1 Maximizing Remote Work in Flooding-based P2P Systems Qixiang Sun Neil Daswani Hector Garcia-Molina Stanford University.

1

Maximizing Remote Work in Flooding-based P2P Systems

Qixiang Sun

Neil Daswani

Hector Garcia-Molina

Stanford University

2

P2P File Sharing

• Gnutella, KaZaA, etc.

Internet Internet

3

Architecture

• Super-node network with flooding-based search

Search Query

4

Problem

• Accept new queries from local clients• Handle remote queries from other super-nodes

Where is the balance?

?

5

Problem (2)

• Objective: Remote Work – process as many queries from other nodes as

possible.

Query

0 11

1

0 0

6

Problem (3)R

emot

e w

ork

done

Number of new queries injected

Where is the optimal?

7

Simple Model

Super-nodes operate in rounds

Capacity C

Accepts new queriesfrom local clients

Handles remote queries

8

When Overloaded

• Choose queries with the highest TTL first

• Ties can be broken randomly

Has a steady state and

is optimal in remote work

?

9

Example

• 3 super-nodes with TTL = 1

= ?A B

C

13

local

neighbor 1 neighbor 2

10

Example (2)

= ?13

= ?14

= ?12

• 6 super-nodes with TTL = 1

11

Solution

• 6 super-nodes with TTL = 1

2 4

12

Solution (2)

• 6 super-nodes with TTL = 1

2 4 24

3

3

{ 2, 2, 3, 3, 4, 4 }

2 4 7

= 13

13

Another Example

• 5 super-nodes with TTL = 2

3

4

4

5 5

3 + 4 = 7 > 5

{ 3, 4, 4, 5, 5 }

= 14

14

Intuition

. . . . . .

2 5 7 10

= 16

Unsaturated Saturated

15

Intuition (2)

. . . . . .

2 5 7 10

= 17

Unsaturated Saturated

Loss = Gain?

16

Different

• Each super-node could use a different More work done in the network!

Spare capacity

17

Example

• Star topology with TTL = 1

Identical = 0.5Remote work = 3.5 C

Different Remote work = 6 C

1

0 0

0

0 0

0

18

Penalty of using identical

. . . . . .D1 i nDD

Maximum remote work is at most n C

Pick = all nodes saturatedD1

1

penalty is D

1

1

D1

1 remote work = n C (1 - )

19

Penalty of using identical (2)

D

1

1• How big is ?

D1 TTL + 1

D1 50 penalty is less than 2%

In practice:

20

Solving for different

1

0 0

0

0 0

0

Similar to findingthe dominatingset for the graph

w1 w2

w3

w4

Minimize sum of all weights

21

Why?

. . . . . .

Unsaturated Saturated

Boost unsaturated nodes

22

Future Directions

• Nodes of different capacities

• Incremental algorithm for computing at each node

• An incentive mechanism so that each node will forward neighbors’ queries

23

Conclusion

• Controlling rate of query injection leads to better efficiency

• Solutions for finding the optimal rate

For other P2P related work, google for “Stanford Peers”