I am testing a way to speed up "high level" computations using multiprocessing. The idea is to have multiple processes (let's call them G) doing different executions of the same task T. This can be a long task (playing a whole board game) which in the end returns some results. It can be asynchronous and I know already how to gather all results using multiprocessing and apply_async.
However, at some point, T needs a call to GPU functions. My idea was to create another process S which would act as a service, gathering data from registered Gs, calling GPU function (tensorflow NN evaluation) on the gathered data (while putting Gs on halt) and "sending" back the results to all corresponding Gs.
I looked at the answers of How do I make processes able to write in an array of the main program?. However, the difference here is that the gathering does not happen only at the end of the tasks.
Do you think this is possible? I also tried different approaches using ctypes and OpenMP without success.
Here is a pseudo code of what I would like to do:
shared_service = Service()
class Worker():
def __init__(self):
shared_service.register(self)
...
def run(self):
finished = False
while (not finished):
... do my stuff ...
... gather data to "send" to GPU ...
shared_service.request(data, self.callback)
... wait for result ...
... use result ...
... do more stuff ...
def callback(self, result):
... store result ...
class Service():
...
def register(self, o):
... register new "client" ...
def request(self, o, data):
... add request to current buffer(and keep track of requester)...
def run(self):
while(True):
... wait for full buffer ...
... call GPU function ...
... dispatch results to "clients" ...
main:
... init one "Service" ...
... init N "Worker" ...
... run N Workers asynchronously ...
Thank you for your help!
