README_parallelization.md

Parallelization

Due to the global interpreter lock Python can only run one simulation thread at time. To run multiple simulations (for example of a parameter study) in parallel multiple processes need to be used. A simulation process can be encapsulated by objects of the class SimProcess (offered by module queuesim). The constructor of SimProcess requires as parameter a Dict containing the model.

An list of SimProcess objects can be run in parallel by using the run_parallel function (also from queuesim).

from queuesim import SimProcess, run_parallel
from queuesim.models import mmc_model

# Input parameters
mean_i = 100
mean_s_range = range(70, 93)
c = 1
count = 100_000

# Generate processes
processes=[SimProcess(mmc_model(mean_i, mean_s, c, count)) for mean_s in mean_s_range]

# Run processes and wait for termination of all processes
models, simulators = run_parallel(processes)

# Get a list of the average queue lengths of the different models
ENQ = [model['Process'].statistic_queue_length.mean for model in models]

Differences between plain Python and Jupyter notebook mode

In plain Python code you will need to add a

if __name__ == '__main__':

above the simulation code to ensure it is only executed in the main process. See example_sim_mmc_series.py for this.

Limitations

To send models to sub processes the data need to be serialized. Since only simple objects (numbers, strings, etc.) can be serialized, no lambda expressions for getting inter-arrival times, service times etc. can be used in the model. When using the convenience functions from queuesim.random_dist (see Random number distributions) this is no problem. The generator functions can return lambda expressions or strings that can be evaluated to lambda expressions. If the as_lambda=True parameter is not used, they will return strings. All stations of QueueSim can handle lambda expressions as well as strings for inter-arrival times, service times etc. The strings will be evaluated to function object on first use (this means: in the sub process, after serialization and unserialization).

Example files

See example_sim_mmc_series.ipynb or example_sim_mmc_series.py for a complete example of using multiple simulation processes in parallel.