"""A way over-simplified SRUMMA matrix multiplication implementation.
Assumes square matrices with the shape as a multiple of the block size.
"""
import mpi4py.MPI
import ga
import numpy as np
CHUNK_SIZE = 256
MULTIPLIER = 3
N = CHUNK_SIZE*MULTIPLIER
me = ga.nodeid()
nproc = ga.nnodes()
class Task(object):
def __init__(self, alo, ahi, blo, bhi, clo, chi):
self.alo = alo
self.ahi = ahi
self.blo = blo
self.bhi = bhi
self.clo = clo
self.chi = chi
def __repr__(self):
return "Task(%s,%s,%s,%s,%s,%s)" % (
self.alo, self.ahi, self.blo, self.bhi, self.clo, self.chi)
def get_task_list(chunk_size, multiplier):
count = 0
task_list = [None]*multiplier**3
import mpi4py.MPI
import ga
import numpy as np
me = ga.nodeid()
nproc = ga.nnodes()
def parallel_task():
me = ga.pgroup_nodeid()
nproc = ga.pgroup_nnodes()
### print a message from the master of the group
g_a = ga.create(ga.C_DBL, (3, 4, 5))
ga.randomize(g_a)
### sum the g_a and print the sum
### -OR- do something else with g_a...
midproc = nproc // 2
### assign to 'proclist_first' the first half of the process range
### assign to 'proclist_last' the last half of the process range
### create the 'group_id_first' process group
### create the 'group_id_last' process group
if me in proclist_first:
### set the default group to 'group_id_first'
parallel_task()
### reset the default group to the world group
### synchronize
if me in proclist_last:
### set the default group to 'group_id_last'
parallel_task()
def __init__(self):
self.me = ga.nodeid()
self.stdout = sys.stdout
"""Use ga.access() to sum locally per SMP node."""
import mpi4py.MPI
import ga
import numpy as np
world_id = ga.nodeid()
world_nproc = ga.nnodes()
node_id = ga.cluster_nodeid()
node_nproc = ga.cluster_nprocs(node_id)
node_me = ga.cluster_procid(node_id,ga.nodeid())
g_a = ga.create(ga.C_DBL, (3,4,5,6))
if world_id == 0:
ga.put(g_a, np.arange(3*4*5*6))
ga.sync()
if node_me == 0:
sum = 0
for i in range(node_nproc):
smp_neighbor_world_id = ga.cluster_procid(node_id,i)
buffer = ga.access(g_a, proc=smp_neighbor_world_id)
sum += np.sum(buffer)
print sum
def __init__(self):
self.me = ga.nodeid()
self.stdout = sys.stdout
"""Use ga.access() to sum locally per SMP node."""
import mpi4py.MPI
import ga
import numpy as np
world_id = ga.nodeid()
world_nproc = ga.nnodes()
node_id = ga.cluster_nodeid()
node_nproc = ga.cluster_nprocs(node_id)
node_me = ga.cluster_procid(node_id, ga.nodeid())
g_a = ga.create(ga.C_DBL, (3, 4, 5, 6))
if world_id == 0:
ga.put(g_a, np.arange(3 * 4 * 5 * 6))
ga.sync()
if node_me == 0:
sum = 0
for i in range(node_nproc):
smp_neighbor_world_id = ga.cluster_procid(node_id, i)
buffer = ga.access(g_a, proc=smp_neighbor_world_id)
sum += np.sum(buffer)
print sum
return n
def comp_pi(n, myrank=0, nprocs=1):
h = 1.0 / n;
s = 0.0;
for i in xrange(myrank + 1, n + 1, nprocs):
x = h * (i - 0.5);
s += 4.0 / (1.0 + x**2);
return s * h
def prn_pi(pi, PI):
message = "pi is approximately %.16f, error is %.16f"
print (message % (pi, abs(pi - PI)))
nprocs = ga.nnodes()
myrank = ga.nodeid()
g_pi = ga.create(ga.C_DBL, [1])
one_time = False
if len(sys.argv) == 2:
n = int(sys.argv[1])
one_time = True
while True:
if not one_time:
if myrank == 0:
n = get_n()
n = ga.brdcst(n)
else:
n = ga.brdcst(0)
import mpi4py.MPI # initialize Message Passing Interface import ga # initialize Global Arrays print "hello from %s out of %s" % (ga.nodeid(),ga.nnodes())