DynaSoar A Scalable Architecture for High Performance AI Applications

Preview:

DESCRIPTION

DynaSoar A Scalable Architecture for High Performance AI Applications. Syed Enam-ur-Rehman Department of Computer Engineering Sir Syed University of Engineering & Technology. AI Trends & Applications. Large Simulations Behavioral / Social Modeling Control Systems Resource Management. - PowerPoint PPT Presentation

Citation preview

Syed Enam-ur-RehmanSyed Enam-ur-Rehman

Department of Computer EngineeringDepartment of Computer EngineeringSir Syed University of Engineering & TechnologySir Syed University of Engineering & Technology

DynaSoar A Scalable Architecture for High Performance AI Applications

AI Trends & ApplicationsAI Trends & Applications

Large SimulationsLarge Simulations

Behavioral / Social ModelingBehavioral / Social Modeling

Control SystemsControl Systems

Resource ManagementResource Management

Multi Agent SystemsMulti Agent Systems

Computation IntensiveComputation Intensive

Data IntensiveData Intensive

Communication IntensiveCommunication Intensive

Goals & ObjectivesGoals & Objectives

High PerformanceHigh Performance

TransparencyTransparency

Single System ImageSingle System Image

Lower cost to performance ratioLower cost to performance ratio

The ArchitectureThe Architecture

. . . INCM

DSAPI

Environment

EnvironmentNode 1

EnvironmentNode n

INCM

DSAPI

Environment

Utility Server

Management MonitoringRegistration

INCM

SK Soar KernelRA Resource AnalyzerSch SchedulerINCM Inter-Node Communication Module

DSAPI Distributed Soar API

. . .

DynaSoar Node 1

SK

Soar API

SchRA

INCM

DynaSoar Node n

SK

Soar API

SchRA

INCM

SoarSoar

Soar KernelSoar Kernel Core AI EngineCore AI Engine Agent Maintenance and ExecutionAgent Maintenance and Execution

Soar APISoar API Wrapper to Soar KernelWrapper to Soar Kernel AbstractionAbstraction

SchedulerScheduler

Predective Task SchedulingPredective Task Scheduling

Load BalancingLoad Balancing

Asymmetric SystemsAsymmetric Systems

Distributed CentralizedDistributed Centralized

Resource AnalyzerResource Analyzer

SpecificationsSpecifications

AvailabilityAvailability

Usage HistoryUsage History

INCM (Inter Node Communication Module)INCM (Inter Node Communication Module)

GatewayGateway Request DispatchRequest Dispatch Request BlockRequest Block SearchingSearching Allocation Tables & UpdatingAllocation Tables & Updating Abundant I/O ChannelsAbundant I/O Channels

DSAPI(Distributed Soar API)DSAPI(Distributed Soar API)

InterfaceInterface Utilizes INCMUtilizes INCM AbstractionAbstraction TransparencyTransparency SSISSI Agent CreationAgent Creation Agent InteractionAgent Interaction

EnvironmentEnvironment

User-spaceUser-space

Utilizes DSAPIUtilizes DSAPI

Single Entry PointSingle Entry Point

Distributed & ConcentratedDistributed & Concentrated

Utility ServerUtility Server

MonitoringMonitoring ConfigurationConfiguration RegistrationRegistration Master Allocation TableMaster Allocation Table BackupBackup

Development PhasesDevelopment Phases

Scheduling ParametersScheduling Parameters

Scheduling PolicyScheduling Policy

If TRTO = 0

XR

T EXPACT

If TRTO > 0

ACT

RTOACT R

TT

For simulation purpose:

XR

T ACTRTO

newnACTnEXP

nEXP RRR

R

2)()1(

)(

ACTEXPXPE TRT

SpeculationSpeculation

ConclusionConclusion

Separate Execution of Soar & EnvironmentSeparate Execution of Soar & Environment

Uniform gainUniform gain

Easier to implement Large AI ApplicationsEasier to implement Large AI Applications

EnquiriesEnquiries

Syed Enam-ur-Rehman (senam@ieee.org)

Usman Azeem Usmani (usman@vetolimits.com)

Nabeel Shaheen (nabeel@vetolimits.com)

Qazi Raheel Akhtar (raheel@vetolimits.com )

http://sourceforge.net/projects/dynasoar/

Recommended