1. Modeling Grid Based Computer Services for Public Health Applications
Martin Cryer
Catherine Staes, BSN, MPH, PhD
Lewis Frey, PhD

2. Public Health Applications
Pressure to adopt Grid platforms for Public Health applications…
Porting legacy applications
Taking advantage of newly available Silos
Developing native Grid applications
Developing Native Grid PH Applications
Enhancing Legacy PH Applications
Porting Legacy PH Applications
New Silos More Data

3. Adoption Risks
Significant risks in migrating legacy applications to Grid:
Performance bottlenecks
Impact on legacy applications
Impact on other Grid users
Remote error handling for large queries
Design decisions made without sufficient experience
Early adopters have steep learning curve 3

4. Consequences of Adoption Risks
Failure to handle risks can lead to:
Projects voting with their feet
Withdraw from Grid project when own application is impacted
Removed from Grid project when Grid is impacted
Loss of credibility for Grid provider
Loss of credibility for project owners
Reduced cooperative opportunities
Squandering the benefits of a Grid
Widespread financial risks 4

5. Financial Risk Large cost overruns
$100m’s at risk, even more for cross institution developments
Viability of projects threatened
Failures in porting legacy applications
Cost of fixes and upgrades
Cost of lost production
Cost to the user community
Threat to future Grid projects

6. Modeling and Simulation
Creating a model and running simulations
Build a formal model of the system
Run simulations for what-if analysis
Output drives decision making process
Run lots of iterations for most-likely outcome
Doesn’t replace experience, validates designs
We need to predict…
System performance
Availability
Legacy application errors – Query timeouts
The effects of new applications and components on the Grid

7. Example of Application Migration
Model a production Grid, NIH caGRID
Integrate an NCI SEER system
Incorporate enhanced application functionality
Simulation predicts impact on combined system

8. Getting Started: Combined Model
Calibrated model of NIH caGRID Network
Use subset of Grid
caTISSUE / caARRAY services
Calibrated model of NCI SEER System
Subset of services
Integrate SEER and caGRID models
Run simulations on integrated model
NCI SEER System(s) Model
caTISSUE caARRAY NIH caGRID Workflows Model
Calibration
Calibration
NCI SEER Calibrated Model
NIH caGRID Calibrated Model
NIH caGRID Predictive Model Incorporating NCI SEER Functionality 8

9. Modeling Techniques Three main modeling techniques…
Continuous / Numeric Simulation
Continuous process: soln. differential equations
Discrete Event Simulation
Queuing theory, event driven
Agent Based Modeling (ABM)
Agents derived from Von Neumann, Cellular Automata
Agents represent functional components of system
Program the interactions and behaviors of agents
Agents interact with other agents and environment
Observe the outcome
Repeat for many iterations for consensus

10. Agent Based Modeling Purpose of simulation: Types of simulation:
Analysis validation? – Calibrates regardless of model?
Design verification - Calibrates as a f(model variations)
Types of simulation:
Trace driven? - Ordered set of trace data / model outcome?
Stochastic methods? - Driven by probability distributions?
Combined - Trace driven simulation of load used within a randomized model
Agents represent:
Process orientated (process or threads) model?
Distributed (message passing) model?
Combination of process orientated and distributed models at use-case not process level

11. Using ABM for Simulating the Grid
Use-case modeling
Simplification
Node agents
Model node functionality
Network agents
Model network functionality
Workflow agents
Define each use-case
Run simulations and collect results
Workflow “A” Agent
Workflow “B” Agent
Node Agents
Workflo w “B” Agent
Workflo w “A” Agent
Workflows “A” and “B” Interact w/ Node Agents

12. Simulation Output Model is XML document driven
XJ Technology’s AnyLogic™ software used to define model
Output displayed via a Python based analysis and graphing application
Combines repeated simulation run outputs to provide most likely outcome
matplotlib()

13. Early Results Modeled basic components of node and network components
Simplified agent model without SEER, low level building blocks only, simplified Grid
Model scales as expected against real world; with concurrent load…
Response time for users (system latency) increases
Transaction rate (TPS) decays
Future model under development is more sophisticated

14. Further Investigations
Additional functionality
Model to reflect operational costs
What-if simulation of Reliability, Availability, Serviceability, Usability and Maintainability (RASUM) variations
Failure domain modeling
Incorporation of real time sensor components to Grid, Query Timeout issues
Addition of “real-time” surveillance data
Real time model feedback
Workflow scheduler
Grid services
System Administrator (SA) training

15. Conclusions
Migration of legacy systems to Grid environments requires prediction of…
Sizing of required systems
Impact of system on Grid and Grid on system
Planning of new applications
Cost impact
Use-case simulations using Agent Based Modeling appears suitable for modeling of a Grid based environment
Modeling Grid based systems allows for a reduction in risk

16. Acknowledgments Contact Support provided by Acknowledgments
Martin Cryer
Univ. of Utah, School of Medicine
Support provided by
National Library of Medicine (NLM) Training Grant No. 1T15LM
Acknowledgments
Dr. Lewis Frey (PhD Committee Chair)
Univ. of Utah P.I. caBIG™
Dr. Catherine Staes (PhD Committee Member)
Univ. of Utah CoE in Public Health Informatics
Dr. Alun Thomas
Suggestions regarding concurrent access latency calculations

17. Questions 17