Resource Addressable Network: Resource Addressable Network: An Adaptive Peer-to-Peer Discovery An Adaptive Peer-to-Peer Discovery Substrate for Internet-Scale Service Substrate for Internet-Scale Service Platforms Platforms Balasubramaneyam Maniymaran Balasubramaneyam Maniymaran Ph.D. Student, Ph.D. Student, Department of Electrical & Computer Engineering, Department of Electrical & Computer Engineering, McGill University McGill University Supervisor: Dr. Muthucumaru Maheswaran Supervisor: Dr. Muthucumaru Maheswaran
22
Embed
Resource Addressable Network: An Adaptive Peer-to-Peer Discovery Substrate for Internet-Scale Service Platforms Balasubramaneyam Maniymaran Ph.D. Student,
Advanced Networking Research Laboratory, The School of Computer Science, McGill University, Montreal, QC, Canada. 3 RAN discovery substrate ODC Service Physical Resources Location-based discovery Landmark-aided positioning Profile-based discovery Network positioning mechanism, assigning coordinates for each node in the network delay space Resource Addressable Network RAN: middle layer between services and resources.RAN: middle layer between services and resources. Attribute-based and location-based discovery.Attribute-based and location-based discovery. Naming the resources based on their attributes Profile-based naming
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Resource Addressable Network: Resource Addressable Network: An Adaptive Peer-to-Peer Discovery An Adaptive Peer-to-Peer Discovery Substrate for Internet-Scale Service Substrate for Internet-Scale Service PlatformsPlatforms
Balasubramaneyam ManiymaranBalasubramaneyam ManiymaranPh.D. Student,Ph.D. Student,Department of Electrical & Computer Engineering,Department of Electrical & Computer Engineering,McGill UniversityMcGill University
Supervisor: Dr. Muthucumaru MaheswaranSupervisor: Dr. Muthucumaru Maheswaran
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
2
IntroductionIntroduction• On-demand computingOn-demand computing (ODC) an emerging model for (ODC) an emerging model for
next generation systems.next generation systems.• Peer-to-peerPeer-to-peer (P2P) is one way of building ODC (P2P) is one way of building ODC
systems.systems.– P2P Grid, P2P CDNs, public computing utilities.P2P Grid, P2P CDNs, public computing utilities.
• To assemble ODC from P2P resource base.To assemble ODC from P2P resource base.– Need a generalized resource discovery scheme.Need a generalized resource discovery scheme.– Discover resources based on given requirements.Discover resources based on given requirements.
• Resource addressable networkResource addressable network (RAN). (RAN).– Discovers resources based on attributes and location.Discovers resources based on attributes and location.– One of the major concerns in RAN is scalability:One of the major concerns in RAN is scalability:
• Low overhead in managing overlay and information.Low overhead in managing overlay and information.• Three design concepts: fully decentralized, distributed Three design concepts: fully decentralized, distributed
knowledge, and adaptive design.knowledge, and adaptive design.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
Network positioning Network positioning mechanism, assigning mechanism, assigning coordinates for each node in coordinates for each node in the network delay spacethe network delay space
Resource Addressable Resource Addressable NetworkNetwork• RAN: middle layer between services and RAN: middle layer between services and
resources.resources.• Attribute-based and location-based discovery.Attribute-based and location-based discovery.
Naming the resources Naming the resources based on their based on their attributesattributes
Profile-based namingProfile-based naming
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
for the nodes in a virtual Cartesian space, for the nodes in a virtual Cartesian space, from which real network delay can be from which real network delay can be predicted. predicted.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
5
Landmark Aided PositioningLandmark Aided Positioning• Landmark aided positioning Landmark aided positioning (LAP): the (LAP): the
network positioning scheme for RANnetwork positioning scheme for RAN• Using a set of Using a set of landmarkslandmarks..• Other nodes:Other nodes:
– Select a subset of the total landmarks and ping Select a subset of the total landmarks and ping them.them.
– Run optimization algorithm to position themselves Run optimization algorithm to position themselves to minimize the total error in distance prediction.to minimize the total error in distance prediction.
• Two phases of LAP:Two phases of LAP:– Landmark positioningLandmark positioning: : positioning the landmarks.positioning the landmarks.– Node positioningNode positioning: : positioning the normal nodes.positioning the normal nodes.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
6
LAP:: Landmark PositioningLAP:: Landmark Positioning• Each landmark calculates its coordinate Each landmark calculates its coordinate
relative to other landmarks.relative to other landmarks.• Landmark positioning involves two loops:Landmark positioning involves two loops:
– Inner loopInner loop contains the iteration for node contains the iteration for node positioning.positioning.• Mostly affects the computational complexity.Mostly affects the computational complexity.
– Outer loopOuter loop contains many node positioning phases. contains many node positioning phases.• Between each node positioning phase, nodes have to Between each node positioning phase, nodes have to
contact others to get their new coordinates contact others to get their new coordinates message message complexity.complexity.
• Simplex and Spring both found to be producing high outer Simplex and Spring both found to be producing high outer loop iterations.loop iterations.
• Introducing new algorithm called Introducing new algorithm called SpringEqSpringEq..
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
7
• SpringEq (short for “spring in equilibrium”):SpringEq (short for “spring in equilibrium”):– Inspired from Spring; same spring system Inspired from Spring; same spring system
concept.concept.– But instead of minimizing the deformations in But instead of minimizing the deformations in
the spring, SpringEq consider the equilibrium the spring, SpringEq consider the equilibrium condition.condition.• The resultant force applied at each node is zero.The resultant force applied at each node is zero.
– A spring system at equilibrium can be modelled A spring system at equilibrium can be modelled by a set of simultaneous equations.by a set of simultaneous equations.
– SpringEq solves this simultaneous equation SpringEq solves this simultaneous equation using fast iterative process.using fast iterative process.
• Random network configuration; 100 landmarks.Random network configuration; 100 landmarks.• Distance correlationDistance correlation: correlation between the ping and calculated : correlation between the ping and calculated
distance matrices.distance matrices.• Simplex – good prediction, but too many iterations; Spring – Simplex – good prediction, but too many iterations; Spring –
comparatively few iterations, but bad prediction;comparatively few iterations, but bad prediction;– SpringEq outperforms both Simplex and Spring.SpringEq outperforms both Simplex and Spring.
SimplexSimplex SimplexSimplex
SpringSpring
SpringSpringSpringEqSpringEq
SpringEqSpringEq
Distance correlation vs. ping error.Distance correlation vs. ping error.No. of iteration vs. ping error.No. of iteration vs. ping error.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
centroidcentroidtune SLAP coordinates so tune SLAP coordinates so that it lies within cluster that it lies within cluster
diameterdiameter
CLAP
CL
AP
adju
stm
ent
adju
stm
ent
clust
er
clust
er
initi
aliza
tion
initi
aliza
tion
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
10
CLAP PerformanceCLAP Performance
SLAPSLAP
CLAPCLAP
CLAP is relatively CLAP is relatively robustrobust
CLAP’s minimum CLAP’s minimum performance is performance is better than SLAP’s better than SLAP’s maximum maximum performance.performance.
Variation of distance correlation with Variation of distance correlation with increasing network congestion.increasing network congestion.
Experiment with Experiment with Planetlab data.Planetlab data.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
11
CLAP PerformanceCLAP Performance (cont…) (cont…)
Cumulative distribution of relative distance error in the Cumulative distribution of relative distance error in the system for different amount of network congestion.system for different amount of network congestion.
SLAPSLAP
CLAPCLAP
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
12
Location-based DiscoveryLocation-based Discovery• Finding a resource at specific coordinate/range:Finding a resource at specific coordinate/range:
– Multidimensional search – Multidimensional search – spatial data structurespatial data structure..– Chose Chose Hilbert curveHilbert curve as the data structure. as the data structure.
• Hilbert curve:Hilbert curve:
– Provides a Provides a d-d-D to D to 11-D mapping.-D mapping.– Preserving proximity.Preserving proximity.– Hierarchical Hilbert index Hierarchical Hilbert index location IDlocation ID (LID). (LID).
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
Hilbert mapping of the nodes in Planetlab Hilbert mapping of the nodes in Planetlab network (n = 133, approximation level = 7)network (n = 133, approximation level = 7)
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
• Routing table for location-based discovery.Routing table for location-based discovery.– Non-zero error in pings justifies fixed length LIDs.Non-zero error in pings justifies fixed length LIDs.– Ring pointersRing pointers ensuring connectivity; ensuring connectivity; jump pointersjump pointers
enhancing route complexity.enhancing route complexity.• Average search hop complexity = Average search hop complexity = h h (approx. level) (approx. level) O(1)O(1)..
Routing table at node with LID = 2.3.3.1.0Routing table at node with LID = 2.3.3.1.0
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
15
Profile-based DiscoveryProfile-based Discovery• Discovery systems implements naming Discovery systems implements naming
schemes:schemes:– Label-based namingLabel-based naming (LBN): DNS, IP Address. (LBN): DNS, IP Address.
• Scalable, but not flexible.Scalable, but not flexible.– Description-based namingDescription-based naming (DBN): LDAP. (DBN): LDAP.
• Flexible, but with high overhead due to information Flexible, but with high overhead due to information maintenance, complex matching algorithms.maintenance, complex matching algorithms.
• Introducing Introducing profile based namingprofile based naming (PBN): (PBN):– Labels popular attribute-value combinations.Labels popular attribute-value combinations.
• Combines the goods of LBN and DBN.Combines the goods of LBN and DBN.• Can not discover all the attribute-value combinations.Can not discover all the attribute-value combinations.• Trading off flexibility (performance) for scalability.Trading off flexibility (performance) for scalability.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
•Profile-based routing table is very similar to location-based routing table.Profile-based routing table is very similar to location-based routing table.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
17
Related WorksRelated Works• Network positioning:Network positioning:
– GNP: GNP: • Centralized implementation, fixed set of landmarks.Centralized implementation, fixed set of landmarks.
– Vivaldi: Vivaldi: • Dynamic landmarks: anybody can be a landmark.Dynamic landmarks: anybody can be a landmark.• New node disturbing others, requires RPC calls.New node disturbing others, requires RPC calls.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
18
Related WorksRelated Works (cont…)(cont…)
• Location-based discovery:Location-based discovery:– SkipNet: proximity based on DNS names – fails outside DNS SkipNet: proximity based on DNS names – fails outside DNS
structure.structure.– Pastry, expressway of CAN: document discovery.Pastry, expressway of CAN: document discovery.– RAN:RAN:
• Proximity information is available at any resolution.Proximity information is available at any resolution.• No indirection.No indirection.• Fixing the search hop complexity.Fixing the search hop complexity.
– Architecture:Architecture:• Extending the concept of structured-document Extending the concept of structured-document
discovery to resource discovery:discovery to resource discovery:– Extracting a structure out of the unstructured Extracting a structure out of the unstructured
metric space using Hilbert curve.metric space using Hilbert curve.• First discovery structure combining attribute-based and First discovery structure combining attribute-based and
• Efficient overlay design using Hilbert indices.Efficient overlay design using Hilbert indices.• Fixing the search complexity by fixing the search Fixing the search complexity by fixing the search
• Trading off flexibility for scalability.Trading off flexibility for scalability.• Efficient profile-based routing overlay design.Efficient profile-based routing overlay design.• Profile-based search complexity depends on popularity Profile-based search complexity depends on popularity
distribution.distribution.
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.
21
• Roadmap to completion:Roadmap to completion:– LAP:LAP:
• Analysis of SpringEq for its convergence and stability. Analysis of SpringEq for its convergence and stability. (Sep. (Sep. 2005)2005)
– Architecture:Architecture:• The deficiencies the routing mechanism can face due to the The deficiencies the routing mechanism can face due to the
non-uniformity of metric space will be studied. non-uniformity of metric space will be studied. (Oct. 2005)(Oct. 2005)– Location-based discovery:Location-based discovery:
• A practical value for search resolution will be found based on A practical value for search resolution will be found based on errors in pings and the applications requirements. errors in pings and the applications requirements. (Nov. 2005)(Nov. 2005)
• Analysis of other possible schemes that can map description Analysis of other possible schemes that can map description onto profile space. onto profile space. (May 2006)(May 2006)
• Impact of incorporating virtual profiles. Impact of incorporating virtual profiles. (July 2006)(July 2006)
Conclusion Conclusion (cont…)(cont…)
Advanced Networking Research Laboratory,The School of Computer Science,McGill University, Montreal, QC, Canada.