Lewis & Clark College Portlan d, OR Slide 1 of 18 Request Algorithms in Freenet-style Peer-to-Peer Systems Jeff Lesh Lewis & Clark College Portland, OR USA Dr. Jens Mache, Jason Guchereau, Melanie Gilbert, Felix Ramli, Matt Willkensin
Jun 26, 2015
Lewis & Clark College Portland, OR
Slide 1 of 18
Request Algorithms in Freenet-style Peer-to-Peer Systems
Jeff Lesh
Lewis & Clark College
Portland, OR USADr. Jens Mache, Jason Guchereau, Melanie Gilbert,
Felix Ramli, Matt Willkensin
Lewis & Clark College Portland, OR
Slide 2 of 18
Overview
Introduction to Freenet Performance issues
Modified Request Algorithms Results
Conclusions
Lewis & Clark College Portland, OR
Slide 3 of 18
What is Freenet?
completely decentralized forwards requests to one neighbor dynamically changes the overlay network and
replicates files on demand. Adapts to network usage Improves when used
Global storage space
Lewis & Clark College Portland, OR
Slide 4 of 18
=Data Request =Data Return
Node AReferences Data Store225 - B
932 -M
940-R
925
930
931
Node DReferences Data Store100 - G
101 - H
113 - J
099
111
113
Node BReferences Data Store025 - X
115 - C
230 - F
224
225
233
Node CReferences Data Store111 - D
112 - E
119 - L
115
118
120
111 - D
225 - B
932 -M
940-R
025 - X
115 - C
230 - F
111 - D
112 - E
119 - L
099
111
113
025 - X
111 - D
115 - C
230 - F
225 - B
932 -M
940-R
Request AlgorithmRequest Algorithm
•These changes in the overlay network improves the performance of subsequent queries
•Some caching as well.
Lewis & Clark College Portland, OR
Slide 5 of 18
Freenet’s design goals
Privacy Privacy for information producers, holders, and
consumers; Resistance to information censorship;
Performance High availability and reliability through decentralization;
and Efficient, scalable, and adaptive storage and routing.
Lewis & Clark College Portland, OR
Slide 6 of 18
This talk focuses on efficiency.
Our measure of efficiency is pathlength. Important for:
Users - delay Chance of slow links
Network – bandwidth usage
Lewis & Clark College Portland, OR
Slide 7 of 18
Our Questions
How good are the existing algorithms? “As the network is used, pathlength decreases”
[Hong’01]. But, this assumes 50% inserts and 50% requests.
Can we improve it?
Lewis & Clark College Portland, OR
Slide 8 of 18
Figure 2 50:50 vs. 99:1
Lewis & Clark College Portland, OR
Slide 9 of 18
Why does this happen?
Insert and request algorithms creates specialized references. Insert does more.
Requests don’t change the overlay network if they fail (and they do a lot at first).
We want to change these weaknesses.
Lewis & Clark College Portland, OR
Slide 10 of 18
Small-world effect
The overlay network created by the routing tables is an example of a small-world network.
Specialization as well as references to nodes that have far different specialization is key.
Real-world example. Also common in:
Social networks - film actor collaboration Neural Networks - C. Elegans (worm) Power Grid [Duncan J. Watt et.al. Nature ‘98]
Lewis & Clark College Portland, OR
Slide 11 of 18
The Experiments
Aurora Simulator – written by Theodor Hong It simulates nodes interacting with other nodes Inserting and requesting existing documents and
recording the number of hops it takes to find them.
The data we present is an average from 100 simulation runs.
Lewis & Clark College Portland, OR
Slide 12 of 18
Modified Request Algorithms
Learning from failed requests. Announcing successful requests.
Breadth-random fashion Breadth-neighbor fashion
Lewis & Clark College Portland, OR
Slide 13 of 18
Failed requests
As is they don’t add any short-cuts. Our goal was to take advantage of the work
already done up to the expiration of the HTL. Short-cut from the requesting node to the last
node’s data item with key closest to requested one.
Lewis & Clark College Portland, OR
Slide 14 of 18
Example
Lewis & Clark College Portland, OR
Slide 15 of 18
Successful Requests
The main idea is to add extra reference pointing to the fulfiller for each remaining HTL.
Breadth-random Breadth-neighbor approach worked best.
Here is announces itself to its neighbors with key close to the one it fulfilled.
This promotes specialization.
Lewis & Clark College Portland, OR
Slide 16 of 18
Example - Successful Requests.
Reference Table155 Z202 B901 C912 H939 V
Who to announce to?
Data Store
900
Lewis & Clark College Portland, OR
Slide 17 of 18
Results
Lewis & Clark College Portland, OR
Slide 18 of 18
Conclusions
The original request algorithm is not very effective at changing the overlay network.
They way to improve the performance of Freenet is to enhance the small-world properties that it already has.
Our combined algorithm was able,in simulation, to lower the pathlength under 99 to 1 conditions from 225 to 25.
Lewis & Clark College Portland, OR
Slide 19 of 18
Backup Slides
Lewis & Clark College Portland, OR
Slide 20 of 18
Differences from distributed hash table. Data is placed deterministically
Items located within bounded number of hops, but Securing against attack, load balancing, and (for some)
exploiting network proximity becomes more difficult.
Open source We can modify, learn.
Lewis & Clark College Portland, OR
Slide 21 of 18
50,000 Actions (99:1)
Original median pathlength: 2.94 Recycled
100% : 6.25 25% : 2.88