FriesischScott
Saving the state of subset simulation
Often, when running subset simulation on HPC, the algorithm can get quite far (many levels) and then can fail due to a simulation crashing.
After fixing, we need to restart, and this is resource inefficient. It would be useful if we could save the state of the algorithm at every (successful) iteration, and reload this state if it crashes and we need to continue.
Perhaps all we need to do is save the samples and performance DataFrames in the main loop?
This is also useful for initialising the algorithm, as all the subset methods begin with a Monte Carlo sample.
This might be something worth having for all simulations. Running a large MC simulation with SLURM could for example also benefit from a restart with data if something goes wrong. Requires some thought for a good system.
Yes, ideally in the most non-intrusive way possible.
One suggestion (but quite intrusive) is to replace the for loop with while i <= sim.levels, and save the i, samples and performance (and sim?) at each iteration, with a user option of passing a saved state to the probability_of_failure function.
Yes, ideally in the most non-intrusive way possible.
One suggestion (but quite intrusive) is to replace the for loop with
while i <= sim.levels, and save thei,samplesandperformance(andsim?) at each iteration, with a user option of passing a saved state to theprobability_of_failurefunction.
I'm getting Vietnam flashbacks of the SimulationData in Cossan 🙈