Ditto - A System for Opportunistic Caching in Multi-hop Mesh Networks

DITTO - A SYSTEM FOR OPPORTUNISTIC CACHING IN MULTI-HOP MESH NETWORKS

CONTENTS

• Introduction to WMN• Throughput issues in WMN• Related Work• Challenges and Opportunities• Ditto Design

– Chunk Based Transfers– Ditto Proxy– Sniffer

• Evaluation• Limitations

2

WIRELESS MESH NETWORKS (WMNS)

3

Cost Effective,Last mile technology

Greater Coverage Robust , Self Healing

A communications network made of radio nodes in a mesh topology

Router

Router

Router

RouterGateway

Infrastructure / Backbone WMN

MOTIVATION :THROUGHPUT PROBLEM IN WMNS

4

• Interference

• Congestion at hotspots (GW)

P1P3

THROUGHPUT PROBLEMS (CONTD..)

Increase in number of hops -> increase in probability of loss due to interference

Has an adverse effect on the throughput at Transport layer

5

HIERARCHICAL CACHING [WEB CACHING TECHNIQUE]

6

Disadvantages 1) Application specific 2) Need to cache the whole files.

Bottom(Local Cache)

Institutional Cache

Regional Cache

National Cache

Server

Takes advantage of the similarity and locality of workload.

RTS-ID

7

00

0

0

00RTS

for packet id 0

CTSI already have it

1) Packet Level Caching 2) Cache overheard packets

Disadvantages3) Too Granular, high overhead4) Not scalable for large networks.

EXPLOITING LOCALITY THROUGH CACHING

8On-Path + Opportunistic Caching -> Ditto

P1

P2

P3

Path of the transfer: Alice, P1, P3, GW

P1 and P3 perform on-path caching

P2 can perform opportunistic caching

CHALLENGES IN WIRELESS NETWORKS FOR OPPORTUNISTIC CACHING

• High loss rates due to interference / collision.– MAC layer takes care of this by re-transmissions.

• There could be cases where overhearing node cannot hear properly while the recipient node does.

9

111

1

22 2

2 ObservationMulti-hop transmission creates more opportunities for over-hearing thus mitigating the effect of lossy transmissions in wireless networks.

DITTO : KEY POINTS

Based on Opportunistic caching and exploits the broadcast nature of wireless networks.

“Chunk based” caching Application independent. Exploits re-transmissions for advantage.

10

FILE TRANSFER MECHANISM IN DITTO Ditto uses “Data Oriented Transfer” Factors needed to be considered while devising an

efficient file transfer mechanism: 1) Packet level caching is too granular. Incurs a lot of overhead for re-assembly . 2) File-level caching is also not efficient as overhearing a complete file is less probable.Solution: Divide file into smaller units called “Chunks”[8 – 32 kb]

Use chunks as data transfer units. Use the hash value of contents to identify the chunk. 11

ADVANTAGES OF DOT

Fairly good probability of overhearing, easy to Re-assemble

Application independent : uses Rabin fingerprinting to mark boundary regions for splitting files into chunks. Increases the probability of finding identical chunks in similar files.

Self verifying mechanism as chunks are named using the hash value of their contents.

12

DITTO DESIGN

Components:o Ditto Proxies

• Cache and serve data to their previous hop proxies.

• Next – Hop towards Gateway• Per-Hop TCP connection.

Sniffers• Over hear data• Reconstruct overheard chunks• Pass the reconstructed chunks to proxies

13

DITTO DESIGN

14

DITTO TRANSFER - ILLUSTRATION

Step 1

Step 2 Receiver

Sender

15

File.txt

Chunk 1

Chunk 2

Chunk 3

Chunk 4

Hash Function

Chunk ID2

Chunk ID1

Chunk ID3

Chunk ID4

Application Application

DOT Service DOT ServiceTransport layer Transport layer

File Request

ChunkIds:1,2,3,4

Chunk request

Chunks

DOT Service

Sniffer

SNIFFER

Overhears the ongoing transmission Identifies the TCP transmission using the IP

and Port pairs(Src IP , Src Port , Dest IP , Dest Port)

Uses sequence number for re-assembly.Inter Stream Re-Assembly TCP STREAM A

Chunk Header

TCP STREAM B

Thus Exploits multiple hearing oppurtunities

16

1 6542

1 32 5 6

1 2 3 4 5 6

X

EMULAB WIRELESS TESTBED

17

Gateway

EXPERIMENTAL SETUP

Wireless Network used : 802.11b File Sizes : 1- 5 MB Chunk Sizes – 8 KB, 16 KB, 32 KB

18

RECONSTRUCTION EFFICIENCY

19

Around 50% observers are able to reconstruct at least 50% chunks

THROUGHPUT EVALUATION

Leaf Nodes request the same file from the gateway e.g: software update on all nodes

Different request patterns: Sequential, staggered Random order of receivers

Schemes Ditto’s comparison with On-Path and E2E

20

THROUGHPUT IMPROVEMENT IN DITTO

21

Median = 540 Kbps

Campus Testbed

Median = 1380 Kbps

Median = 540 Kbps

Median = 5370 Kbps

Median = 1380 Kbps

Median = 540 Kbps

RESULTS (CONTD)

Effect Of Chunk Size - As noted earlier , smaller chunk size

implies better reconstruction efficiency. - Results show that with 8Kb chunk size 20

% of observers were able to reconstruct 80% of chunks. With 32Kb chunk size , this

reduces to 60 %. Effect of Inter Stream Re-assembly 10 % improvement over without inter-

stream re- assembly.

22

EVALUATION RESULTS / CONCLUSION

Hotspot formation near gateways avoided as nodes near gateway have high overhearing effectiveness.

A choice 8-32 KB of chunk size gives a good reconstruction efficiency.

Significant throughput improvement – up to 7X compared to On Path and 10X compared to no caching

At points where-ever inter stream assembly by overhearing was possible there was an improvement of 10% in reconstruction efficiency 23

LIMITATIONS OF THE APPROACH Ditto requires that a TCP connection between each

proxy , if there are multiple Transport layer timeouts it would severely affect the performance.(It is quite possible in a wireless environment).

While fetching a file , a lookup is made at each proxy which results in considerable delay if there are many proxies on the path [especially if the file chunks are not present in any of the proxies].

The method is not energy efficient , overhearing results in a lot of energy wastage.

If Encapsulated packets are used as in secure tunnels, it would be complex to perform inter stream re-assembly (as all packets are addressed to the same destination – which is the secure gateway).

24

Thank You

Q?25

EXTRA SLIDES: SNIFFER IMPLEMENTATION

26

SNIFFER IMPLEMENTAITON (CONTINUED)

27

Chunk Table

SNIFFER IMPLEMENTATION (CONTD)

Flow Chunk Map

28