alexander Afanasyev Tutors: Seung-Hoon Lee, Uichin Lee Content Distribution in VANETs using Network...
Alexander Afanasyev
Tutors: Seung-Hoon Lee, Uichin Lee
Content Content Distribution in VANETs using Network Distribution in VANETs using Network Coding: Coding:
Evaluation of the Generation Selection Evaluation of the Generation Selection AlgorithmsAlgorithms
March 2, 2009
Applications ◦ Software updates and patches (e.g., navigation map, games)◦ Multimedia data downloads (e.g., videos, news, etc.)
Content Distribution in Content Distribution in Vehicular Ad-Hoc Networks (VANETs)Vehicular Ad-Hoc Networks (VANETs)
◦ High mobility (i.e., highly dynamic networks)◦ Error-prone channel (due to obstacles, multi-path fading, etc.)
Content Distribution Challenges Content Distribution Challenges
CarTorrent: BitTorrent-like CarTorrent: BitTorrent-like Cooperative Content Distribution in Cooperative Content Distribution in VANETsVANETs
Web Server
A file is divided into pieces
Exchange pieces via Vehicle-to-Vehicle Communications
Download a file (piece by piece)
Problem: Peer & Piece selection coupon collection problem
Cannot complete download!
Not useful!
Using Network Coding: Using Network Coding: CodeTorrentCodeTorrent
Web Server
A file is divided into pieces
Any linearly independent coded packet is helpful
1 more?
Network Coding ProblemProcessing Overhead
Delay without O/H◦ Small # of generations is a better choice◦ Larger # of generations more severe coupon collection problem
Single Generation
Single Generation
5/10/50 Generations5/10/50 Generations
Overhead
Solution: divide a file into small generations◦ Problem: too many generation causes a coupon
collection problem◦ Conflicting goals: maximizing benefits of NC vs.
minimizing coding O/H
Mitigating Coding Overheads
50MB
Solution: divide a file into small generations◦ Problem: too many generation causes a coupon
collection problem◦ Conflicting goals: maximizing benefits of NC vs.
minimizing coding O/H
Mitigating Coding Overheads
1 4
10MB x 5
What is optimal strategy for generation downloading?
Checking neighbor rank improve chances of linearly independent block, but◦ Low-rank cars can also have valuable blocks
Back to the BitTorrent problem of piece/generation selection◦ Local status based decision (i.e., the least/the most
downloaded generation, sequential order)?◦ Neighbor status based decision?◦ Random?
Gen1 Gen2 Gen3Local:(my status)
Global:(neighbor status)
Gen1 Gen2 Gen3
Request to ??
Generation Selection Generation Selection StrategiesStrategies
Virtual “Global” Completeness Vector
Lo
cal Min
Gen
2
Lo
cal Max
Gen
1
Neig
hb
or M
in
Gen
4
Neig
hb
or M
ax
Gen
3Global Min: Gen 4Global Max: Gen 3Random: RandomSequential: Gen 1
Simulation Setup Communications
◦ 802.11b; 11Mbps + Two-ray ground propagation Mobility
◦ Random Waypoint model w/ speed range of [0,20] m/s◦ Westwood area map: 2400m*2400m
Nodes◦ 3 APs: file sources◦ 200 nodes/40% interest level:
80 nodes are downloading a file Download parameters
◦ 50 megabyte file◦ 10 generations
Westwood area mapWestwood area map
Downloading all generations in Downloading all generations in parallelparallelGeneration ProgressGeneration Progress
Global MinGlobal Min Local MinLocal Min
RandomRandomNeighbor Min
Neighbor Min
Downloading all generations in Downloading all generations in parallelparallelOverall ProgressOverall Progress
Neighbor-aware strategy improves at the beginning of
downloading
* confidence interval is calculated with probability 95% using 8 simulations
* confidence interval is calculated with probability 95% using 8 simulations
Local-aware and random strategies
has smaller tail
Conclusions: Network-aware strategy has long tail of finishing times Local and random strategies behave almost as good as global status-aware
Downloading all generations in Downloading all generations in parallelparallelFinishing times histogramFinishing times histogram
Downloading generations Downloading generations (semi-)sequentially(semi-)sequentiallyGeneration ProgressGeneration Progress
Global Max
Global Max
Neighbor Max
Neighbor Max
Local MaxLocal Max
SequentialSequential
Downloading generations Downloading generations (semi-)(semi-)sequentiallysequentiallyOverall ProgressOverall Progress
* confidence interval is calculated with probability 95% using 8 simulations
* confidence interval is calculated with probability 95% using 8 simulations
!!! Neighbor-aware strategy
outperforms local and global one
Conclusions: Network-aware strategy outperforms other strategies Average finishing time for global/local max strategies 1.5 times worse
than neighborhood status aware policy
Downloading generations Downloading generations (semi-)sequentially(semi-)sequentiallyFinishing times histogramFinishing times histogram
Parallel vs Sequential DownloadingParallel vs Sequential DownloadingOverall progress of the best Overall progress of the best strategiesstrategies
Conclusions: Neighbor-aware generation choosing considerably improves chances for
helpful block (linearly independent) at the beginning Local or random strategy improves download finishing time
Parallel vs Sequential DownloadingParallel vs Sequential DownloadingFinishing times histogramFinishing times histogram
Conclusions: Neighbor-aware strategies have on average 20% worse finishing times
than local max strategy
Checking generation rank of the available generation greatly improves performance for neighbor status aware strategies
Integer vector gossiping decrease overall download performance
Interesting FactsInteresting FactsRank
checkingRank
checking
No rank check
No rank check
BoolvectorBool
vector
Intvector
Intvector
Conclusion Generation selection strategy in multi-generation
CodeTorrent downloads have big impact on the overall download performance
Local status aware strategies (local-min, random) have the best finishing performance
Neighbor status aware strategies have the best start-up performance
It is important to check rank for neighbor status aware strategies
Future work◦ Investigate performance of combined strategies◦ Check performance using different node mobility models