Download view - LaBRI

Ordered slicing of very large scale
overlay networks
Mark Jelasity
University of Bologna, Italy
Anne-Marie Kermarrec
INRIA Rennes/IRISA, France
20/10/2006
ALPAGE
1
Motivation
• Potential of P2P technology
• Several applications running on the same
infrastructure
• Need for network partitioning
•
•
•
•
Decentralized
Robust
Handle dynamics
Customizable
Gossip-based approach to “slice” the network
20/10/2006
ALPAGE
2
Outline
1. Introduction
2. Background of gossip-based
membership algorithms
3. Ordered slicing
4. Simulation results
5. Conclusion
20/10/2006
ALPAGE
3
Objective
• Targeted applications
• Deskstop grids
• Testbed platforms (PlanetLab)
• Telco applications
• Resource assignment
• Objective
• Create and maintain partitions (slices) as subsets of the
network in a fully decentralized manner
• Ordered nature: along a single attribute (memory, bandwidth,
computing power)
• Our approach: Use a gossip-based approach to estimate to
which partition a node belongs
• Scalable
• Robust
• Based on local knowledge
20/10/2006
ALPAGE
4
Gossip-based protocols
• Unstructured peer to peer networks
• Highly resilient to failure and dynamics
• Gossip-based membership protocols
• Periodic exchange of information between nodes
• Basic functionality: Peer sampling
• Provide a sample of peers given a metric
• Random sampling
• Applications
• Event dissemination
• Recovery protocols
• Aggregation
20/10/2006
ALPAGE
5
Gossip-based generic protocol
1295
1
3
2
2 6 10 3
10
4
9
8
6
5
7
20/10/2006
ALPAGE
6
Gossip-based generic protocol
3
2
1 2 9 5 6 10 3
1
10
4
9
8
6
5
7
20/10/2006
ALPAGE
7
Gossip-based generic protocol
2 9 10
3
2
1
10
4
9
8
6
5
7
20/10/2006
ALPAGE
8
Design space
• Periodically each peer initiates communication
with another peer
• Peer selection
Selectpeer(): return the IP@ of an alive peer
• View propagation
How peers exchange their membership
information
• View selection: Select (c, buffer)
• c: size of the resulting view
• Buffer: information exchanged
20/10/2006
ALPAGE
9
Peer sampling algorithm
• Common framework for existing
gossip-based protocols
• Lpbcast [Eugster & al, ACM TOCS 2003]
• Cyclon [Voulgaris & al JNSM 2005]
• Newscast [Jelasity & van Steen, 2002]
• Peer selection: Select a peer at random from the
view
• View propagation: Pushpull
• View selection: Select the freshest entries
20/10/2006
ALPAGE
10
Random slices
3
2
1
10
4
9
8
6
5
7
20/10/2006
ALPAGE
11
Random slices
3
0.52
2
0.11
1
0.21
10
0.98
8
0.67
9
0.43
6
0.55
4
0.22
5
0.87
7
0.09
20/10/2006
ALPAGE
12
Random slices
3
0.52
2
0.11
1
0.21
10
0.98
8
0.67
9
0.43
6
0.55
4
0.22
5
0.87
7
0.09
20/10/2006
ALPAGE
13
Random slices
3
0.52
2
0.11
1
0.21
10
0.98
8
0.67
9
0.43
6
0.55
4
0.22
5
0.87
7
0.09
20/10/2006
ALPAGE
14
Ordered slices: System model
• Problem
• Automatically assign a node to a slice
• Along an attribute metric
• System model
• Crash-only nodes
• Each node i has
• an attribute i
• a random number i
• a view of c entries (peer sampling)
• a time stamp
• Each node belongs to one slice
x
20/10/2006
r
ALPAGE
15
Ordered slicing algorithm: basic
operation
Node a
xi
ri
xi
ri
xi
ri
20/10/2006
12
34
22
0.6
0.4
0.3
12
34
22
0.6
0.4
2
12
0.12 0.21
Node b
35
56
78
98
2
37
13
0.45 0.87 0.77 0.21 0.65 0.98 0.12
35
56
78
98
0.65 0.45 0.87 0.77 0.21
13
22
0.3
0.4
34
35
37
2
0.3
56
37
13
0.98 0.12
78
98
0.45 0.6 0.65 0.77 0.87 0.98
ALPAGE
16
Ordered slicing algorithm
Wait (t)
p<- random element from view
buffer<-view U {(my@,
my_stamp, xq , rq )}
Send buffer to p
Receive buffer(p)
view<-freshest c entries from
{buffer(p) U view}
i<- peer such that ( xi  xq )( ri  rq )  0
Send ( xq , rq ) to i
Receive buffer(q) from q
buffer<-view U {(my@,
my_stamp, x p , rp)}
Send buffer to q
view<-freshest c entries from
{buffer(p) U view}
rq  ri
Active thread at peer q
20/10/2006
ALPAGE
Passive thread at peer p
17
Ordered slicing algorithm: maintenance
• New nodes discovered using the random
peer sampling
• Random number ensures uniform spread
• Once the order stabilizes: each node knows
to which slice it belongs
• Example
• A peer with a number <0.5 knows in the first
50% of the nodes according to the metric
• Slice creation and maintenance
20/10/2006
ALPAGE
18
Disorder measure
 (t , ri (t ) ,...., ri
1
20/10/2006
N
2
)


(
t
)

1
/
N
(
j

i
(
t
))
 j
N (t )
j 1
ALPAGE
19
Analogy with average
• Weight conserving property
N
N
N
j 1
j 1
j 1
1 / N  j  i j (t )  1 / N  j  1 / N  i j (t )  0
• The swapping does not influence this
value (=0) but always reduces the
disorder value
20/10/2006
ALPAGE
20
Exponential decrease of the disorder
20/10/2006
ALPAGE
21
Swaps over time
20/10/2006
ALPAGE
22
Simulation set-up
• PeerSim simulator
• Configurations
• Network size: 30,000 – 100,000 – 300,000 node
overlay
• Views: c=20, 40 and 80
• Unreliable communication channels
• Churn
• Age bias: peer selection based on age similarity
• Metric: disorder
20/10/2006
ALPAGE
23
Churn
0.1%
20/10/2006
1%
ALPAGE
24
Age-based technique
Young nodes disordered
Old nodes protected
20/10/2006
ALPAGE
25
Example
• Quick stabilization
• Relatively welldefined slices
• constant churn
• massive failure
• massive join
• Stabilizes as soon
as churn stops
20/10/2006
ALPAGE
26
Conclusion & future work
• Gossip-based solution to partition a
dynamic network according to a given
metric
• Resilient to churn
• Future work
• Slices maintenance
• Improve the peer selection
20/10/2006
ALPAGE
27
Thank you !
20/10/2006
ALPAGE
28