Just been curious about parallel computation. Clusters. Gives me a little nerd hard-on.
Working my way up to running some stuff on AWS (Amazon Web Services).
So I’ve been goofing around with mpi. Mpi (message passing interface) is sort of an instant messager for programs to pass around data . It’s got some convenient functions but its mostly pretty low level.
I’ll jot some fast and incomplete notes and examples
Tried to install mpi4py.
sudo pip install mpi4py
but it failed, first had to install openmpi
To install on Mac I had to follow these instructions here. Took about 10 minutes to compile
give this code a run
#mpirun -np 3 python helloworld.py from mpi4py import MPI comm = MPI.COMM_WORLD rank = comm.Get_rank() size = comm.Get_size() name = MPI.Get_processor_name() print "Hello. This is rank " + str(rank) + " of " + str(size) + " on processor " + name`
the command mpirun runs a couple instances. You know which instance you are by checking the rank number which in this case is 0 through 2.
Typically rank 0 is some kind of master
lower case methods in mpi4py work kind of like how you’d expect. You can communicate between with comm.send and comm.recv
#mpirun -np 2 python helloworld.py from mpi4py import MPI comm = MPI.COMM_WORLD rank = comm.Get_rank() size = comm.Get_size() name = MPI.Get_processor_name() if rank == 0: comm.send("fred",dest=1) else: counter = comm.recv(source=0) print counter
However I think the these are toy methods. Apparently they use pickle (python’s fast and dirty file storage library) in the background. On the other hand, maybe since you’re writing in python anyhow, you don’t need the ultimate in performance and just want things to be easy. On the third hand, why are you doing parallel programming if you want things to be easy? On the fourth hand, maybe you
The capital letter mpi functions are the ones that are better, but they are not pythony. They are direct translations of the C api which uses no returns values. Instead you pass along pointers to the variables you want to be filled.
from mpi4py import MPI import numpy as np comm = MPI.COMM_WORLD rank = comm.Get_rank() size = comm.Get_size() name = MPI.Get_processor_name() nprank = np.array(float(rank)) result = np.zeros(1) comm.Reduce(nprank, result, op=MPI.SUM, root=0) if rank == 0: print result