Upload
rodney-martin
View
216
Download
2
Embed Size (px)
Citation preview
Survivability of Large Scale Networks and Design Research
Soundar R.T. KumaraDistinguished Professor of Industrial and
Manufacturing EngineeringThe Pennsylvania State University
University Park, PA [email protected]
NSF-EXCITED WorkshopFebruary 28, 2005
CYBER DESIGN NET(CD-NET)
PSU
VT
NASA
NIST
BU
MIT
GT
UMSU
Idaho
Univ. Agent
RepositoryAgent
Company BAgent
DesignAgent A
DesignAgent B
Web service
Company AAgent
Repository / Digital Libraries
PSU
VT
NASA
NIST
BU
MIT
GT
UMSU
Idaho
Univ. Agent
RepositoryAgent
Company BAgent
DesignAgent A
DesignAgent B
Web service
Company AAgent
Repository / Digital Libraries
CustomizationAgent
DesignAgent 1
CoordinatorAgent
RetrievalAgent
DesignAgent n
DesignAgent 3
RetrievalAgent
DesignAgent 2
DesignAgent 4
RetrievalAgent
RetrievalAgent
RetrievalAgent
CustomizationAgent
DesignAgent 1
CoordinatorAgent
RetrievalAgent
DesignAgent n
DesignAgent 3
RetrievalAgent
DesignAgent 2
DesignAgent 4
RetrievalAgent
RetrievalAgent
RetrievalAgent
Agent based Design Network
Agent based Design
NSF-ITR : An Information Management Infrastructure for Product Family Planning and Mass Customization, PI: Timothy W. Simpson (PSU), Co-PIs: Soundar R.T. Kumara (PSU),S.B. Shooter (Bucknell), J.P. Terpenny (Virginia Tech), R.B. Stone (U. Missouri-Rolla),August 2003 – July 2006
CustomizationAgent
DesignAgent 1
CoordinatorAgent
RetrievalAgent
DesignAgent n
RetrievalAgent
Customer Needs
Functional Requirements
Design Repository
NewProduct
Design Repository
Product Analysis
DesignAgent 2
RetrievalAgent
Logistics Network
Agent Based Logistics Network
General Motors: Development of Wireless based Automatic Deployment and Load Makeup System PI: Soundar R. T. Kumara (PSU). (January 2001 – current)
Sensor Networks
NSF SST : Self-Supporting Wireless Sensor Networks for In-Process and In-Service Integrity Monitoring Using High Energy-Harvesting Nonlinear Modeling Principles. PI: Soundar R. T. Kumara (PSU) Collaborators: S. Bukkapatnam (Oklahoma State), S.G. Kim (MIT) and X. Zhang (UC Berkely) (September 2004 – August 2007);
Marine Corps: Integrated Diagnostics: Soundar Kumara and Barney Grimes
Military Logistics (UltraLog)
•Secureagainst cyber attack
•Robustagainst damage
•Scalable to wartime data loads
UltraLog: Extremely survivable net-centric logistics information systems for the modern
battlefieldDARPA - ULTRALOG : Chaos, Situation Extraction, and Control: A Novel Integrated Approach to Robust and Scalable Cognitive Agent Design
PI: Soundar R. T. Kumara (PSU) (Jan. 2001 to July 2005)
UltraLog Challenges (PSU)
Situation Identification Performance Estimation Adaptive Control Hierarchical Control Robustness
Infrastructure level Application level
Network Survivability Security
Methodologies
Chaos based time series analysis, Machine learning
Digital sensors Model predictive control Auction mechanisms Mathematical optimal control Queueing theory Complex networks theory
Situation Identification
Objective: Estimate global stress environments at TAO
Methodologies: Time series analysis (Chaos), Machine learning
1
2
TAO 4
8
9
10
11
12
5 7
13
14 15
6
16 17
100%
95%
52%
62%
62%
59%
64%
64%
Adaptive Control
N1
A1
A2
A3 A4
A8
A9
A10
A12
A5A7
A11
A13 A14
A6
A15 A16
N2
N4
N3
N5
1000
200
500
500 500100
100
500
300
CPU
LP
Heuristic
Objective: Build distributed adaptive control policy for the stress environment
Control facilities: Resource allocation, Alternative algorithms
Adaptive Control
Stress Environment
Auction
Agent 1Sensor
Agent 2Sensor
Agent 3Sensor
Sensor Design
Mathematical Programming
Decentralized Coordination
Continuous Modeling
Periodic Auctioning
Methodologies: Model predictive control, Auction
DMAS Implementation: CPE Society
Military logistics Command and Control Structure Distributed, continuous planning and execution Stressful Environment: Stresses range from heavy
computational loads to infrastructure loss
Objective: Identify and demonstrate key concepts in the argument for and concept of “design for survivability
Specification and Performance Estimation
Methodology: XML based distributed specification (TechSpecs), Queueing theory based performance modeling.
Description: TechSpecs described
agent attributes, measurement points and control parameters.
BCMP network and Whitt QNA employed to estimate the end-to-end app-layer response times and remove infeasible operating modes.
Control of the DMAS
Methodology: Application-Layer control using queueing theory, and other learned models.
Description: Trading off QOS (plan quality) for performance (response time) using estimates gained from Queueing network models. Regression models used to assess the impact of model prediction on application utility.
Designing a Network Infrastructure
Hierarchical Agent Society Satisfying Constraints with Minimum Total Infrastructure Set-up Cost
Methodology: Optimization using GA.
Description: Represent the entire network of agents as a math programming model with constraints on resources with an objective to minimize the total set-up costs.
Mathematical Formulation
Load Control Problem for Agent Systems Optimal resource control to optimize long run
performance. Piecewise deterministic Markov process for
dynamic environment (workload and CPU availability)
CPU
Workload Agent
workload stress
d1
d2
)(,1 tZr
)(,2 tZr
1l
2l
)(tZDB
B
Z(t) ~ finite state, CTMC
CPU stress
))(( th x
))((2 tc l
)(1 tx
)(2 tx
d1, d2 : CPU time allocation
l1, l2 : algorithm control
Survivability: Topological perspective
Objective: Survivability of large-scale network Methodology: Complex networks theory
Cyber Design Network (CD-NET)
Challenges:
•Securityagainst cyber attacks, hackers
•Robustnessagainst damage (infrastructure and application)
•Scalability to growth and load of the network
Scalability
Sec
urity
Rob
ustn
ess
PSU
VT
NASA
NIST
BU
MIT
GT
UMSU
Idaho
Univ. Agent
RepositoryAgent
Company BAgent
DesignAgent A
DesignAgent B
Web service
Company AAgent
Repository / Digital Libraries
Scalability
Sec
urity
Rob
ustn
ess
PSU
VT
NASA
NIST
BU
MIT
GT
UMSU
Idaho
Univ. Agent
RepositoryAgent
Company BAgent
DesignAgent A
DesignAgent B
Web service
Company AAgent
Repository / Digital Libraries
PSU
VT
NASA
NIST
BU
MIT
GT
UMSU
Idaho
Univ. Agent
RepositoryAgent
Company BAgent
DesignAgent A
DesignAgent B
Web service
Company AAgent
Repository / Digital Libraries
Distributed Large Scale Networks Research- Lessons Learnt and their usefulness to CD-NET
Distributed Agents – Agent definitions, communication and platform are critical
Agent Composition to solve a problem is feasible through TechSpecs (meta-data) and dynamic service discovery
Ontologies are the foundation for TechSpecs Infrastructure Survivability – Optimization
approaches Application Survivability – Through CAS analysis