[zeromq-dev] pyzmq and multiprocessing

[Jon Dyte]

On 22/12/11 15:03, Ivo Danihelka wrote:
> On Thu, Dec 22, 2011 at 2:48 PM, Wiliam Souza<wiliam at mandriva.com.br>  wrote:
>> [1] http://pastebin.mandriva.com/38061
> I don't get the assert on zmq 2.1.4.
> The ControllerServer process does not receive any msg for me.
> I guess, multiprocessing is not compatible with zmq threads.
> The example worked, when I moved the zmq initialization inside run().
>
> I don't know multiprocessing internals. I would avoid mixing it with zeromq.
>
That's correct. It's not going to work creating the zmq Context in the 
__init__ method of the Process object.

On unix Python Multiprocess works by forking the script for each Process 
object and communicating over some IPC.
The main script coordinates all the other 'Process' objects.

The problem in the example is that the 0MQ context's are created before 
the fork, in the __init__ method.
These contexts create zmq threads in the background.

Then when the 'start' methods are executed on each controller, the 
forking happens. Threads don't generally survive fork
and thus the process is in an undefined state.

If you put a massive sleep statement ie (time.sleep(86400)) at the 
beginning of each start method, you
can clearly see using ps you will see something like

jon      24938 18475  0 17:50 pts/1    00:00:00 python /home/jon/mp.py
jon      24943 24938  0 17:50 pts/1    00:00:00 python /home/jon/mp.py
jon      24944 24938  0 17:51 pts/1    00:00:00 python /home/jon/mp.py

24938 is the actually executed script
24943 and 24944 are the two children executing a Process object each.

However if you look under the /proc/<pid>/tasks

you will see
24938 has 5 threads
ls /proc/24938/task
24938  24939  24940  24941  24942
These are the 'main' thread and two threads per 0MQ context
the child  processes both have one thread and hence 0MQ isn't going to 
work as they have no zmq threads!

ls /proc/24943/task
24943

ls /proc/24944/task
24944

If you want the script as written to have a chance of working, the 0MQ 
context and sockets need to be created in the run method,
or at least after the fork.

Jon