Evaluating Centralized,Evaluating Centralized, Hierarchical, and Networked Hierarchical, and Networked

Architectures for Rule SystemsArchitectures for Rule Systems Benjamin CraigBenjamin Craig

University of New BrunswickUniversity of New Brunswick

Faculty of Computer ScienceFaculty of Computer Science

Fredericton, NB, CanadaFredericton, NB, Canada

Senior Technical Report Presentation Senior Technical Report Presentation

November 20, 2008November 20, 2008

22

OutlineOutline Defining the TerminologyDefining the Terminology

Rules, Distributed Systems, Topologies, OO Rules, Distributed Systems, Topologies, OO jDREW, Rule ResponderjDREW, Rule Responder

Topologies for distributed ArchitecturesTopologies for distributed Architectures Star Topology Advantages and Star Topology Advantages and

DisadvantagesDisadvantages P2P Topology AdvantagesP2P Topology Advantages

Knowledge Maintenance for Rule SystemsKnowledge Maintenance for Rule Systems Knowledge OrganizationKnowledge Organization Knowledge MaintenanceKnowledge Maintenance

ConclusionConclusion

33

What is a rule?What is a rule? Fact (POSL format):Fact (POSL format):

spending(Peter Miller, min 5000 euro, last year).spending(Peter Miller, min 5000 euro, last year).

Rule (POSL format):Rule (POSL format):

premium(?Customer)premium(?Customer) :- :- spending(?Customer, min 5000 euro, last year).spending(?Customer, min 5000 euro, last year).

A deductive rule engine can deduce that A deductive rule engine can deduce that Peter Miller is a premium customer from his Peter Miller is a premium customer from his spendingspending

44

Distributed SystemsDistributed Systems

A distributed system is a set of computer processes that appear to the user as a single system

The distributed system must coordinate all of these processes

Distributed systems are implemented using middleware that creates a communication topology

55

Hierarchical - Star TopologyHierarchical - Star Topology

Single level hierarchySingle level hierarchy

Connects all spokes with Connects all spokes with a centralized huba centralized hub

All information must be All information must be sent through the hub to sent through the hub to the spokesthe spokes

66

Networked (P2P-Like Architecture)

Fully connected networkFully connected network Connects all nodes together Connects all nodes together

with a direct connectionwith a direct connection Full mesh topologyFull mesh topology

Partially connected Partially connected networknetwork Only a subset of nodes are Only a subset of nodes are

connected togetherconnected together Partial mesh topologyPartial mesh topology

77

OO jDREWOO jDREW

Centralized Rule SystemCentralized Rule System Object Oriented Java Deductive Object Oriented Java Deductive

Reasoning Engine for the Web – Reasoning Engine for the Web – extensions of jDREWextensions of jDREW

Supports rules in two formats:Supports rules in two formats: POSL: POSL: Positional Slotted presentation syntaxPositional Slotted presentation syntax RuleML: RuleML: XML interchange syntaxXML interchange syntax

(can be generated from POSL) (can be generated from POSL)

88

Rule ResponderRule Responder Distributed Rule System Is currently implemented as a hierarchical

rule system Rule Responder is a prototypical

multi-agent system for virtual communities

Supports rule-based collaboration between the distributed members of community

Members are assisted by semi-automated rule-based agents, which use rules to describe the decision and behavioral logic

99

1010

Topology PerformanceTopology Performance

When building a distributed system a When building a distributed system a topology is requiredtopology is required

Distributed topologies all have Distributed topologies all have communication over head that communication over head that centralized systems do not havecentralized systems do not have

A key design goal for distributed A key design goal for distributed systems is to minimize this systems is to minimize this communication over headcommunication over head

1111

Star AdvantagesStar Advantages Isolation of spokes from other spokesIsolation of spokes from other spokes

If one spoke fails then it does not affect If one spoke fails then it does not affect othersothers

Adding and removing spokes in the hub is Adding and removing spokes in the hub is trivialtrivial

Hub provides single point of inspection of Hub provides single point of inspection of all traffic through the topologyall traffic through the topology Improved SecurityImproved Security

Trouble shooting is easyTrouble shooting is easy Easy to understand and implementEasy to understand and implement

1212

Star DisadvantagesStar Disadvantages

Scalability, reliability and performance of the star topology rely on the hub

If the hub fails then the entire system fails

The hub can become overloaded and the system will experience slowdown

To prevent the bottleneck of the star topology a P2P topology can be used

1313

P2P AdvantagesP2P Advantages

Removes bottleneck performance Removes bottleneck performance issues of the star topologyissues of the star topology

Whenever a node is added the total Whenever a node is added the total bandwidth capacity is increasedbandwidth capacity is increased

When a node fails the system will be When a node fails the system will be able to recoverable to recover A peer can act in place of another peerA peer can act in place of another peer

1414

Knowledge Maintenance forKnowledge Maintenance for Rule Systems Rule Systems

A distributed system can have many A distributed system can have many different knowledge bases different knowledge bases distributed across the systemdistributed across the system Each knowledge base acts as a moduleEach knowledge base acts as a module Many files and databasesMany files and databases

A centralized system has all of the A centralized system has all of the knowledge stored in a single locationknowledge stored in a single location Either a file or a databaseEither a file or a database

1515

Knowledge Organization - IKnowledge Organization - I

When deciding how to group modules When deciding how to group modules one of two ways can be usedone of two ways can be used Predicate CentricPredicate Centric

All clauses of a predicates are stored in one All clauses of a predicates are stored in one modulemodule

Person CentricPerson Centric All clauses about one person or thing is stored All clauses about one person or thing is stored

in one modulein one module Rule Responder uses person centric Rule Responder uses person centric

organizationorganization example on next slideexample on next slide

1616

Knowledge Organization - IIKnowledge Organization - II

Predicate Centric: phoneOf(ben, 1-506-270-3403) phoneOf(jim, 1-506-275-9712) emailOf(ben, ben.craig@unb.ca) emailOf(jim, jim.lorde@unb.ca)

Person Centric: phoneOf(ben, 1-506-270-3403) emailOf(ben, ben.craig@unb.ca) phoneOf(jim, 1-506-275-9712) emailOf(jim, jim.lorde@unb.ca)

1717

Module BoundariesModule Boundaries

When querying modules sometimes When querying modules sometimes information from multiple modules is information from multiple modules is requiredrequired

Example Query Example Query ““What are the phone numbers of everyone What are the phone numbers of everyone

in the organization?”in the organization?” This query must backtrack across multiple This query must backtrack across multiple

modules when using person centric modules when using person centric storagestorage

1818

Centralized MaintenanceCentralized Maintenance

All knowledge is stored in a single All knowledge is stored in a single locationlocation Updating knowledge is simpleUpdating knowledge is simple Can better avoid/repair knowledge Can better avoid/repair knowledge

inconsistenciesinconsistencies All knowledge is stored in a single All knowledge is stored in a single

formatformat No translation steps when using a rule No translation steps when using a rule

engine to execute the rules and factsengine to execute the rules and facts

1919

Distributed MaintenanceDistributed Maintenance

Knowledge is stored in a many locationsKnowledge is stored in a many locations Each agent can separately update their own Each agent can separately update their own

knowledgeknowledge Knowledge bases could be incomplete or Knowledge bases could be incomplete or

inconsistentinconsistent Integrity rules can be used to test if the knowledge Integrity rules can be used to test if the knowledge

is complete and consistentis complete and consistent

Knowledge is stored in many formatsKnowledge is stored in many formats Translation steps are required when sending a Translation steps are required when sending a

query from one rule engine to anotherquery from one rule engine to another An interchange language is requiredAn interchange language is required

2020

Benchmarking Use CaseBenchmarking Use Case

RuleML-20xy SymposiaRuleML-20xy Symposia An organizational agent acts as the single An organizational agent acts as the single

point of entry to point of entry to assistassist with symposium with symposium planning:planning:

Currently, query answering about the symposium Currently, query answering about the symposium Ultimately, preparing and running the symposium Ultimately, preparing and running the symposium

Personal agents have supported Personal agents have supported symposium chairs since 2007 (deployed symposium chairs since 2007 (deployed as as Q&AQ&A in 2008) in 2008)

General Chair, Program Chair, Panel Chair, General Chair, Program Chair, Panel Chair, Publicity Chair, etcPublicity Chair, etc..

2121

Queries UsedQueries Used

1) Sponsoring the symposium1) Sponsoring the symposium

2) Check panel participants2) Check panel participants

3) View symposium sponsors3) View symposium sponsors

4) View organization partners4) View organization partners

5) 5) Check panel timeCheck panel time

2222

OO jDREW (centralized) OO jDREW (centralized) BenchmarkingBenchmarking

Query: Computation Time (ms):1) 141 2) 313) 224) 185) 16 Results show that a centralized system

does not take much computation time Queries do not require heavy computation

2323

Rule Responder (Rule Responder (Hierarchical) BenchmarkingBenchmarking

Same 5 queries used as in the OO jDREW bench Same 5 queries used as in the OO jDREW bench markingmarking

Query: Computation Time (ms):1) 3430 2) 48613) 40574) 90485) 2780 Increase in computation time due to sequential Increase in computation time due to sequential

delivery of answers to the queriesdelivery of answers to the queries Communication overhead of distributed system Communication overhead of distributed system

not compensated by workload distributionnot compensated by workload distribution

2424

Network Performance Network Performance ConsiderationsConsiderations

Speed ups can be obtained using a P2P Speed ups can be obtained using a P2P topologytopology

Instead of all communication going Instead of all communication going 'vertically' through the hub, direct 'vertically' through the hub, direct 'horizontal' communication between 'horizontal' communication between spokes could be often usedspokes could be often used Will reduce the amount of communication Will reduce the amount of communication

steps in the distributed systemsteps in the distributed system The bottleneck issue of a hierarchical The bottleneck issue of a hierarchical

system does not exist in a networked system does not exist in a networked systemsystem

2525

ConclusionConclusion

A rule system can be either distributed or A rule system can be either distributed or centralizedcentralized

When using a distributed system the When using a distributed system the communication topology must be decidedcommunication topology must be decided

The topology should reflect the modularization The topology should reflect the modularization decision about the distributed rule systemdecision about the distributed rule system

The advantages and disadvantages of distributed The advantages and disadvantages of distributed knowledge maintenance must be weighted when knowledge maintenance must be weighted when building a rule systembuilding a rule system

Our initial benchmarks, not requiring heavy Our initial benchmarks, not requiring heavy computation, show increase in computation time computation, show increase in computation time for a distributed Hierarchal systemfor a distributed Hierarchal system

Only distributed networked system will *scale* to Only distributed networked system will *scale* to the Webthe Web