The Ergodicity Question- Time averaging vs Ensemble averaging

11am – 3pm

  • Tried time averaging over the entire time window (set sample_size to 1 less than max time); in close agreement with the ensemble predictions (getting around 1355).
  • Mystery doesn’t seem to actually be a problem in the method, but just innate variability! Even over reasonable time-averages and ensemble averages there’s significant deviation from the expected variance! This variation around the warning signal/leading indicator is exactly what might frustrate the warning signal. So within a run, the observed variance averaged over a given window may fluctuate significantly.
image
image
image
image
  • Note that variance almost doubles just by chance, and the average looks pretty close to the analytic calculation for these parameters (1352). This samples every unit time on a window of 149 time units (single replicate).
  • Code that produces this found in version 26 (run warning_signals and then plot data with that version of plot.py). Note: seed was randomized in this compile.
  • Gillespie header file has a flag to turn on random seed!

Next Steps

  • Parallelization – restructure my_pars and record classes.
  • Compute variance in variance (and in other warning signals) across ensembles. Shows not just the expected signal, but the collection of possible signals. Think about analytic version of this calculation.
  • Update records class for these kinds of ensemble averaging. Ensemble Variance in temporal mean with windowsize 1 will then give the variance across ensembles alone, rather than calculating separately (which I’m currently using the ar1 and arN blocks to compute).


### Parallelizing Code

  • Large ensemble averaging can benefit from relatively trivial parallelization. Parallelizing other sections of the code will be more demanding and less obviously an efficiency boost. I have been doing basic parallelization for multicore, shared memory machines using OpenMP, allowing me to take advantage of the up to eight cores we have on the server (one.)
  • Makefile now has a make parallel option to call the compiler with the openmp flag, -f openmp
  • My first try added an OpenMP command to the Gillespie code to parallelize the for loop over ensembles:

  1. pragma omp parallel for default(none) shared(rng, rate_fn, outcome) private(lambda, t, tmp, i, check) firstprivate(rates_data, my_pars)

for(l = 0; l < ensembles; l++){

I haven’t worked with firstprivate before, but seems necessary for dynamically allocated private variables. This isn’t going to work, as my_pars contains information that must be shared (like where the global data is collected) and private (like the state variables). A successful parallelization of this algorithm will require being more explicit. Basically everything in record needs to be shared and the rest of my_pars needs to be private…

As the random number generator is called in a different order in parallel and serial runs I don’t think I can compare values simply by keeping constant seed, but I can compare convergent values.

  • Will need to separate my parameters data structure into global recording data and private state variable data before the Gillespie code can be effectively parallelized. Will need to consider checks to prevent user from mismatching these.

edit on Fri 26th

  • Anything declared private in openmp must be initialized in the loop or declared firstprivate instead. If it’s dynamically allocated compiler will complain about private but not firstprivate. Making a copy of a constant pointer (as in promotions from void pointers) lets you change the values of things it points to.

  • Parallel code not very compatible with GSL’s creative use of void pointers as agnostic ways to pass parameters to functions. Fine for shared parameters, creates a problem for private parameters. Using restrict declaration might help with this, will need to investigate further. Meanwhile, experimental directory has simplified code for exploring the parallel error. currently cannot get time series to cumulatively sum results from different replicates rather than overwrite.

edits Sun Feb 28

  • Parallel code works if I can avoid firstprivate. this means specifying predetermined size arrays for checkpts and my_pars explicitly. Will have to figure out how to get openmp to work with dynamically allocated variables. Time to actually spend some time in openmp tutorials…