Resource Addressable Network: An Adaptive Peer-to-Peer Discovery Substrate for Internet-Scale Service Platforms

1
Resource Addressable Network An Adaptive
Peer-to-Peer Discovery Substrate for
Internet-Scale Service Platforms

• Balasubramaneyam Maniymaran
• Ph.D. Student,
• Department of Electrical Computer Engineering,
• McGill University
• Supervisor Dr. Muthucumaru Maheswaran

2
Introduction

• On-demand computing (ODC) an emerging model for
  next generation systems.
• Peer-to-peer (P2P) is one way of building ODC
  systems.
• P2P Grid, P2P CDNs, public computing utilities.
• To assemble ODC from P2P resource base.
• Need a generalized resource discovery scheme.
• Discover resources based on given requirements.
• Resource addressable network (RAN).
• Discovers resources based on attributes and
  location.
• One of the major concerns in RAN is scalability
• Low overhead in managing overlay and information.
• Three design concepts fully decentralized,
  distributed knowledge, and adaptive design.

3
Resource Addressable Network

• RAN middle layer between services and resources.
• Attribute-based and location-based discovery.

ODC Service
Naming the resources based on their attributes
Profile-based naming
Network positioning mechanism, assigning
coordinates for each node in the network delay
space
Landmark-aided positioning
Physical Resources
4
Network Positioning

• Network positioning assigning coordinates for
  the nodes in a virtual Cartesian space, from
  which real network delay can be predicted.

Internet
l12
l12 v(x1-x2)2(y1-y2)2
y
(x2, y2)
Cartesian space
(x1, y1)
x
5
Landmark Aided Positioning

• Landmark aided positioning (LAP) the network
  positioning scheme for RAN
• Using a set of landmarks.
• Other nodes
• Select a subset of the total landmarks and ping
  them.
• Run optimization algorithm to position themselves
  to minimize the total error in distance
  prediction.
• Two phases of LAP
• Landmark positioning positioning the landmarks.
• Node positioning positioning the normal nodes.

6
LAP Landmark Positioning

• Each landmark calculates its coordinate relative
  to other landmarks.
• Landmark positioning involves two loops
• Inner loop contains the iteration for node
  positioning.
• Mostly affects the computational complexity.
• Outer loop contains many node positioning phases.
• Between each node positioning phase, nodes have
  to contact others to get their new coordinates ?
  message complexity.
• Simplex and Spring both found to be producing
  high outer loop iterations.
• Introducing new algorithm called SpringEq.

7
LAP Landmark Positioning (cont)

• SpringEq (short for spring in equilibrium)
• Inspired from Spring same spring system concept.
• But instead of minimizing the deformations in the
  spring, SpringEq consider the equilibrium
  condition.
• The resultant force applied at each node is zero.
• A spring system at equilibrium can be modelled by
  a set of simultaneous equations.
• SpringEq solves this simultaneous equation using
  fast iterative process.

8
LAP Landmark Positioning (cont)
SpringEq
Simplex
Simplex
Spring
SpringEq
Spring
Distance correlation vs. ping error.
No. of iteration vs. ping error.

• Random network configuration 100 landmarks.
• Distance correlation correlation between the
  ping and calculated distance matrices.
• Simplex good prediction, but too many
  iterations Spring comparatively few
  iterations, but bad prediction
• SpringEq outperforms both Simplex and Spring.

9
Clustered LAP

• Non-random network errors severely impact
  positioning.
• Data smoothing and optimization can not handle
  this.
• Clustered landmark aided positioning (CLAP)
• CLAP assumptions
• Network errors are created by abnormal
  congestion.
• Peers within the same network segment can be
  trusted.
• Congestion does not affect pings within a network
  segment.

join a cluster
ping others and share the values with others
cluster initialization
calculate cluster diameter
find coordinates using simple LAP (SLAP)
inter exchange SLAP coordinates
CLAP adjustment
calculate cluster centroid
tune SLAP coordinates so that it lies within
cluster diameter
10
CLAP Performance
CLAP
Experiment with Planetlab data.
CLAPs minimum performance is better than SLAPs
maximum performance.
SLAP
CLAP is relatively robust
Variation of distance correlation with increasing
network congestion.
11
CLAP Performance (cont)
CLAP
SLAP
Cumulative distribution of relative distance
error in the system for different amount of
network congestion.
12
Location-based Discovery

• Finding a resource at specific coordinate/range
• Multidimensional search spatial data structure.
• Chose Hilbert curve as the data structure.
• Hilbert curve
• Provides a d-D to 1-D mapping.
• Preserving proximity.
• Hierarchical Hilbert index ? location ID (LID).

13
Location-based Discovery (cont)
Hilbert mapping of the nodes in Planetlab network
(n 133, approximation level 7)
14
Location-based Discovery (cont)
Routing table at node with LID 2.3.3.1.0

• Routing table for location-based discovery.
• Non-zero error in pings justifies fixed length
  LIDs.
• Ring pointers ensuring connectivity jump
  pointers enhancing route complexity.
• Average search hop complexity h (approx. level)
  ? O(1).

15
Profile-based Discovery

• Discovery systems implements naming schemes
• Label-based naming (LBN) DNS, IP Address.
• Scalable, but not flexible.
• Description-based naming (DBN) LDAP.
• Flexible, but with high overhead due to
  information maintenance, complex matching
  algorithms.
• Introducing profile based naming (PBN)
• Labels popular attribute-value combinations.
• Combines the goods of LBN and DBN.
• Can not discover all the attribute-value
  combinations.
• Trading off flexibility (performance) for
  scalability.

16
Profile-based Discovery (cont)
profiles
profile space
description
1
2
3
Profile IDs
description space
Profile 1 Intel/AMD, 512MB 0. Profile 2
Intel with 1GB 1.0 Profile 3 Intel/AMD,
gt 1GB 1.1,1.2

• Profile-based routing table is very similar to
  location-based routing table.

17
Related Works

• Network positioning
• GNP
• Centralized implementation, fixed set of
  landmarks.
• Vivaldi
• Dynamic landmarks anybody can be a landmark.
• New node disturbing others, requires RPC calls.
• Others NPS, PIC, big-bang simulation, PCoord.
• LAP
• Semi-dynamic landmarks.
• Low message overhead design.
• CLAP improvement.
• RAN infrastructure helps selecting landmarks.

18
Related Works (cont)

• Location-based discovery
• SkipNet proximity based on DNS names fails
  outside DNS structure.
• Pastry, expressway of CAN document discovery.
• RAN
• Proximity information is available at any
  resolution.
• No indirection.
• Fixing the search hop complexity.
• Attribute-based discovery
• Directory services LDAP, MDS.
• Intentional naming scheme/Twine
• Document discovery for resource discovery.
• RAN
• Trading off performance for scalability.
• No indirection.

19
Conclusion

• Expected contributions
• Architecture
• Extending the concept of structured-document
  discovery to resource discovery
• Extracting a structure out of the unstructured
  metric space using Hilbert curve.
• First discovery structure combining
  attribute-based and location-based discovery.
• Network positioning
• CLAP resilient to network congestion.
• SpringEq providing low message complexity.
• Spring vs. Simplex comparison.

20
Conclusion (cont)

• Expected contributions (cont)
• Location-based discovery
• Efficient overlay design using Hilbert indices.
• Fixing the search complexity by fixing the search
  resolution.
• Profile-based naming
• Trading off flexibility for scalability.
• Efficient profile-based routing overlay design.
• Profile-based search complexity depends on
  popularity distribution.

21
Conclusion (cont)

• Roadmap to completion
• LAP
• Analysis of SpringEq for its convergence and
  stability. (Sep. 2005)
• Architecture
• The deficiencies the routing mechanism can face
  due to the non-uniformity of metric space will be
  studied. (Oct. 2005)
• Location-based discovery
• A practical value for search resolution will be
  found based on errors in pings and the
  applications requirements. (Nov. 2005)
• Simulation study. (Mar. 2006)
• Profile-based discovery
• Analysis of other possible schemes that can map
  description onto profile space. (May 2006)
• Impact of incorporating virtual profiles. (July
  2006)

22
Thank you
23
Outline

• Introduction and motivation
• Resource addressable network (RAN)
• Network positioning
• Landmark aided positioning (LAP)
• Clustered LAP (CLAP)
• Location-based discovery
• Profile-based discovery
• Related works
• Contributions and conclusion

24
Structured discovery
Creating structured discovery scheme
Objects
Resources
Structured metric space
Network delay space fit with HC
Resources arranged using Hilbert indices
Structured overlay of objects
Resource discovery
25
Landmark Aided Positioning

• Landmark aided positioning (LAP) the network
  positioning scheme for RAN

• Using a set of landmarks.
• Stable nodes, expandable membership.
• Other nodes
• ping landmarks.
• Run optimization algorithm to position themselves
  to minimize the total error in distance
  prediction.

l2
l1
l3
Error in positioning related to L2
26
LAP Node Positioning

• Two algorithms used in the literature
• Simplex downhill method
• Spring
• Considers nodes are connected using springs.
• Ping value between nodes is the natural length of
  the spring.
• Adjust the position of the nodes such that the
  total deformation in all the springs is minimized.

27
LAP Node Positioning (cont)
Simplex
Simplex
Spring
Spring
No. of iteration vs. ping error
Relative distance error vs. ping error

• Random network configurations no. of landmarks
  5.
• Spring performs better than Simplex therefore,
  it is chosen for node positioning in LAP.

28
Clustered LAP (cont)

• CLAP operation
• Nearby nodes collaborate together to form
  clusters.
• In my experiments, I assumed clustering of nodes
  within the same/nearby AS domains.
• Nodes within a cluster collects all-to-all pings
  and compute the diameter of the cluster.
• Nodes undergo simple LAP (SLAP) procedure to find
  their coordinates.
• Each node has different set of landmarks.
• Then, nodes perform a CLAP adjustment routine.
• Nodes within cluster exchange their coordinates.
• Calculates the centroid.
• If the distance from centroid is more than ½ of
  the diameter, the nodes coordinates are adjusted
  towards the centroid.

29
CLAP Performance (cont)
CLAP
Rank error error in ranking the nodes in the
network based on proximity.
SLAP
Cumulative distribution of rank error for various
amount of network congestion.
30
CLAP Performance (cont)
Max. occupancy diameter the diameter of the
minimum area circle that covers all the points a
node occupied in a window of positioning
operations.
SLAP
CLAP
Note occupancy diameter is always non-zero
Variation of max. occupancy diameter in the
system with increasing network congestion.
31
CLAP Congestion Detection
nodes in the congested segment
SLAP
nodes in the non-congested segment
CLAP
Point congestion
32
Location-based Discovery (cont)

• Why Hilbert curve?
• Names in flat space Hilbert indices
• No bottleneck as in other hierarchical data
  structures.
• Recursive structure hierarchical
  representation
• Possible distributed hash table (DHT) type
  overlay creation.
• Preserving proximity in naming
• Possible false-negative, but no false-positive.
• Index of a point is fixed for a specific approx.
  level
• Addition/deletion of nodes do not affect others.
• Availability of mathematical constructs for
  finding the Hilbert index of a coordinate.
• Hierarchical Hilbert index ? location ID (LID).

33
Profile-based Discovery (cont)

• Profile-based naming (PBN)
• Combining of description-based label-based
  naming.
• Trading off flexibility for scalability.
• Justifying PBN
• Analysis on CNET.
• Shows popularity is
• largely skewed.

383 (13.6)
110 (3.6)
2808
34
Profile-based Discovery (cont)
Profile-based routing table at node with PID
7.5.1

• Profile-based routing table similar to
  location-based one
• Differences
• Not necessarily 2-D.
• Aggregation of profiles.
• Inclusion of LIDs with PIDs ? for neighbourhood
  rings.

35
RAN Overlay
Location ID
Neighborhood pointers connect the rings
Hilbert indexing
decides the location
LAP
decides the ring
PBN/Hilbert indexing
Type rings
Profile ID
Resources with the same profile ID form a ring
Route pointers in the nodes creates the overlay
structure