www.dvs1.informatik.tu-darmstadt.de bubblestorm: resilient, probabilistic, and exhaustive...
TRANSCRIPT
![Page 1: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/1.jpg)
www.dvs1.informatik.tu-darmstadt.de
BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search
Wesley W. Terpstra, Jussi Kangasharju, Christof Leng, Alejandro P. Buchmann
Databases and Distributed Systems Group
Technische Universität Darmstadt
Germany
![Page 2: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/2.jpg)
2
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Classification of P2P Search
…
Centralized
EDonkey2000
Chord
CAN
Kademlia
P-Grid
Tapestry Pastry
…
Napster
Structured Unstructured
Gnutella
Random Walks Gia
BubbleStorm
…
FastTrack
…
![Page 3: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/3.jpg)
4
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Why unstructured search?
![Page 4: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/4.jpg)
5
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Query processing is independent from routing
This simplifies application development:
Implementing a query language locally distributed implementation “for free”
Reuse existing libraries for query languages SQLite, XPath, Lucene, …
No need to invent a new algorithm per query language
Separation of Concerns
![Page 5: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/5.jpg)
6
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Expressive Searches are Easy
Any selective query can be supported
One operation in unstructured systems can perform a full-text search range restriction on file size hierarchical type selection
Structured systems break queries into small pieces e.g. DHTs must transform the query into key-value Cannot simply compare the cost of operations
![Page 6: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/6.jpg)
7
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Example: Full text search
One op. fetches documents May contact more peers
Perform multiple lookups Transfer word lists (not via DHT) Fetch documents
Keyword 1
Keyword 2
Keyword 3
Doc 1
Doc 2
Doc 2
Doc 1
Unstructured Overlay DHT with inverted index
Latency favours the unstructured approach Relative bandwidth requirements highly parameter dependent
![Page 7: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/7.jpg)
8
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Not everything is uniform
Natural load balancing
P2P applications often handle Zipfian loads Human text has =1 YouTube has =0.5
An unstructured request can be served by any peer Heterogeneity is accommodated by irregular degree In comparison, adding a keyspace creates hot
spots
![Page 8: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/8.jpg)
9
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Example: Zipfian Load
For natural languages (e.g. full text search) in a keyspace: Expected load on most the loaded peer is 7000x average The loaded peer probably has only average capacity
1000
10000
100
10
1
0.1 100000 10000
Node Rank
1000 100 101
Load
com
pare
d t
o load
avera
ge
alpha=1.0
alpha=0.5
alpha=0.0
100,000 peers1 million documents
![Page 9: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/9.jpg)
10
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
BubbleStorm Intuition
Replicate both queries and data copies each (hidden constants unequal)
Data and queries rendezvous in the network€
O n( )
![Page 10: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/10.jpg)
11
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
BubbleStorm: Random Replication
Place data replicas on random nodes Nodes evaluate query replicas on all stored data
Where both data and query go, matches are found
Collisions result from the birthday paradoxData
Query
![Page 11: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/11.jpg)
12
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
BubbleStorm: Exploiting Heterogeneity
Peers have different capacities Faster peers receive more traffic
This is beneficial! Contribution is squared
Data
Query
![Page 12: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/12.jpg)
13
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
BubbleStorm Components
Random Topology Allows efficient sampling of peers at random
Topology Measurement Computes network size and statistics
See PODC’07 brief announcement
Bubblecast Replicates queries/data onto peers quickly
Bubble Maintainer Preserves the correct number of replicas
Covered in this talk
![Page 13: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/13.jpg)
14
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Random Multigraph Topology
Random graphs support the birthday paradox Exploring an edge leads to a randomly sampled peer creation of random node subset (bubble) is cheap
Node degree is chosen proportional to bandwidth As random walks (and bubblecasts) follow edges
with equal probabilityUtilization will be balanced for heterogeneity
![Page 14: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/14.jpg)
15
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Random Multigraph Topology
The topology is a random permutation of its edges It is modified only when peers join or leave
Topology Eulerian Cycle
![Page 15: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/15.jpg)
16
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Join Algorithm
Topology Eulerian Cycle
Joining Node
Contact bootstrapping node Random walk finds a random edge Split the edge and insert in between Multiple joins are executed in parallel or
iteratively
![Page 16: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/16.jpg)
17
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Leave Algorithm
Leave splices two neighboring edges together Join and leave do not change degree of neighbors
Topology Eulerian Cycle
Leaving Node
![Page 17: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/17.jpg)
18
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Bubblecast Motivation
Random Walk
- high latency- unreliable+ precise length+ balanced link load
Bubblecast
+ low latency+ reliable+ precise node count+ balanced link load
Flooding
+ low latency+ reliable- imprecise node count- unbalanced link load
node counter (not hops)fixed branch factor
branch in every step
![Page 18: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/18.jpg)
19
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Example Bubblecast Execution
Decrement the counter for matching locally
17
8 8
4 3 4 3
2 1 1 1 2 1 1 1
1 1
- 1
- 1 - 1
- 1 - 1- 1 - 1
- 1 - 1
Split the counter between two neighbors Counters are always integral
Forwarding terminates when counter reaches 0 Final routing depth differs by at most one hop
A counter specifies the number of replicas to create
![Page 19: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/19.jpg)
20
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Bubblecast Properties
Used for query and data replication
Fixed branch factor balances load Same stationary distribution as a random walk
Counter for edges crossed, not hops Precisely controls replica count
Logarithmic routing depth Slightly deeper than flooding
Message loss reduces replication by log(size)
Samples random nodes… due to random topology
![Page 20: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/20.jpg)
21
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Complexity and Correctness
BubbleStorm costs roughly bandwidth / op.to provide exhaustive search with
The full equation ( ) is complicated by Heterogeneous peer capacity (H) Dependent sampling (due to repeated withdrawals) Unequal query and post traffic (; BS optimizes this)Full details in the paper
€
c n
€
P(failure) < e−c 2
c 1 2 3 4
P(success) 63.21% 98.17% 99.99% 99.99999%
€
e−c 2 +c 3HΥ
![Page 21: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/21.jpg)
22
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Heterogeneous System
Simulation parameters 1 million peers with heterogeneous upstream:
60% 16kB/s, 25% 32kB/s,10% 128kB/s, 5% 1.2MB/s
100B query every 5 user minutes (80/20 injection) 2kB meta data stored every 30 user minutes
Exponential lifetime, mean 60 minutes 10% of leaves are crash failures
Target reliability is 98.2% (c=2)
![Page 22: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/22.jpg)
23
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Apples to Apples Performance: Success
0
0.2
0.4
0.6
0.8
1
100 1000 10000 100000 1e+06
Success Probability
Network size (nodes)
BubbleStormRandom Walk
Gnutella
Search success remains unaffected by increasing the network size
![Page 23: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/23.jpg)
24
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Apples to Apples Performance: Latency
0.1
1
10
100
1000
100 1000 10000 100000 1e+06
Latency (s)
Network size (nodes)
RW PostRW QueryRW MatchGnu QueryGnu Match
BS PostBS QueryBS Match
BS = BubbleStormGnu = GnutellaRW = Ferreira P2P’05
Post = Data replicated Query = Query completed
Match = First hit found
![Page 24: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/24.jpg)
25
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Apples to Apples Performance: Bandwidth
0.01
0.1
1
10
100
1000
10000
100 1000 10000 100000 1e+06
Total System Traffic (MB/s)
Network size (nodes)
Uplink CapacityGnutella
TopologyRandom Walk
Bubblecast
![Page 25: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/25.jpg)
26
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Homogeneous Network: Leave 50%
Give all peers 10kB/s upstream (heterogeneity would help) Set 50% to depart (gracefully) after 1 minute
0.95
0.96
0.97
0.98
0.99
1
27:00 29:00 31:00 33:00 35:00
Probability/Fraction reached
Time (mm:ss)
Unique peersSuccess
![Page 26: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/26.jpg)
27
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Homogeneous Network: Crash 50%
Crash 50% of peers after 1 minute Echo effect: posted data is missing
0
0.2
0.4
0.6
0.8
1
27:00 29:00 31:00 33:00 35:00
Probability/Fraction reached
Time (mm:ss)
Unique peersSuccess
![Page 27: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/27.jpg)
28
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
Current and Future Work
Replica preservation in persistent bubbles Sustain bubble sizes under churn Scale bubble sizes with network size / composition
Update content Non-destructive, versioned updates Delete with death certificates
Release implementation
![Page 28: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/28.jpg)
29
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
BubbleStorm Properties in Recap
Unstructured Queries may be in any language
Heterogeneous Exploits peers with varied capacity
Load Balanced Stationary utilization distribution is flat
Resilient Survives 50% crash fail and 90% leave
Exhaustive All matches of an operation can be retrieved
Probabilistic Success is a tunable guarantee
![Page 29: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/29.jpg)
30
DA
TA
BA
SE
S A
ND
DIS
TR
IBU
TE
D S
YS
TE
MS
TE
CH
NIS
CH
E U
NIV
ER
SIT
ÄT
DA
RM
ST
AD
T
SIGCOMM’07: 1. Motivation 2. Overview 3. Topology 4. Bubblecast 5. Evaluation
When does BubbleStorm fit?
Complex query languages Keyword search and beyond
Zipfian load with large Partitioning data will create an all-pairs sub-problem
Mostly static data Allows us to trade post traffic for search traffic
Highly volatile networks Unstructured topology recovers quickly
![Page 30: Www.dvs1.informatik.tu-darmstadt.de BubbleStorm: Resilient, Probabilistic, and Exhaustive Peer-to-Peer Search Wesley W. Terpstra, Jussi Kangasharju, Christof](https://reader035.vdocument.in/reader035/viewer/2022062410/5697bffa1a28abf838cc06e0/html5/thumbnails/30.jpg)
www.dvs1.informatik.tu-darmstadt.de
?Questions
Thanks for listening!