Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · –...

37
Navigation in Decentralized Online Social Networks Sarunas Girdzijauskas Docent Seminar 07 November 2016

Transcript of Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · –...

Page 1: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Navigation in Decentralized Online Social Networks

Sarunas GirdzijauskasDocent Seminar

07 November 2016

Page 2: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Shortly About me

• PhD from EPFL, Switzerland 2009

• Research Intern at IBM Haifa 2008

• Post-doc at SICS 2009

• Senior researcher at SICS 2011

• Assistant Professor at LCN/EES 2012

• Associate Professor at SCS/ICT 2016

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 2

Page 3: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Research Topics

• Networks and Graphs

– Graph algorithms for Cloud Computing, Distributed and

Decentralized Systems, Social Networks, Privacy preservation.

• Research Areas

– Partitioning and community detection (Fatemeh)

– Storing Social Network (linked) data in Cloud Environments

(Anis)

– Streaming graph partitioning (Anis)

– Anomaly detection in Social graphs (Amira)

– Gossip learning for decentralized Online Social Networks

(Amira)

– NLP using Graph Partitioning (Kambiz)

– VM consolidation in clouds using Gossip (w. University of Umeå)

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

• This talk: on Navigation in DOSNs

– Decentralized Online Social Networks

– Community Cloud

3

Page 4: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

DOSNs: Motivation Slide

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 4

Page 5: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Why DOSNs?

• Challenging environments: Decentralized, highly heterogeneous(both resources and demand)

• Promises for: Scalability, availability, robustness, efficiency…• E.g,. DIASPORA, PeerSoN, Safebook, Vis-à-Vis, NetTube, DECENT,

PrisM, GoDisco, SocialTube

Give back the control to users:Decentralized Online Social Networks

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

Current Centralized Online Social Networks:

Data Collection

Data Mining

5

Page 6: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

DOSN basic services

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

• Data Dissemination/Search & Storage

• More advanced services: global aggregation/learning, analytics,

recommendation etc.

• How do we arrange decentralized nodes together (i.e., design the

topology of such decentralized system) where the above services

perform the best.

• I.e., Minimize traffic load/relaying and latency.

• How can decentralized search even work?

• Milgram’s Experiment

6

Page 7: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Milgram’s Experiment

• Not the one on obedience and authority!

• Milgram’s Experiment on the “topology of our Social

Networks”

– No Online Social Networks in 1960s

– 296 random people in USA forwarding a letter to a “target”

person in Boston.

– Personal info on the “target” (including address and

occupation)

– Forward only to a person known by first-name basis.

– Result?

– 64 chains succeeded.

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 7

Page 8: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Avg. number of steps?

– We live in a Small-World: average length of the

chains that were completed lied between 5 and 6

steps;

– Coined as “Six degrees of separation” principle.

– Similar results have been found in many other

social networks

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 8

Page 9: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Topology of Social Graph

• How to interpret such network?

• Maybe it is a Random Graph?

– low diameter!

– But very little clusteristaion i.e., very few triangles (common

friendships)

• Watts-Stogatz Small World Model

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 9

High clusteristaion and Short diameter!High clusteristaion but large diameter!

Weak links

Strong Links

Page 10: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Milgram’s Small-World Experiment (cont.)

• In 2011 Facebook shrunk 'Six Degrees Of Separation‘ to just 4.74 (721m users, 69b friendships)

– Twitter’s 5,91 of 12,8M friendships

• Does the low diameter of the SN graph answer our question?

– Surprise is that we can find these paths in a

decentralized manner, i.e., navigate

successfully with no “map”, no central authority,

no “Big Brother”.

• Why it Works?

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 10

Page 11: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

So why Milgram’s experiment worked?

• Social network is not a bare graph of vertices and edges – Nodes come with certain implicit

“labels” representing various dimensions of our life

– Hobbies, work, geographical distribution etc.

• There is (are multiple) “concept” space(s)– E.g., Geographical, Occupational,

Hobby etc.

– Each with a explicit or implicit distance metric!!!

– We can greedily minimize the distance!!

– Decentralized search: a greedydistance minimizing routingalgorithm

John,Stockholm;Neighbor;Musician;Likes photography;Etc.

Peter,Stockholm;Colleague;Computer scientist;Loves movies;Etc.

Simon,Paris;Friend;Stamp collector;Loves climbing;Etc.

Bill

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 11

Page 12: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Fundamental Navigability Rules

• Kleinberg’s Navigable Small-World model

– Very rough insight (”rule of thumb”):

• Connect to nodes that are inversely proportional to the

distance from you.

– With O(log(N)) links we can navigate in O(log(N)) hops

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

A4

A3

A2

A1

12

Weak links

Strong Links

Page 13: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Traditional DHTs and Kleinberg model

• P2P networks, DHTs

• Kleinberg’s model

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 13

Page 14: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

How do we build Structured P2P (DHT)? (recap)

• Navigable Overlay is

a graph, ”cleverly”

embedded in the ID

space, where efficient

routing is possible.

• The resulting

topology is a fixed

degree small-world

graph with high

clusterization and low

diameter.

0

0.5

0.7

5

0.2

5

ID SpaceEfficient Routing

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 14

Page 15: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Choice of Topology for DOSNs?

• Structured P2P networks (DHTs)

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 15

Page 16: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Event Dissemination in DOSNs

Social Network (F2F Network)

P2P Structured Overlay (DHT)

Message/Event Routing (Dissemination Structures):

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 16

Page 17: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Navigable Overlays (DHTs) as the backbone for DOSNs

• Navigable (e.g., . Search in O(logN)

number of hops) and has very low

degree (O(logN)) graph

• Each node gets uniform random

IDs (e.g., on a unit ring)

• Connect by some predefined rules

to k other nodes based on their IDs

• As discussed in previous slides

• E.g, Chord, Pastry, Symphony etc.

Greedy routing O(logN) hops

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 17

Page 18: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Event Publishing on Navigable Overlays

• Building efficient data

dissemination structures

• Creating a dissemination tree

with a fanout max of DHT

degree O(logN), and a depth

max of expected search cost

O(logN)

• Scalable

• Robust

• Bounded delay

• Message overhead and relays!

• Privacy!

• Services on top! (e.g,. Storage)

Publisher

Social Friends of the publisher

Greedy routing

Relay Nodes (overhed)

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 18

Page 19: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Why is DHT so inefficient in messages and dissemination cost?

• We build a very efficient (navigable) overlay that is based

ONLY on node IDs and is completely oblivious to Friend-to-

Friend (F2F) network and node localities

• Friends that are close in “social graph” are uniformly

distributed in the DHT ring – scrambled forever

• Any group communication induced by social activities

on expectation will not be local and will induce O(log N)

communication cost (likely non-interested relays).

• In effect we end with:

• No locality,

• (almost) no direct friend-to-friend communication

• Dissemination structures (trees) that we build will have

on expectation around O(logN) relay nodes for every

friendship

• i.e., vast majority of relays for each “newsfeed” or other action.

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 19

Page 20: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Choice of Topology for DOSNs?

• Structured P2P networks (DHTs)

• Unstructured P2P

– Friend-to-Friend based networks

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 20

Page 21: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Overlay Design for DOSNs (2nd attempt)

• Let’s connect them all: DOSN overlay that mirrors

“social network”, i.e., build Friend-to-Friend overlay.

• Low latency, Low communication cost

• Still global search might be problematic…

• Most of social networks have relatively high avg.

degree and have power-law degree distributions.

• Issue of handling most popular nodes!

• E.g., tens of millions of links in Twitter and even

5000 friendship links in Facebook!

• E.g, no go for WebRTC

• Challenge: Keep the node degree fixed!

•ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 21

Page 22: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Dilemma…

• Using only Unstructured Overlays (F2F based):

• Too large node degrees

• Search/routing is not effective

• hard to build other services on top: i.e., storage,

recommendation, analytics

• Using Structured Navigable overlays (DHT based):

• locality is lost,

• high relay traffic.

• Is there a way out?

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 22

Page 23: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Locality Aware Structured P2Ps?

• Naïve solutions

• Use DHT solutions that provide some ”degrees of freedom”

while selecting neighbors (e.g., Pastry, Symphony etc)• Castro et al. ”Topology-aware routing in structured peer-to-peer overlay networks” 2003

• Antaris et al. “A Socio-Aware Decentralized Topology Construction Protocol” HotWeb2015

• Chen et al. “Design of Routing Protocols and Overlay Topologies for Topic-based Publish/Subscribe on

Small-World Networks” Middleware 2015

• Stick to the ”rules”, but choose only those peers that are

more ”useful”Log (N) exponentialy

decreasing in size partitions of ID space

Fair amount of choice in the largest partitions

• Take 1M network and

20 friends

• ~20 partitions

– 1st partition: ~10

friends

– 2nd partition : ~5

– 3rd partition : ~2.5

– 4th partition : ~1

– 5th to 20th :1 friend on

expectation

That’s whre it is important:

locality!Problem: Marginal improvements only!

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 23

Almost no choice here!

Page 24: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

F2F social graph vs. Navigable P2P Overlay

• Both networks are non-random, but ”small-world”

like.

• High clusterisation, low diameter.

• Differences:

• F2F: power-law degree distribution

• Navigable Overlay: fixed degree distribution

Navigable Overlay (DHT)Both are of

Small-World topology!

Social Network (F2F)

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 24

Page 25: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

• What if:

• 1) we take a subgraph of F2F with fixed degree that is

topologically similar to the graph induced by the Navigable

Overlay?

• 2) embed it into ID space so that the routing is efficient?

• i.e., ”cleverly assign IDs for each node”.

0

0.7

5

0.2

5

ID Space

Making F2F Network Navigable

Navigable Overlay (DHT)

Social Network (F2F)

Subgraph of F2F

Similar Topological Properties

Efficient Routing

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

Navigable Overlay constructed out of

F2F links

25

Page 26: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

0

0.7

5

0.2

5

How to assign IDs to F2F Network ?

• Option 1:

• Start with random ID assignment

• Expected poor initial routing performance

• Keep on exchanging IDs between pairs of nodes to “improve” the

routing.

• Option 2:

• Try to “infer” what are the topological “clusters” in the graph and

allocate similar IDs for the nodes in those clusters.

ID Space

Subgraph of F2F

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

Efficient routing

26

Page 27: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Option 1: random ID assignment

• Let’s start “easier”:

• Take existing Navigable Overlay

• “forget” ID space (remove IDs)

• Try to reassign IDs just by looking into the topology of the

graph.

• NP-hard problem – at least as hard as community detection,

0

0.7

5

0.2

5

ID Space

Subgraph of F2F

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

0

0.7

5

0.2

5

Reassigned ID Space

Can we still get efficient routing?

Navigable Overlay (DHT)

27

Page 28: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Proposed Solutions

Sandberg et al. “Distributed Routing in Small-World Networks”

ALENEX2006

• Each peer gets a random IDs

• Each peer periodically exchanges info of their IDs with a random

peer and decides whether to swap the IDs.

• Better than random, but largely fails to “discover” right ID

allocation.

• For larger graphs (100k nodes) up to 50% of queries failed to

reach the destination

• Reason: All links are of the same “importance”

– Too many “degrees of freedom”, too many dependencies: position

improvements in one pair “damages” many other positions.

• Our Solution:

• The links are not the same (remember strong and weak

links)?! Treat each link differently!

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

0

0.7

5

0.2

5

0

0.7

5

0.2

5

28

Page 29: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Weak and Strong Ties

• Each node orders all the neighbors by the “strength of their ties”

– Weak vs. Strong links

– E.g., counting friendship triangles, gossip to detect local communities etc.

• Graph pruning:

– Consider only top k strongest links (representing communities), ignore weakest

links.

– A graph with large diameter emerges,

• i.e, less “degrees of freedom” -> easier to converge to local optima

Removal of weakest links

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 29

Page 30: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Socially-Aware Distributed Hash Tables (Nasir et al, P2P2015)

• Each peer gets a random IDs

• Each peer periodically exchanges info of their IDs with a random peer and

decides whether to swap the IDs.

– Decision: based on a cost function that (locally) improves the positions of the

two nodes as compared to their neighbors

– Cost function: biases the preference towards the strong neighbors (prefer

to have strong links with IDs that are as close as possible in the ID space,

while disregarding the weak links)

– Gossip based, Integrated with Symphony Overlay.

– Reduces lookup latency by ~30%

• Can we do even better?

– Antaris et al. “SELECT: A Distributed Publish/Subscribe Notification System for Online

Social Networks” (collaboration with University of Cyprus)

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 30

Page 31: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Option 2: ID assignment on the fly

• F2F graph: nodes arrive with associated edges.

• Bootstrapping: Only first nodes get random IDs.

• Subsequent arriving nodes identify the

“strongest” existing friendship communities

and “join” them by selecting ID centered in

these communities.

• Arriving node assumes ID as a centroid between k

strongest links (e.g., nodes that share most of the

friends)

– ID ranges identify communities

• We have to cap the degree (take a subgraph)

• Following Kleinbergian rule: identify most

dissimilar communities/regions to point to

• +Bias toward more reliable nodes

• +Bias toward particular workloads in F2F graph.

Few extra links: to maintain the ring, and fill up

the finger table for nodes with low social degree

(people with few friends)

0

0.7

5

0.2

5

Social Network (F2F)

Subgraph of F2F Efficient

Routing

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

31

Page 32: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Option 2: ID assignment on the fly (cont.)

• Which community should

the newcomer join?

– Which ID should it take?

• Join between “strongest”

existing friendships in a

particular community

– e.g., id - a centroid of k-

strongest links (nodes that

shares most of the friends)

– Communities and Strength

discovered by gossiping

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

Community 1

Community 2

Community 3

0.05 0.090.12

0.18

0.810.790.750.670.63

0.530.5 0.48

0.430.4

0.35

?

32

Page 33: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Capping the degree

• Usually a node has way more social friends than ”connection quota”.– E.g., max 20 out of 500 friends.

• Which ones to keep?– Only the closest friends?

• We lose long range links

– Random friends?• Similar to Watts&Strogats model:

Small-World but not navigable...

– Ideas from Kleinberg?• Can not apply directly since we do not

want to create new links and ID space is not uniform!

– Serch performance should be biased toward Friends nodes

– Detect k friends, representing k-most dissimilar regions (Kleinbergian partitions, e.g., using gossiping or LSH technique) and establish connections to them

• Few extra links if necessary

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

Community 1

Community 2

Community 3

0.05 0.090.12

0.18

0.810.790.750.670.63

Which ones to keep?

33

0.530.5

0.480.43

0.40.35

0.49

Page 34: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

SELECT: A Distributed Publish/Subscribe

Notification System for Online Social Networks

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

Google+ (110k nodes, 12M edges, avg degree 157)

Twitter (5M nodes, 800M edges,avg degree 127)

Average size of dissemination tree (NewsFeed)

DHT (Symphony) 468 869

SELECT 169 (↓64%) 143 (↓84%)

Average Relay nodes

DHT (Symphony) 311 769

SELECT 12 (↓96%) 16 (↓98%)

More shallow dissemination trees (up to

3x reduction on Delay)

Almost no relay nodes (Improvement on message

overhead)

34

• Some Results:

Page 35: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Take Aways

• The World is not only Small (6-degrees of separation),

but also navigable in a completely decentralizedfashion.

• Community aware assignment of IDs and selection of

links enables efficient navigation in F2F networks

• in turn improving privacy/security, getting rid of majority of

relay traffic and allowing locality aware services

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 35

Page 36: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

• Overlay navigation, dissemination and pub/sub

1. Muhammad Anis Uddin Nasir, Nicolas Kourtellis, Sarunas Girdzijauskas: Highly

Scalable, Heterogeneous and Reliable Overlays for Decentralized Online Social

Networks. P2P 2015

2. Mansour Khelghatdoust, Sarunas Girdzijauskas: Short: Gossip-Based Sampling in

Social Overlays. NETYS 2014: 335-340

3. Fatemeh Rahimian, Thinh Le Nguyen Huu, Sarunas Girdzijauskas: Locality-

Awareness in a Peer-to-Peer Publish/Subscribe Network. IFIP International

Conference on Distributed Applications and Interoperable Systems, DAIS 2012: 45-

58

4. Fatemeh Rahimian, Sarunas Girdzijauskas, Amir H. Payberah, Seif Haridi: Vitis: A

Gossip-based Hybrid Overlay for Internet-scale Publish/Subscribe Enabling

Rendezvous Routing in Unstructured Overlay Networks. IPDPS 2011: 746-757

5. Fatemeh Rahimian, Sarunas Girdzijauskas, Amir H. Payberah, Seif Haridi:

Subscription Awareness Meets Rendezvous Routing, The Fourth International

Conference on Advances in P2P Systems, AP2PS 2012.

6. Sarunas Girdzijauskas, Gregory Chockler, Ymir Vigfusson, Yoav Tock, Roie

Melamed: Magnet: practical subscription clustering for Internet-scale

publish/subscribe. DEBS 2010: 172-183

7. Sarunas Girdzijauskas, Gregory Chockler, Roie Melamed, Yoav Tock: Gravity: An

Interest-Aware Publish/Subscribe System Based on Structured Overlays, [fast

abstract] DEBS 2008.

8. Fabius Klemm, Sarunas Girdzijauskas, Jean-Yves Le Boudec, Karl Aberer: On

Routing in Distributed Hash Tables. Peer-to-Peer Computing 2007: 113-122

9. Anwitaman Datta, Sarunas Girdzijauskas, Karl Aberer: On de Bruijn Routing in

Distributed Hash Tables: There and Back Again. Peer-to-Peer Computing 2004: 159-

166

10. Sarunas Girdzijauskas, Wojciech Galuba, Vasilios Darlagiannis, Anwitaman Datta,

Karl Aberer: Fuzzynet: Ringless routing in a ring-like structured overlay. Peer-to-Peer

Networking and Applications 4(3): 259-273 (2011)

Bibliography

36ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH

• Overlay Construction

1. Sarunas Girdzijauskas, Anwitaman Datta, Karl Aberer: Oscar: A Data-

Oriented Overlay For Heterogeneous Environments. ICDE 2007: 1365-1367

2. Sarunas Girdzijauskas, Anwitaman Datta, Karl Aberer: Oscar: Small-World

Overlay for Realistic Key Distributions. DBISP2P 2006: 247-258

3. Sarunas Girdzijauskas, Anwitaman Datta, Karl Aberer: On Small World

Graphs in Non-uniformly Distributed Key Spaces. ICDE Workshops 2005:

1187

4. Karl Aberer, Luc Onana Alima, Ali Ghodsi, Sarunas Girdzijauskas, Seif Haridi,

Manfred Hauswirth: The Essence of P2P: A Reference Architecture for

Overlay Networks. Peer-to-Peer Computing 2005: 11-20

5. Sarunas Girdzijauskas, Anwitaman Datta, Karl Aberer: Structured overlay for

heterogeneous environments: Design and evaluation of oscar. TAAS 5(1)

(2010)

Page 37: Navigation in Decentralized Online Social Networkssarunasg/Documents/NavigationDOSNs.pdf · – Gossip learning for decentralized Online Social Networks (Amira) ... Sandberg et al.

Thank you!

ŠARŪNAS GIRDZIJAUSKAS - 2016 NOV 07, KTH 37