High Performance Computing Solution to Pathway Parameter Estimation

1
High Performance Computing Solution to Pathway
Parameter Estimation

• Peter Henning
• April 25, 2005

2
Overview

• 1) Reliability
• 2) Customization
• 3) Performance
• 4) Issues and Concerns

3
Reliability 1
4
Reliability 2
5
Reliability 3 52Node
6
Customization

• Eliminate Non Physical Solutions
• Is Continuous Minimization Possible?
• Can Solutions Be Constrained Using Monte Carlo
  Techniques to reduce parameter space?
• Enable the Construction of Hybrid Algorithms
• Conjugate Gradient Methods
• Downhill Simplex
• Davidon-Flecther-Powell (DFP)

7
Performance 1
8
Performance 2
9
Performance 3
10
Monte Carlo 6 Node Trials

• 160 Million Trials in 27 Hours
• 3.25 of trials produced viable solutions
• 22 Trials with zero error, 42 with error less
  than 20

11
Monte Carlo 52 Node Trials

• 34 Million Trials in 16 hours
• Produced only 2 viable solutions
• Successful Trials were not exceedingly less than
  arbitrary guesses such as making all parameters
  the same.

12
Genetic Algorithm 6 Node Trials

• Execution Time varies widely likely due to
  increased file writing. The fastest process
  finished 1 Million Generations of 1000 in 12.6
  Hours. Most trials were much slower.
• Significant Improvements on the parameters of the
  Monte Carlo Trials were not found in this case.

13
Issues and Concerns

• File Writing is now the bottleneck of the
  process. Writing to the NFS drive is the major
  source of network traffic.
• A master process responsible for generating
  statistics and all file writing could potentially
  improve this process.
• More investigation into what causes a
  non-physical solution and if they have to be
  considered in optimization schemes to get to
  physical minima.