In [ ]:
import numpy as np
In [1]:
import ipyparallel as ipp
In [ ]:
# c = ipp.Client(profile='mpi',sshserver='marcin.kostur@155.158.128.149')
In [ ]:
# c = ipp.Client(profile='default',sshserver='marcin.kostur@155.158.128.148')
In [ ]:
#c = ipp.Client(profile='mpi',sshserver='marcin.kostur@155.158.128.98')
In [ ]:
c = ipp.Client(profile='mpi')
In [4]:
c = ipp.Client(profile='gpu4')
In [5]:
c.ids
Out[5]:
In [6]:
view = c[:]
In [7]:
view.activate()
In [8]:
%%px --block
import socket
print(socket.gethostbyname(socket.gethostname()))
In [9]:
%%px --block
import random
a=random.random()
In [10]:
view['a']
Out[10]:
In [11]:
view['a'] = 2.0
In [ ]:
%%px --block
import sys
sys.version
In [16]:
%%px --block
import sys
print(sys.version)
In [ ]:
In [ ]:
In [ ]:
import socket
socket.gethostbyname(socket.gethostname())
In [ ]:
In [12]:
%%px --block
import numpy as np
import pycuda.driver as cuda
import pycuda.autoinit
from pycuda.compiler import SourceModule
mod = SourceModule("""
__global__ void kernel(float *a)
{
int idx = threadIdx.x + blockDim.x*blockIdx.x;
if(threadIdx.x>2)
a[idx] = 444.0f;
}
""")
a = np.zeros(10).astype(np.float32)
func = mod.get_function("kernel")
print (np.linspace(0,9,10))
print ("----------------")
print (a)
func(cuda.InOut(a),block=(5,1,1),grid=(2,1,1))
print (a)
In [13]:
view['a'][0]
Out[13]:
In [ ]:
%%px --block
import numpy as np
import pycuda.driver as cuda
import pycuda.autoinit
from pycuda.compiler import SourceModule
import pycuda.gpuarray as gpuarray
from pycuda.curandom import rand as curand
cuda.init()
device = cuda.Device(0)
ctx = device.make_context()
Nx = 1000
Na = 1000
a = np.linspace(3.255,4,Na).astype(np.float32)
a = np.repeat(a,Nx)
a_gpu = gpuarray.to_gpu(a)
x_gpu = curand((Na*Nx,))
from pycuda.elementwise import ElementwiseKernel
iterate = ElementwiseKernel(
"float *a, float *x",
"x[i] = a[i]*x[i]*(1.0f-x[i])",
"iterate")
Niter = 10000
for i in range(Niter):
iterate(a_gpu,x_gpu)
ctx.synchronize()
print (Niter*Na*Nx/1e9)
a,x = a_gpu.get(),x_gpu.get()
In [ ]:
%matplotlib inline
import matplotlib.pyplot as plt
a=view['a'][0]
x=view['x'][0]
In [ ]:
plt.figure(num=1, figsize=(30, 20))
every = 1
plt.plot(a[::every],x[::every],'ob',markersize=1,alpha=0.1)
plt.plot([3.83,3.83],[0,1])
In [ ]:
In [ ]:
In [ ]:
In [ ]:
view.block = True
In [ ]:
%autopx
In [ ]:
import socket
print(socket.gethostbyname(socket.gethostname()))
import sys
sys.version
In [ ]:
In [ ]:
import pycuda.gpuarray as gpuarray
from pycuda.curandom import rand as curand
from pycuda.compiler import SourceModule
import pycuda.driver as cuda
try:
ctx.pop()
ctx.detach()
except:
print ("No CTX!")
cuda.init()
device = cuda.Device(0)
ctx = device.make_context()
print (device.name(), device.compute_capability(),device.total_memory()/1024.**3,"GB")
print ("a tak wogóle to mamy tu:",cuda.Device.count(), " urządzenia")
In [ ]:
%px from ipyparallel import bind_kernel; bind_kernel()
In [ ]:
import sys
sys.version
In [ ]:
%matplotlib inline
In [ ]:
%%px --block
%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np
plt.plot(np.random.randn(1000),np.random.randn(1000),'yo')
In [ ]:
%%px --block
%%sh
pwd
In [ ]:
%%sh
pwd
In [ ]:
c = ipp.Client(profile='default')
In [ ]:
In [ ]:
c.ids
In [ ]:
dview = c[:]
activate - make magics use the curent view
In [ ]:
%%px
%matplotlib inline
from pylab import *
plot([1,2,9],[2,9,3],'o-')
In [ ]:
%pxresult
In [ ]:
In [ ]:
In [ ]:
@dview.remote(block=True)
def getpid():
import os
return os.getpid()
In [ ]:
getpid()
In [ ]:
@dview.remote(block=True)
def iterate_logistic(a,N,Niter):
import numpy as np
x = np.random.random(N)
for i in range(Niter):
x = a*x*(1.0-x)
return x
In [ ]:
In [ ]:
iterate_logistic(4.0,1,10000)
In [ ]:
In [ ]:
import numpy as np
x = np.random.random((2**12))
print(x.nbytes/1024**2,x.size)
a = np.ones_like(x)*4.0
In [ ]:
@dview.parallel(block=True)
def iterate_logistic(x,a):
import numpy as np
x = np.copy(x)
for i in range(100000):
x = a*x*(1.0-x)
return x
In [ ]:
%%time
for i in range(100000):
x = a*x*(1.0-x)
In [ ]:
%time x = iterate_logistic(x,a)
In [ ]:
In [ ]:
In [ ]:
dview?
In [ ]:
import ipyparallel.client.remotefunction
ipyparallel.client.remotefunction.ParallelFunction?
In [ ]:
In [ ]:
In [ ]:
In [ ]:
Configure mpi profile first:
In [ ]:
In [ ]:
c = ipp.Client(profile='mpi')
In [ ]:
c.ids
In [ ]:
c[:].apply_sync(lambda : "Hello, World")
Interesting might be:
c = Client('/path/to/my/ipcontroller-client.json', sshserver='me@myhub.example.com')
In [ ]:
import mpi4py
In [ ]:
%%px
In [ ]:
%%writefile psum.py
from mpi4py import MPI
import numpy as np
def psum(a):
locsum = np.sum(a)
rcvBuf = np.array(0.0,'d')
MPI.COMM_WORLD.Allreduce([locsum, MPI.DOUBLE],
[rcvBuf, MPI.DOUBLE],
op=MPI.SUM)
return rcvBuf
In [ ]:
view = c[:]
In [ ]:
view.activate()
In [ ]:
view.run('psum.py')
In [ ]:
view.scatter('a',np.arange(16,dtype='float'))
In [ ]:
view['a']
In [ ]:
%px totalsum = psum(a)
In [ ]:
view['totalsum']
In [ ]:
In [ ]:
%%writefile psum.py
from mpi4py import MPI
import numpy as np
import time
def psum(a):
locsum = np.sum(a)
rcvBuf = np.array(0.0,'d')
MPI.COMM_WORLD.Allreduce([locsum, MPI.DOUBLE],
[rcvBuf, MPI.DOUBLE],
op=MPI.SUM)
return rcvBuf
rank = MPI.COMM_WORLD.Get_rank()
size = MPI.COMM_WORLD.Get_size()
if rank==0:
a = np.arange(16,dtype='float')
else:
a = None
a_local = np.empty(16/size,dtype='float')
MPI.COMM_WORLD.Scatter( a, a_local )
time.sleep(rank*0.1)
print(rank,":::",a_local)
totalsum = psum(a_local)
print(totalsum)
In [ ]:
!mpirun -n 4 /opt/conda/envs/py27/bin/python psum.py
In [ ]:
In [ ]: