MPI(Message Passing Interface),消息传递接口,是一个标准化和轻便的能够运行在各种各样并行计算机上的消息传递系统。消息传递指的是并行执行的各个进程拥有自己独立的堆栈和代码段,作为互不相关的多个程序独立执行,进程之间的信息交互完全通过显示地调用通信函数来完成。
mpi4py是构建在MPI之上的Python非官方库,使得Python的数据可以在进程之间进行传递。
并行程序是指一组独立、同一的处理过程;
当使用Python,使用n个Python解释器;
mpirun -np 32 python parallel_script.py
并行执行模型如下所示,
rank:给予每个进程的id;
Communicator:包含进程的群组;
所有的变量和数据结构都是进程的局部值;
进程之间通过发送和接收消息来交换数据;
from mpi4py import MPI
comm = MPI.COMM_WORLD #Communicator对象包含所有进程
size = comm.Get_size()
rank = comm.Get_rank()
print "rank = %d,size = %d"%(rank,size)
MPI Python环境搭建
MPI Windows集群环境搭建
工作方式主要有点对点和群体通信两种;点对点通信就是一对一,群体通信是一对多;
example 1
点对点发送Python内置dict对象;
#Broadcasting a Python dict
from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:
data = {"a":7,"b":3.14}
comm.send(data,dest = 1,tag = 11)
print "send data = ",data
elif rank == 1:
data = comm.recv(source = 0,tag = 11)
print "recv data = ",data
任意的Python内置对象可以通过send和recv进行通信,目标rank和源rank和tag都要互相匹配;
send(data,dest,tag)
recv(source,tag)
example 2
点对点发送Python内置dict对象,非阻塞通信;
#point to point communication Python objects with non-blocking communication
from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank ==0:
data = {"a":7,"b":3.14}
req = comm.isend(data,dest = 1,tag = 11)
req.wait()
print "send data = ",data
elif rank == 1:
req = comm.irecv(source = 0,tag = 11)
data = req.wait()
print "recv data = ",data
example 3
发送Numpy数组;
#point to point communication Python objects Numpy arrays
from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
# automatic MPI datatypes discovery
if rank == 0:
data = np.arange(100,dtype = np.int)
comm.Send(data, dest = 1,tag = 13)
print "send data = ",data
elif rank == 1:
data = np.empty(100,dtype = np.int)
comm.Recv(data, source = 0,tag = 13)
print "recv data = ",data
当发送消息时,任意的Python对象转换为字节流;
当接收消息时,字节流被转换为Python对象;
Send(data,dest,tag),Recv(data,source,tag),连续型数组,速度快;
send(data,dest,tag),recv(source,tag),Python内置对象,速度慢;
群体通信分为发送和接收,发送是一次性把数据发给所有人,接收是一次性从所有人那里回收结果;
example 1
root进程新建data dict,然后将data数据广播给所有的进程,这样所有的进程都拥有这个data dict;
#Broadcasting a Python dict
from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:
data = {"key1":[7,2.72,2+3j],"key2":("abc","xyz")}
else:
data = None
data = comm.bcast(data,root = 0)
print "rank = ",rank," data = ",data
example 2
root进程新建了一个list,然后将它散播给所有的进程,相当于对这个list做了划分,每个进程获得等分的数据,这里就是list中的每一个数字(主要根据list的索引来划分,list索引为第i份的数据就发送给第i个进程),如果是矩阵,那么久等分的划分行,每个进程获得相同的行数进行处理;
MPI的工作方式是每个进程都会执行所有的代码,每个进程都会执行scatter这个指令,但是只有root进程执行它的时候,它才兼备发送者和接收者的身份(root进程也会得到数据它自己的那份数据),对于其他进程来说,他们都只是接收者而已;
#Scattering Python objects
from mpi4py import MPI
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
if rank == 0:
data = [(i+1)**2 for i in range(size)]
else:
data = None
data = comm.scatter(data,root = 0)
assert data == (rank+1)**2
print "rank = ",rank," data = ",data
example 3
gather是将所有进程的数据收集回来,然后合并成一个列表;
#Gathering Python objects
from mpi4py import MPI
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
data = (rank+1)**2
data = comm.gather(data,root = 0)
if rank == 0:
for i in range(size):
assert(data[i] == (i+1)**2)
print "data = ",data
else:
assert data is None
example 4
广播Numpy数组;
#Broadcasting Numpy array
from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
if rank == 0:
data = np.arange(100, dtype = 'i')
else:
data = np.empty(100,dtype = 'i')
comm.Bcast(data,root = 0)
for i in range(100):
assert(data[i] == i)
print "rank = ",rank," data = ",data
example 5
散播Numpy数组;
#Scattering Numpy arrays
from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
senbuf = None
if rank == 0:
senbuf = np.empty([size,100],dtype = 'i')
senbuf.T[:,:] = range(size)
recvbuf = np.empty(100,dtype = 'i')
comm.Scatter(senbuf,recvbuf,root = 0)
assert np.allclose(recvbuf,rank)
print "rank = ",rank," recvbuf = ",recvbuf
example 6
收集Numpy数组;
#Gathering Numpy array
from mpi4py import MPI
import numpy as np
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
sendbuf = np.zeros(100, dtype='i') + rank
recvbuf = None
if rank == 0:
recvbuf = np.empty([size, 100], dtype='i')
comm.Gather(sendbuf, recvbuf, root=0)
if rank == 0:
for i in range(size):
assert np.allclose(recvbuf[i,:], i)
mpi4py tutorial
Python多核编程mpi4py实践