Pipes are getting stuck--no other solution on stack overflow working

Question

(UPDATED) I am building a module to distribute agent based models, the idea is to split the model over multiple processes and then when the agents reach a boundary they are passed to the processor handling that region. I can get the processes set up and working with no communication, but cannot get the data to pass through the pipes and update the model segment on the other processor.

I have tried the solutions on stackoverflow and built a simple version of the model. As soon as I put in a model object into the pipe the model hangs (it works with python standard data types). The simple version just passes agents back and forth.

 from pathos.multiprocessing import ProcessPool
 from pathos.helpers import mp
 import copy



 class TestAgent:

 "Agent Class-- Schedule iterates through each agent and \
  executes step function"

 def __init__(self, unique_id, model):
      self.unique_id = unique_id
      self.model = model
      self.type = "agent"

 def step(self):

       pass 
       #print ('     ', self.unique_id, "I have stepped")

class TestModel:

   "Model Class iterates through schedule and executes step function for \
   each agent"

   def __init__(self):
       self.schedule = []
       self.pipe = None
       self.process = None

       for i in range(1000):
           a = TestAgent(i, self)
           self.schedule.append(a)


   def step(self):


       for a in self.schedule:
           a.step()

if __name__ == '__main__':

   pool = ProcessPool(nodes=2)

   #create instance of model
   test_model = TestModel()
   #create copies of model to be run on 2 processors
   test1 = copy.deepcopy(test_model)
   #clear schedule
   test1.schedule = []
   #Put in only half the schedule
   for i in range(0,500):
       test1.schedule.append(test_model.schedule[i])  
   #Give process tracker number
   test1.process = 1
   #repeat for other processor
   test2= copy.deepcopy(test_model)
   test2.schedule = []
   for i in range(500,1000):
       test2.schedule.append(test_model.schedule[i])
   test2.process = 2

   #create pipe
   end1, end2 = mp.Pipe()

   #Main run function for each process
   def run(model, pipe):

      for i in range(5):
          print (model.process)#, [a.unique_id for a in model.schedule])
          model.step() # IT HANGS AFTER INITIAL STEP
          print ("send")
          pipe.send(model.schedule)
          print ("closed")
          sched = pipe.recv()
          print ("received")
          model.schedule = sched



   pool.map(run, [test1, test2], [end1,end2])

The agents should switch processors and execute their print functions. (My next problem will be synchronizing the processors so they stay on each step, but one thing at a time.)

Answer 1

I got it to work. I was exceeding the pipe buffer limit in python (8192). This is particularly true if the agent holds a copy of the model as an attribute. A working version of the above code, which passes the agents one at a time is below. It uses Pympler to get the size of all the agents.

from pathos.multiprocessing import ProcessPool
from pathos.helpers import mp
import copy

# do a blocking map on the chosen function

class TestAgent:

"Agent Class-- Schedule iterates through each agent and \
executes step function"

   def __init__(self, unique_id, model):
       self.unique_id = unique_id
       self.type = "agent"

   def step(self):
       pass 


class TestModel:

   "Model Class iterates through schedule and executes step function for \
   each agent"

   def __init__(self):
       from pympler import asizeof 

       self.schedule = []
       self.pipe = None
       self.process = None
       self.size = asizeof.asizeof


       for i in range(1000):
           a = TestAgent(i, self)
           self.schedule.append(a)


   def step(self):


       for a in self.schedule:
           a.step()

if __name__ == '__main__':

   pool = ProcessPool(nodes=2)

   #create instance of model
   test_model = TestModel()
   #create copies of model to be run on 2 processors
   test1 = copy.deepcopy(test_model)
   #clear schedule
   test1.schedule = []
   #Put in only half the schedule
   for i in range(0,500):
       test1.schedule.append(test_model.schedule[i])  
   #Give process tracker number
   test1.process = 1
   #repeat for other processor
   test2= copy.deepcopy(test_model)
   test2.schedule = []
   for i in range(500,1000):
       test2.schedule.append(test_model.schedule[i])
   test2.process = 2

   #create pipe
   end1, end2 = mp.Pipe()

   #Main run function for each process
   def run(model, pipe):

      for i in range(5):
        agents = []
        print (model.process, model.size(model.schedule) ) 
        model.step() # IT HANGS AFTER INITIAL STEP
        #agent_num = list(model.schedule._agents.keys())
        for agent in model.schedule[:]:
            model.schedule.remove(agent)
            pipe.send(agent)
            agent = pipe.recv()
            agents.append(agent)
        print (model.process, "all agents received")
        for agent in agents: 
            model.schedule.append(agent)

        print (model.process, len(model.schedule))



   pool.map(run, [test1, test2], [end1,end2])

Mike McKerns and Thomas Moreau --Thanks for the help you put me on the right path.