LessLog A Logless File Replication Algorithm for Peer-to-Peer Distributed Systems

Preview:

DESCRIPTION

LessLog A Logless File Replication Algorithm for Peer-to-Peer Distributed Systems. Kuang-Li Huang Tai-Yi Huang Jerry C. Y. Chou Embedded Operating Systems (EOS) Lab http://eos.cs.nthu.edu.tw/ Department of Computer Science National Tsing Hua University, Taiwan. Outlines. Introduction - PowerPoint PPT Presentation

Citation preview

LessLogA Logless File Replication Algorithm for Peer-to-Peer

Distributed Systems

Kuang-Li HuangTai-Yi Huang

Jerry C. Y. Chou

Embedded Operating Systems (EOS) Labhttp://eos.cs.nthu.edu.tw/

Department of Computer ScienceNational Tsing Hua University, Taiwan

EOS Lab, National Tsing Hua University IPDPS 2004

Outlines

IntroductionSystem model

Basic model Advanced model Fault-tolerant model

Experimental resultsConclusions and future work

EOS Lab, National Tsing Hua University IPDPS 2004

Related Work Chord: A Scalable Peer-to-Peer Lookup Protocol for Internet

Applications. IEEE/ACM Transactions on Networking, 2003

Pastry: Scalable, Distributed Object Location and Routing for Large-Scale Peer-to-Peer Systems.  IFIP/ACM International Conference on Distributed Systems Platforms (Middleware), 2001

Tapestry: A Resilient Global-Scale Overlay for Service Deployment. IEEE Journal on Selected Areas in Communications, January 2004

SCAN: A Dynamic, Scalable, and Efficient Content Distribution Network. International Conference on Pervasive Computing, 2002

EOS Lab, National Tsing Hua University IPDPS 2004

LessLog Contributions

An efficient bitwise replication algorithm Independent with underlying lookup protocols Make no use of client-access logs

A complete set of file-access APIFault-tolerant featuresSelf-organized for dynamic system change

EOS Lab, National Tsing Hua University IPDPS 2004

Basic System Model

There are totally N live nodes, N = 2m.Each node is assigned a unique ID in [0, N-1]

Randomly assigned or any user-specified way Such ID is called the PID of the node Denoted by P(i)

A complete N-node binomial tree is built for each node Totally N different physical trees

EOS Lab, National Tsing Hua University IPDPS 2004

Virtual Tree We use one unique virtual tree to construct each of

N different physical trees

111111

111100

101011

010111

101000

000011

000000

010100

VID

EOS Lab, National Tsing Hua University IPDPS 2004

Physical Tree for Node 3

111

110 101 011

100 001

000

010

3

2 1 7

0 6

4

3=011

key=100

complement

110⊕100=010

101⊕100=001

5

EOS Lab, National Tsing Hua University IPDPS 2004

Physical Tree for Node 6

111111

110110 101101 011011

100100 001001

000000

010010

66

77 44 22

55 33 00

11

EOS Lab, National Tsing Hua University IPDPS 2004

Properties of Virtual Tree

A node has i children nodes if its leftmost i bits in VID are all 1’s

The node of VID i has more or the same offspring nodes than the node of VID j if i > j

Given the PID of the root node in a lookup tree, we can do PID VID mapping for each node in the lookup tree

EOS Lab, National Tsing Hua University IPDPS 2004

LessLog Tree Traversal

111111

110110 101101 011011

100100 001001

000000

010010

33

22 11 77

00 66 55

44

EOS Lab, National Tsing Hua University IPDPS 2004

LessLog INSERT / GET API

111111

110110 101101 011011

100100 001001

000000

010010

33

22 11 77

00 66 55

44ask Hashing(file)=3

connect to P(6)

EOS Lab, National Tsing Hua University IPDPS 2004

LessLog REPLICATE API

111111

110110 101101 011011

100100 001001

000000

010010

33

Each replication reduces 50% load!

EOS Lab, National Tsing Hua University IPDPS 2004

Advanced System ModelThere are totally N live nodes, N ≦ 2m.Each node is assigned a unique ID in [0, 2m-

1] Randomly assigned or any user-specified way Denoted by P(i) Live nodes and dead nodes

A complete 2m-node binomial tree is built for each node Totally 2m different physical trees

EOS Lab, National Tsing Hua University IPDPS 2004

The 1st live node with largest VID

Advanced LessLog INSERT / GET API

111111

110110 101101 011011

100100 001001

000000

010010

33

22 11 77

00 66 55

44 ask Hashing(file)=3

EOS Lab, National Tsing Hua University IPDPS 2004

Overloaded Node with Largest VID

111

110 101 011

100 001

000

010

3

2 1 7

0 6 5

4

100

011

010

001Sorted by VID

EOS Lab, National Tsing Hua University IPDPS 2004

Self-organized LessLog

Joining (Leaving) a node Check whether it is the largest VID Ex. P(6) join, P(1), P(4) are dead nodes

1

111

110 101 011

100 001010

6

7 4 2

5 3 0

100

7111

110 011

010

6 3

2

1015

1004

0011

0000

EOS Lab, National Tsing Hua University IPDPS 2004

Fault-Tolerant LessLogFiles can always be accessedLessLog stores each file in 2b nodes, b ≦ m2b identical subtrees

111

110

101

011

100

001

000

010

3

2 1 7

0 6 5

4

1

1

1

1

0

0 0

0

EOS Lab, National Tsing Hua University IPDPS 2004

22 identical subtrees

1110

1100

1010

0110

1000

0010

0000

0100

2

0 6 10

4 8 14

12

1111

1101

1011

0111

1001

0011

0001

0101

3

1 7 11

5 9 15

13

EOS Lab, National Tsing Hua University IPDPS 2004

0

100

200

300

400

500

600

700

1 3 5 7 9 11 13 15 17 19 21

log-based

LessLog

random

Experimental Results Compared with log-based and random algorithms

at an evenly-distributed model

incoming requests/1000

rep

lica

s

EOS Lab, National Tsing Hua University IPDPS 2004

0

100

200

300

400

500

600

700

1 3 5 7 9 11 13 15 17 19 21

log-basedLessLograndom

Experimental Results Compared with log-based and random algorithms

at an locality model

incoming requests/1000

rep

lica

s

EOS Lab, National Tsing Hua University IPDPS 2004

Conclusions and Future Work

An efficient bitwise replication algorithm Built on top of any existing lookup protocol

Make no use of client-access logs Significant performance improvement

Fault-tolerant & self-organized futures Concurrent joins / leaves / failures

Under implementation at PlanetNet systems Real performance numbers

EOS Lab, National Tsing Hua University IPDPS 2004

Thank You

Recommended