Upload
amia-henderson
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
Universität InnsbruckLeopold Franzens
making semanticsreal. Copyright 2006 DERI Innsbruck www.deri.at
LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Towards Scalable Information Spaces
Reto Krummenacher, Elena Simperl, Dieter Fensel
making semanticsreal.2 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Background: Triple Space Computing
Semantic Spacerd
rd
out
making semanticsreal.3 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Background: Triple Space Computing
Host / Kernel
Host / Kernel Host / Kernel
Host / Kernel
Host / Kernel
InformationSpace
InformationSpace
InformationSpace
InformationSpace
InformationSpace
making semanticsreal.4 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Triple Space Computing at Web Scale
• The Web is distributed, heterogeneous, dynamic, and open.
• Required is a Triple Space platform which copes with these principles at Web scale.
• Classical guarantees encountered in closed environments cannot longer hold.
• Adaptation of the functional properties publish and retrieve is required.
• Feasible non-functional guarantees must be determined from a scalability perspective
making semanticsreal.5 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Outline
1. (Non-) Functional Properties and Their Impact on Scalability of Information Spaces
2. Scalability-driven Trade-Offs
3. Levels of Scalability for Semantic Space Middleware
making semanticsreal.6 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Two Dimensions
• There are two dimensions to a middleware that influence its scalability
– The exposed functionality - the service delivered to the clients - the functional properties.
– The internal behavior of the system that is not obvious or visible to clients at interaction time - the non-functional properties (NFP).
making semanticsreal.7 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Functionals (1/2)
• Publishing information– Writing of one or more RDF triples into an identifiable space.– No guarantee about when the triple is available and
persistently stored.
• Retrieving information– Reading triples by template/query.– No guarantees about when and if a match is returned.– An empty result set does not imply that no matching triple is
available.– Retrieval is possible from an identifiable space, or from any
available spaces.
making semanticsreal.8 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Functionals (2/2)
• Query complexity– Ranging from triple patterns <:tripcom ?p ?o> to arbitrarily
complex queries. – Query resolution (reasoning).
• Explicit vs. inferred data– Publication at data level – no reasoning, multiple copies.– Retrieval at knowledge level – inference, materialization, only
„one“ copy.• Local reasoning vs. distributed reasoning.
• Transactions– E.g. when publishing sets of triples within an atomic
operation.
making semanticsreal.9 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Non-Functionals
• Security – retrieve,publish• Fault-tolerance – retrieve,publish• Availability – retrieve,publish• Reliability – retrieve,publish
• Completeness – retrieve• Correctness – retrieve• Consistency – retrieve• Durability – publish
• Response time – retrieve,publish• Scalability – retrieve,publish
data/query related
TS related dependability
making semanticsreal.
Technical Means for NFP
TS Kernel
replication
triplespace 1
TS Kernel
segment 1cluster
segment 2cluster
Primary CopyPrimary Copy Backup
triplespace 2
TS Kernel Primary
Copy
distribution
• Redundancy– duplication of processes (n-out-of-
m, n < m).
• Recovery– Re-installation of previous system
state.
• Load balancing– decentralization of processes
• by replicating data.• by distributing data (partitioning,
clustering).
10 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
making semanticsreal.11 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
NFP and their Realization (1/3)
• Fault-Tolerance– Redundancy: prevention, 1-out-of-2– Recovery
• Availability– Load balancing: 1-out-of-N
• Reliability– Redundancy: prevention, 2-out-of-3
• Completeness– No distribution of data
making semanticsreal.12 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
NFP and their Realization (2/3)
• Consistency– No replication: no system-caused inconsistency
• Correctness
– No false-positives, a query engine issue, completeness
• Durability– Recovery: persistency is given by storage
• Response time– Load balancing
making semanticsreal.
NFP and their Realization (3/3)
• Scalability– The trade-offs between non-functional properties and
between functionals and non-functionals need to be analyzed in order to to achieve a feasible level of scalability.
– The appropriate configurations for the desired behavioral and structural realization and the corresponding scalability level has to be selected.
13 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
making semanticsreal.14 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Scalability Trade-Offs
Correctness CompletenessPrecision vs. recall
Completeness AvailabilityLoad balancing by distribution
Consistency AvailabilityLoad balancing by replication
Consistency Fault ToleranceReplication could be hidden
Reliability Response TimeCompleteness Response TimeConsistency Response TimeDurability Response Time
Non-local writing results in latency
making semanticsreal.15 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Scalability Levels
• Depending on the desired functionality and non-functionality, the space middleware has different possibilities to provide a scalable infrastructure.
The richer functionality the middleware provides, the less can be guaranteed about the non-functional properties and scalability.
making semanticsreal.16 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Scalability Levels: Discussion (1/3)
• Simplest approach - proved and tested with respect to scalability– The target space is indicated by a resolvable URL– Distribution is client-driven– Discovery by use of a DNS-like procedure– Ensures at least local scalability
making semanticsreal.17 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Scalability Levels: Discussion (2/3)
• Most complex approach – likely not scalable– Access to spaces without indication of a URL– (Knowledge-driven) retrieval on the entire Semantic
Web– Query forwarding, distributed discovery of relevant
spaces– Distributed reasoning to integrate distributed
information sources/spaces or large scale (local) reasoning after transfer of triples
making semanticsreal.18 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Scalability Levels: Discussion (3/3)
• Loosening classical IR guarantees is of benefit to achieve scalability– Access to space(s) without indication of a URL– Best effort retrieval – no completeness– No guarantees about consistency– Rather data than knowledge-driven
• Performance– Search scope– Discovery strategy
making semanticsreal.21 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Conclusions
• Semantic spaces offer at least local scalability and local completeness.
• Scalability at global scale requires the management of scalability trade-offs.
• However…– “The times of 100% complete and correct solutions are gone.”
[D. Fensel, STI International 2007]
– „incomplete, incorrect but requested!“ [D. Fensel, IEEE Intelligent Systems
22(6) 2007]
making semanticsreal.22 LarCK Workshop, ISWC/ASWC 2007 - Busan, KoreaApr 10, 2023
Thank you.
Universität InnsbruckLeopold Franzens
making semanticsreal.LarCK Workshop, ISWC/ASWC 2007 - Busan, Korea
Copyright 2006 DERI Innsbruck www.deri.at Apr 10, 2023