Have Super-Computers Made Parameter Estimation and Other Optimization Processes Obsolete?.

In agronomic and genetic research, estimates of crop model coefficients (C = {c₁, c₂, ..., c_n}) are often required  for thousand of treatments, potentially where C includes parameters for cultivars, soil properties or other factors. Traditional optimizations searching for the C that  minimizes error between measured and simulated data for each treatment can be inefficient and are highly dependent on optimization settings. Simulating all responses for many possible C requires vast numbers of simulation runs, but high performance computing may enable this alternative. Using resources at the Texas Advanced Computing Center (TACC), we simulated the effects of C affecting flowering date in CSM-CERES-Maize for 11 environments corresponding to trials where 5000 lines of the Maize Nested Association Mapping populations were grown in 2006 and 2007. Ex ante analysis showed that assessing 100 values for 7 coefficients was impractical. Assuming a run time of 0.013 s run^-1, the simulations would require 4.1E+09 processor-hours. If each coefficient value and the single output (anthesis date) occupies 2 bytes, the output would be 17.6 petabytes. Thus, we excluded one coefficient and reduced other coefficient ranges based on expected model sensitivity and variability among maize lines. This scenario was estimated to require 107 hours using 320 processors to complete 28.9E+06 simulations. In initial runs at TACC, efficient management of model input and output proved critical.  If queue times are not substantially impacted, performance should increase if 960 processors run simultaneously, so we currently are adjusting the simulation protocol to improve use of processors. Important adjustments have included reducing the model output to the minimum required and suppressing the model's standard output. Final assessment is pending completion of all runs, but the approach appears feasible for model applications involving large numbers of repetitive parameter adjustments due to similarities among treatment combinations.