Transparent Grid Enablement of WRF Using a Profiling, Code Inspection, and Modeling Approach

1
Transparent Grid Enablement of WRF Using a
Profiling, Code Inspection, and Modeling Approach

• S. Masoud Sadjadi, Hugh Willoughby, Javier
  Figueroa, Javier Delgado, Xabriel Collazo, Javier
  Munoz, Diego Lopez, and Selim Kalayci

Presented by Selim Kalayci Florida International
University 11/14/2007
2
Motivation

• Impact of Hurricanes
• Providing accurate and timely information is
  essential for effective planning and response
• WRF model is the latest developed by NCAR
• Adapted by meteorological services in US and
  worldwide
• High requirements of WRF
• Computing nodes with high volume of memory and
  storage
• High-speed network connection

3
Motivation (2)

• Current Approach for executing WRF Models
• Single machine
• Cluster of homogeneous nodes
• Proposed Approach
• Scale out WRF execution to Grid environments

4
Goals

• Enabling WRF to scale out to Grid computing
  environments
• Modeling WRF behavior and its resource
  requirements
• Estimating the runtime
• Predicting the allocation of resources

5
Challenges

• The high latency of Internet compared to
  high-speed LANs
• The high overhead of the Grid middleware software
• Risking compatibility with future WRF versions
• The high volume of the WRF sources code
• Compiling WRF on unsupported platforms

6
Modeling WRF Behavior
Mathematical Modeling
Profiling
Code Inspection Modeling
Parameter Estimation
7
Mathematical Modeling

• Assumptions
• resource consumption consists of any forms of
  static resource parameters such as CPU cycles
• parallelism is independent of the static resource
  parameters
• Generalized Model

• (developed by Shu Shimizu)
• The expression is finally transformed into a
  linear summation of profile parameters (?s),
  which are estimated as an applications
  characteristic.

8
Monitoring and Profiling

• Monitoring Software (amon) - (developed by Shu
  Shimizu)
• runs as a daemon application in each node
• gathers information from executed processes
• sends this data to the profiling software
• Profiling Software (aprof ) - (developed by Shu
  Shimizu)
• runs as a server application in a single node
• receives the reports from the monitoring programs
  
• Uses the mathematical model mentioned before for
  predictions
• Automation Scripts
• Automatic execution of WRF model using different
  number of nodes and processors power percentages
• Automatic formatting of results from WRF
  forecasts or simulations results for use with
  profiling software
• Automatic data gathering and graph generation
  based on prediction results from profiling
  software

9
Parameter Estimation

• Application profiles are derived by executing the
  application on different platforms with varied
  configuration of available resources.
• Regression analysis is then used to fit the data
  into a linear model.
• As more observations are made, the accuracy of
  the model generated improves.
• This model is then subsequently used for
  predicting application execution and resource
  usage on previously unseen platforms.

10
Current Model and Results

• Simplified Model using Inverse Number of CPUs and
  Inverse CPU clock speed
• Execution Results for 2-8 nodes with the above
  model

Texe ( ?0 ?1 / nodes ) ( ?0 ?1 / clock )
- (Shu Shimizu)
11
Code Inspection Modeling

• We use code inspection and modelling to justify
  why WRF behaves as it does
• We provide feedback to the mathematical modelling
  
• may result in adding or removing parameters
• may result in reflecting the dependencies of two
  or more parameters

12
Summary

• Scaling out WRF to the Grid environment
• Should be transparent and compensate the overhead
• We got a simple mathematical model for our
  current simulations, which fits quite well
• Future work
• Execute WRF on a real Grid environment (more than
  two clusters)
• Using profiling, estimate other parameters
• Improve the mathematical model with these
  parameters

13
Acknowledgement (1)

• David Villegas1, Raju Rangaswami1, Shu Shimizu4,
  Hector A. Duran Limon5, Rosa Badia2, Pat Welsh3,
  Jason Liu1, Alex Orta1, and Michael McFail1
• 1 Florida International University, Miami, FL,
  USA
• 2 Barcelona Supercomputing Center, Barcelona,
  Spain
• 3 University of North Florida, Jacksonville, FL,
  USA
• 4 IBM Tokyo Research Laboratory, Tokyo, Japan
• 5 University of Guadalajara, Mexico

14
Acknowledgement (2)

• This work is supported by
• The National Science Foundation grant OCI-0636031
  and REU-0552555.
• IBM (SUR and Student Support awards).
• We would like to thank
• The CIARA staff and engineers support Heidi
  Alvarez, Julio Ibarra, Ernesto Rubi, Diego Obina,
  Ileana Gonzalez, and Omaida Hennessey.
• The SCIS system staff support Steve Luis, Eric
  Johnson, Catherine Hernandez, and Chak Leung.

15
QUESTIONS

• Thank you!
• QUESTIONS?