Vasilios Darlagiannis, EPFL

1
SP3 Overlay Networks and ServicesWP3k
Algorithms for Overlay Networks

• Vasilios Darlagiannis, EPFL
• Partners KTH, SICS, EPFL, HUJI, UCL

2
Contents

• Introduction
• Highlights of previous years
• Achievements during the last year of Evergrow
• Low maintenance structured overlays
• Scalable routing schemes and latency
  approximation
• Overlay network merging
• Self management of large-scale systems
• Conclusions

3
Conclusions
Connectivity

• Overlay Networks
• Significant dimensions

P-Grid
DKS
Oscar
Flexibility for application developers
P-Grid
Jupp
Chord
Routing table lt-gt load balancing
Chord
DKS
Fuzzynet
P-Grid
Constraint topology (Ring)
4
Conclusions

• Services over SONs

P-Grid DKS Chord Oscar Fuzzynet Jupp
Broadcast
DistanceEstimation
BeliefPropagation
Semantics
5
Motivation

• Study overlay networks as Complex System
• Overlay networks already produce more than 60 of
  the Internet traffic
• Overlay networks are complex systems
• Complex evolving structures
• Complex dynamics
• Traditional approaches for distributed algorithms
  and distributed services are not effective
• No ability to largely control the system state
• No ability to enforce constraints or centralized
  coordination
• gt Novel Paradigms inspired by Complex Systems

6
Highlights in 2005

• Extension of Kleinbergs model
• Novel type of structured overlays which support
  load-balancing for imbalanced key and work-load
  distributions
• Case study to evaluate standard methodologies for
  studying stochasticity of probabilistic systems
• Mean value analysis
• Evolution of probability mass distribution
  function
• P2PKit middleware enabling the implementation of
  overlays
• P2PS library for enabling overlay deployment in
  hostile environments
• Julia content distribution network (theoretical
  foundations and deployment)
• Measurement of global properties in complex
  systems with low overhead
• Efficient and adaptive estimation of network size
• MyriadStore design document

7
Highlights in 2006

• Metric Embedding techniques
• Approximation of Internet latencies
• Utilizing labels to estimate distances
• Investigating scalable routing schemes
• Oscar Load-balancing for heterogeneous peer
  populations
• Release simplification assumptions on key
  distributions
• Consideration of heterogeneous peer capabilities
• Small-world principle
• Gradual PDF approximation with local information

8
Highlights in 2006

• Distributed k-ary System (DKS)
• Atomic Lookup Consistency
• Reliable Lookup Algorithms
• Group Communication
• Proximity Routing Schemes
• Usage of index replicas
• Proximity Route Selection (PRS)
• Proximity Neighbor Selection (PNS)
• Parallel lookups

9
Accomplishments during last year
10
Fuzzynet Zero maintenance ringless overlay (1)

• Motivation
• Advantages of the ring
• Easy Navigation (greedy routing)
• Clear responsibility ranges
• Easy to bootstrap long-range links
• BUT!
• Keeping the ring invariant is a difficult task
• Expensive maintenance (periodic, eager)
• Non-transitivity effect (A?B, B? C, but not A? C
  )
• Firewalled peers, NATs
• Routing anomalies

11
Fuzzynet Zero maintenance ringless overlay (2)

• Fuzzynet
• No ring structure (only Small-World long-range
  links)
• No predefined responsibility ranges
• Data is probabilistically stored in the data-key
  vicinity
• Compatible with any Small-World network
• Typical DHT replication rate
• Network construction without the help of the ring
  (peer order is considered)
• Lookup (Read) simple greedy routing
• Publish (Write) greedy routing write burst

12
Write phase 1 Greedy-Approach

• Routing from the originator peer (P0.56) to the
  greedy-closest peer (P0.21) where the greedy
  approach towards the target key 0175
  (actual-closest peer P0.17) is no further
  possible.

13
Write phase 2 Write-Burst

• The greedy-closest peer (P0.21) seeds the
  replicas in the cluster vicinity of the key 0.175
  using the Write-Burst.

14
Lookup (read)

• After writing the data in the vicinity of the key
  0.175, the lookup (read) from any node will have
  very high chance finding at least one of the data
  replicas.

15
Fuzzynet Results

• Theoretical analysis (lower bound on success
  rate)
• Planet Lab Simulation results
• Fuzzynet Zero-maintenance Ringless Overlay.
  Submitted to ACM TAAS.

16
Routing in DHTs

• A novel efficient routing table construction
  algorithm
• arbitrary skews in the peer distribution
• cheap construction
• low routing cost
• easy to implement
• no maintenance
• lost routing entries do not have to be replaced
• hop counts are never updated
• Jupp On Routing in Distributed Hash Tables. The
  Seventh IEEE International Conference on
  Peer-to-Peer Computing

17
Solution Hop Space

• Separate the identifier space from the routing
  topology
• With each link, associate distance in direct
  neighbor hops
• Hop space only used for table construction
• No changes of the routing protocol
• Peers route connection requests in the hop space
  (independent of ID space)
• Table construction
• Create links with hops distances 2, 4, 8, 16,
  etc.

2
4
8
16
1
A
B
C
D
E
0
1
0.5
17
18
Routing Range Queries

• Recursive
• Choose all routing entries that fall into the
  queried range, plus the one before the range.
• Split range query into sub-range queries
• n number of peers
• m number of peers in the range
• Cost O((log n) m)
• Time O(log n)

A
id2
F
id1
E
D
C
B
18
19
Simulation Results
20
Scalable routing schemes and latency approximation

• Several theoretical results
• Algorithm with formal guarantees for finding
  faithful and low dimensional representation of
  data in high dimensional space
• Measure to assign a value to every metric space
• Quantify how close is to a tree metric
• Embeddings with distortion
• depending solely in local structure of the space
• Distributed name-independent routing scheme
• Embedding Metric Spaces in their Intrinsic
  Dimension, (SODA 2008)
• Reconstructing Approximate Tree Metrics, (PODC
  2007)
• Local Embeddings of Metric Spaces, (STOC 2007)
• Strong-Diameter Decompositions of Minor Free
  Graphs, (SPAA 2007)

21
Merging two similar structured overlays

• Multiple networks can result from
• Network partitioning
• Separate construction
• Challenging issues
• How can they merge?
• What is the complexity?
• What is the operational cost?
• Use cases
• Ring based overlay networks
• Structurally replicated overlay networks

22
Merging Illustration on Chord
23
Results on overlay merging

• Ring based overlay networks
• Cannot function at all until merging is complete
• Data transfer complexity
• Structurally replicated overlay networks
• Merging process is graceful and system can
  operate
• Location of all structural replicas can become
  tricky
• Complex replica management
• The challenges of merging two similar structured
  overlays A tale of two networks, International
  Workshop on Self-Organizing Systems (IWSOS)

24
Network Partitions in Rings

• Two algorithms for merging two similar ring-based
  SONs
• Simple Ring Unification
• Periodic process
• Builds on top of Chord lookup
• O(N) time complexity
• Unable to recover from pathological scenarios
• Gossip-based Ring Unification
• Extension of the Simple Ring Unification
• Triggering multiple merging processes (fan-out
  parameter F)
• Handling Network Partitions and Mergers in
  Structured Overlay Networks, IEEE P2P2007

25
r
r.succ
q
SON 1
SON 2
26
Results

• Simulations over SicsSim
• Time complexity with multiple nodes detecting
  the merger
• Message complexitywith multiple nodes detecting
  the merger
• Loopy network cases have beenalso investigated

27
Self management of large-scale systems

• Relaxed Ring Architecture
• Lookup consistency in self-managed SONs
  (ring-based)
• Handling multiple joins and leaves/failures
• Self-organization and Self-healing
• Applied on P2PS
• Feedback loops
• Improving the Peer-to-Peer Ring for Building
  Fault-Tolerant Grids, CoreGRID Workshop on Grid
  Programming Model, Grid and P2P Systems
  Architecture, Grid Systems, Tools, and
  Environments
• A Relaxed-Ring for Self-Organising and
  Fault-Tolerant Peer-to-Peer Networks, XXVI
  International Conference of the Chilean Computer
  Science Society (SCCC 2007)

28
Self management of large-scale systems

• Key-Based Consistency and Availability in SONs
• Evaluation of inconsistency
• Inaccurate failure detectors is the major
  inconsistency factor
• Trade-off between availability and consistency
• The probability of a responsibility inconsistency
  is very low
• Majority-based quorums have been considered
• Key-Based Consistency and Availability in
  Structured Overlay Networks, High-Performance
  Distributed Computing (HPDC)

29
Results

• Majority based quorums
• Probability of getting disjoint majority sets
  for a replica set
• Number of replicas
• Number of inconsistencies
• Probability of disjoint majority sets for two
  majority based operations in one configuration
  of a SON
• Accuracy of the failure detector

30
Conclusions

• Services over SONs
• Broadcast
• Semantic overlays
• Distance estimation
• Belief propagation

31
Summary of achievements

• The research done in WP3l which has been
  completed in the final year of Evergrow produced
  many interesting results including
• Ringless overlay network
• Consistency in Lookup operations
• Merging of overlay networks
• Scalable routing schemes
• Latency approximation mechanisms
• Several publications in Journals, conferences and
  workshops

32
Dissemination

• Publications
• 1 Fuzzynet Zero-maintenance Ringless Overlay.
  Submitted to ACM TAAS.
• 2 On Routing in Distributed Hash Tables. The
  Seventh IEEE International Conference on
  Peer-to-Peer Computing
• 3 Embedding Metric Spaces in their Intrinsic
  Dimension, ACM-Siam Symposium on Discrete
  Algorithms (SODA 2008)
• 4 Reconstructing Approximate Tree Metrics, 26th
  ACM Symposium on Principles of Distributed
  Computing (PODC 2007)
• 5 Local Embeddings of Metrics Spaces, 39th ACM
  Symposium on Theory of Computing (STOC 2007)
• 6 Strong-Diameter Decompositions of Minor Free
  Graphs, 19th ACM Symposium on Parallelism in
  Algorithms and Architectures (SPAA 2007)
• 7 The challenges of merging two similar
  structured overlays A tale of two networks,
  International Workshop on Self-Organizing Systems
  (IWSOS)
• 8 Dealing with Network Partitions in Structured
  Overlay Networks,
• 9 Improving the Peer-to-Peer Ring for Building
  Fault-Tolerant Grids, CoreGRID Workshop on Grid
  Programming Model, Grid and P2P Systems
  Architecture, Grid Systems, Tools, and
  Environments
• 10 A Relaxed-Ring for Self-Organising and
  Fault-Tolerant Peer-to-Peer Networks, XXVI
  International Conference of the Chilean Computer
  Science Society (SCCC 2007)
• 11 Key-Based Consistency and Availability in
  Structured Overlay Networks, High-Performance
  Distributed Computing (HPDC)
• 6 Ph.D. Thesis

33
Plans for the Future

• P-Grid (http//www.p-grid.org )
• Code is used in 5 other EU projects
• Bricks
• Alvis
• Nepomuk
• TEAM
• OCCAM
• DKS http//dks.sics.se
• Grid4All
• SelfMan