As part of the DEAP framework, EAP offers an easy DTM integration. As the EAP algorithms use a map function stored in the toolbox to spawn the individuals evaluations (by default, this is simply the traditional Python map()), the parallelization can be made very easily, by replacing the map operator in the toolbox :
from deap import dtm
tools.register("map", dtm.map)
Thereafter, ensure that your main code is in enclosed in a Python function (for instance, main), and just add the last line :
dtm.start(main)
For instance, take a look at the short version of the onemax. This is how it may be parallelized :
from deap import dtm
creator.create("FitnessMax", base.Fitness, weights=(1.0,))
creator.create("Individual", array.array, typecode='b', fitness=creator.FitnessMax)
toolbox = base.Toolbox()
# Attribute generator
toolbox.register("attr_bool", random.randint, 0, 1)
# Structure initializers
toolbox.register("individual", tools.initRepeat, creator.Individual, toolbox.attr_bool, 100)
toolbox.register("population", tools.initRepeat, list, toolbox.individual)
def evalOneMax(individual):
return sum(individual),
toolbox.register("evaluate", evalOneMax)
toolbox.register("mate", tools.cxTwoPoints)
toolbox.register("mutate", tools.mutFlipBit, indpb=0.05)
toolbox.register("select", tools.selTournament, tournsize=3)
toolbox.register("map", dtm.map)
def main():
random.seed(64)
pop = toolbox.population(n=300)
hof = tools.HallOfFame(1)
stats = tools.Statistics(lambda ind: ind.fitness.values)
stats.register("Avg", tools.mean)
stats.register("Std", tools.std)
stats.register("Min", min)
stats.register("Max", max)
algorithms.eaSimple(toolbox, pop, cxpb=0.5, mutpb=0.2, ngen=40, stats=stats, halloffame=hof)
logging.info("Best individual is %s, %s", hof[0], hof[0].fitness.values)
return pop, stats, hof
dtm.start(main) # Launch the first task
As one can see, parallelization requires almost no changes at all (an import, the selection of the distributed map and the starting instruction), even with a non-trivial program. This program can now be run on a multi-cores computer, on a small cluster or on a supercomputer, without any changes, as long as those environments provide a MPI implementation.
Note
In this specific case, the distributed version would be actually slower than the serial one, because of the extreme simplicity of the evaluation function (which takes less than 0.1 ms to execute), as the small overhead generated by the serialization, load-balancing, treatment and transfer of the tasks and the results is not balanced by a gain in the evaluation time. In more complex, real-life problems (for instance sorting networks), the benefit of a distributed version is fairly noticeable.