Many embarassingly parallel simulationns currently involve launching an ensemble of simulations, each running the same executable program on different unique initial data. In a scenario in which the initialization process involves some common calculation that is expensive, e.g., taking on the order of 30% of the run time, a worthwhile speedup might be obtained from using all of the available resources to compute the initial data rather than having each simulation using a fraction of the computational resources to repeat the common portion of hte initialization process. For example, in an ensemble of 10 simulations with each running on 100 cores, a maximum speedup of 10 for the initialization might be achievable by spreading any common initial initialization functions across all 1000 available cores before partitioning the computation into 10 independent simulations. This could require expensive input/output if the jobs are truly independent processes launched by some a script.
This repository will test the approach of replacing the launch scripts with a Fortran main program that performs the initialization and then launches each individual simulation simultaneously using the Fortran 2015 team feature.