Efficiency of Tree-structured Peer-to-peerService Discovery Systems

Cédric Tedeschi, Eddy Caron, Frédéric Desprez

University of Lyon, FranceLIP laboratory. UMR CNRS-ENS Lyon-UCB Lyon-INRIA 5668

Hot-P2P, April 18, 2008

DLPT overview Protocol Simulation Conclusion

Initial context

Service discovery in grid computingService (binary file, library) installed on serversServers declare their services, client discovers themNeed for maintaining this information

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 2

DLPT overview Protocol Simulation Conclusion

Initial context

Service discovery in grid computingService (binary file, library) installed on serversServers declare their services, client discovers themNeed for maintaining this information

Target platformslarge scaleno central infrastructuredynamic joins and leaves of nodes

P2P systemsPurely decentralized algorithmsScalable algorithms to retrieve objectsFault-tolerance

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 2

DLPT overview Protocol Simulation Conclusion

Trie-based overlays

A promising way to store and retrieve servicesAdvantages

Efficient range queriesAutomatic completion of partial stringsEasy extension to multi-attribute queries

Existing approachesSkip Graphs (Aspnes and Shah – 2003)P-Grid (Datta, Hauswirth, John, Schmidt, Aberer – 2003)PHT (Ramabhadran, Ratnasamy, Hellerstein, Shenker – 2004)DLPT (Caron, Desprez, Tedeschi – 2005)Nodewiz (Basu, Banerjee, Sharma, Lee – 2005)

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 3

DLPT overview Protocol Simulation Conclusion

Trie-based overlays

A promising way to store and retrieve servicesAdvantages

Efficient range queriesAutomatic completion of partial stringsEasy extension to multi-attribute queries

Existing approachesSkip Graphs (Aspnes and Shah – 2003)P-Grid (Datta, Hauswirth, John, Schmidt, Aberer – 2003)PHT (Ramabhadran, Ratnasamy, Hellerstein, Shenker – 2004)DLPT (Caron, Desprez, Tedeschi – 2005)Nodewiz (Basu, Banerjee, Sharma, Lee – 2005)

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 3

DLPT overview Protocol Simulation Conclusion

DLPT : A trie-based indexing system

Logical structure

Greatest Common Prefix Tree

Dynamically constructed

Bounded degree and height

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 4

DLPT overview Protocol Simulation Conclusion

DLPT : A trie-based indexing system

Logical structure

Greatest Common Prefix Tree

Dynamically constructed

Bounded degree and height

Lookup

Exact match

Autocompetion

Range queries

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 4

DLPT overview Protocol Simulation Conclusion

DLPT : A trie-based indexing system

Logical structure

Greatest Common Prefix Tree

Dynamically constructed

Bounded degree and height

Lookup

Exact match

Autocompetion

Range queries

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 4

DLPT overview Protocol Simulation Conclusion

DLPT : Mapping of the system onto the network

PrinciplesAn underlying DHTEach physical node (Peer) runsone or more logical process

DrawbacksNeed for a DHT

Important global costRandomly achieved

Load unbalanceDepth of nodes in the treeHeterogeneous popularity ofservices, hot spots

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 5

DLPT overview Protocol Simulation Conclusion

Objectives

1 Avoid the need for a DHT to reduce the cost2 Inject some load balancing

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 6

DLPT overview Protocol Simulation Conclusion

Build a ring over the physical network

Build a Chord-like ring over peersto distribute the tree nodes

Use the tree links to maintain thephysical network

Complexity : Trie complexities + 2

Peer insertion algorithm

1 Joining peer P contacts arandom tree node

2 Route the request to the treenode N s.t. IDN has the hi-ghest id lower than IDP .

3 The succeqssor of P is eitherPN or succ(PN)

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 7

DLPT overview Protocol Simulation Conclusion

Load balancing heuristics - related work

Karger and Ruhl, 2001periodic random item balancinghomogeneity of peer capacities

Godfrey et al., 2003periodic item redistributionsemi-centralized

Ledlie and Seltzer, 2005Based on the k-choices principleheterogeneity of peer capacities and data popularitychooses the best location for a joining peer among k

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 8

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputObjective function

At time unit τTwo adjacent peers S and P with capacity CS and CP

ντS and ντP the sets of nodes currently managed by S and P.LτS =

∑n∈ντ

Sln

LτP =∑

n∈ντP

lnT τ

S,P = min(LτS,CS) + min(LτP ,CP)

We want to maximize the throughput of time unit τ+1 basedon knowledge of time unit τFind ντ+1

S and ντ+1P that maximizes T τ+1

S,P .

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 9

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputObjective function

At time unit τTwo adjacent peers S and P with capacity CS and CP

ντS and ντP the sets of nodes currently managed by S and P.LτS =

∑n∈ντ

Sln

LτP =∑

n∈ντP

lnT τ

S,P = min(LτS,CS) + min(LτP ,CP)

We want to maximize the throughput of time unit τ+1 basedon knowledge of time unit τFind ντ+1

S and ντ+1P that maximizes T τ+1

S,P .

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 9

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputObjective function

At time unit τTwo adjacent peers S and P with capacity CS and CP

ντS and ντP the sets of nodes currently managed by S and P.LτS =

∑n∈ντ

Sln

LτP =∑

n∈ντP

lnT τ

S,P = min(LτS,CS) + min(LτP ,CP)

We want to maximize the throughput of time unit τ+1 basedon knowledge of time unit τFind ντ+1

S and ντ+1P that maximizes T τ+1

S,P .

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 9

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputObjective function

At time unit τTwo adjacent peers S and P with capacity CS and CP

ντS and ντP the sets of nodes currently managed by S and P.LτS =

∑n∈ντ

Sln

LτP =∑

n∈ντP

lnT τ

S,P = min(LτS,CS) + min(LτP ,CP)

We want to maximize the throughput of time unit τ+1 basedon knowledge of time unit τFind ντ+1

S and ντ+1P that maximizes T τ+1

S,P .

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 9

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputObjective function

At time unit τTwo adjacent peers S and P with capacity CS and CP

ντS and ντP the sets of nodes currently managed by S and P.LτS =

∑n∈ντ

Sln

LτP =∑

n∈ντP

lnT τ

S,P = min(LτS,CS) + min(LτP ,CP)

We want to maximize the throughput of time unit τ+1 basedon knowledge of time unit τFind ντ+1

S and ντ+1P that maximizes T τ+1

S,P .

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 9

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputAlgorithm

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 10

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputAlgorithm

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 10

DLPT overview Protocol Simulation Conclusion

A novel heuristic : Max Local ThroughputAlgorithm

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 10

DLPT overview Protocol Simulation Conclusion

Simulation results

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

Rat

e of

sat

isfie

d re

ques

ts

Time

Load balancing − stable network

Max local throughput [30 run]K−choices [30 run]

No LB [30 run]

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 11

DLPT overview Protocol Simulation Conclusion

Simulation results

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

Rat

e of

sat

isfie

d re

ques

ts

Time

Load balancing − stable network − high load

Max local throughput [30 run]K−choices [30 run]

No LB [30 run]

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 11

DLPT overview Protocol Simulation Conclusion

Simulation results

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

Rat

e of

req

uest

s sa

tisfie

d

Time

Load balancing − dynamic network − low load

Max local throughput [30 run]K−choices [30 run]

No LB [30 run]

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 11

DLPT overview Protocol Simulation Conclusion

Simulation results

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

Rat

e of

req

uest

s sa

tisfie

d

Time

Load balancing − dynamic network − high load

Max local throughput [30 run]K−choices [30 run]

No LB [30 run]

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 11

DLPT overview Protocol Simulation Conclusion

Simulation results

Load Stable network Dynamic networkMax local th. K-choices Max local th. K-choices

5% 39,62% 38,58% 18.25% 32,47%10% 103,41% 58,95% 46,16% 51,00%16% 147,07% 64,97% 65,90% 59,11%24% 165,25% 59,27% 71,26% 60,01%40% 206,90% 68,16% 97,71% 67,18%80% 230,51% 76,99% 90,59% 71,93%

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 11

DLPT overview Protocol Simulation Conclusion

Simulation results

0

0.2

0.4

0.6

0.8

1

0 20 40 60 80 100 120 140 160

Sat

isfie

d re

ques

ts

Time

Load balancing − dynamic network − dynamic load

Max Local Throughput [50 run]K−choices [50 run]

No LB [50 run]

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 11

DLPT overview Protocol Simulation Conclusion

Conclusion and future work

ContributionsA protocol to map a trie on a P2P network

Reduction of the architecture costBetter self-containment

Load BalancingA novel heuristicLocalizing maximizing the throughputGood performance compared to K-choices

Dynamic load balancing in triesOn-going work

Prototype developmentFirst results are promising

C. Tedeschi, E. Caron, F. Desprez Efficiency of Tree-structured P2P Service Discovery Systems 12