Python 高性能并行计算之 mpi4py

时间：2018-08-16 00:48:26 阅读：1053 评论：0 收藏：0 [点我收藏+]

标签：poi details 存在拷贝 cal target 规约 3.1 sdn

MPI 和 MPI4PY 的搭建上一篇文章已经介绍，这里面介绍一些基本用法。

mpi4py 的 helloworld

from mpi4py import MPI
print("hello world")

mpiexec -n 5 python3 x.py

技术分享图片

2. 点对点通信

因为 mpi4py 中点对点的通信 send 语句在数据量较小的时候是把发送数据拷贝到缓存区，是非堵塞的操作，然而在数据量较大时候是堵塞操作，由此如下：

在发送较小数据时：

import mpi4py.MPI as MPI
 
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()
 
# point to point communication
data_send = [comm_rank]*5

comm.send(data_send,dest=(comm_rank+1)%comm_size)

data_recv =comm.recv(source=(comm_rank-1)%comm_size)

print("my rank is %d, and Ireceived:" % comm_rank)
print(data_recv)

技术分享图片

在数据量较大时，比如发送：

# point to point communication
data_send = [comm_rank]*1000000

这时候就会造成各个进程之间的死锁。（因为这时候各个进程是堵塞执行，每个进程都在等待另一个进程的发送数据）

修改后的代码，所有进程顺序执行， 0进程发送给1,1接收然后发送给2，以此类推：

import mpi4py.MPI as MPI
     
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()
     
data_send = [comm_rank]*1000000

if comm_rank == 0:
   comm.send(data_send, dest=(comm_rank+1)%comm_size)

if comm_rank > 0:
   data_recv = comm.recv(source=(comm_rank-1)%comm_size)
   comm.send(data_send, dest=(comm_rank+1)%comm_size)

if comm_rank == 0:
   data_recv = comm.recv(source=(comm_rank-1)%comm_size)

print("my rank is %d, and Ireceived:" % comm_rank)
print(data_recv)

3 群体通信

3.1 广播bcast

一个进程把数据发送给所有进程

import mpi4py.MPI as MPI
 
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()
 
if comm_rank == 0:
   data = range(comm_size)

dat = comm.bcast(data if comm_rank == 0 else None, root=0)

print(‘rank %d, got:‘ % (comm_rank))
print(dat)

技术分享图片

发送方也会收到这部分数据，当然发送方这份数据并不是网络传输接受的，而是本身内存空间中就是存在的。

3.2 散播scatter

import mpi4py.MPI as MPI
 
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()
 
if comm_rank == 0:
   data = range(comm_size)
else:
   data = None

local_data = comm.scatter(data, root=0)

print(‘rank %d, got:‘ % comm_rank)
print(local_data)

3.3 收集gather

将所有数据搜集回来

import mpi4py.MPI as MPI
 
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()
 
if comm_rank == 0:
   data = range(comm_size)
else:
   data = None

local_data = comm.scatter(data, root=0)
local_data = local_data * 2

print(‘rank %d, got and do:‘ % comm_rank)
print(local_data)

combine_data = comm.gather(local_data,root=0)

if comm_rank == 0:
    print("root recv {0}".format(combine_data))

技术分享图片

3.4 规约reduce

import mpi4py.MPI as MPI
 
comm = MPI.COMM_WORLD
comm_rank = comm.Get_rank()
comm_size = comm.Get_size()

if comm_rank == 0:
   data = range(comm_size)
else:
   data = None

local_data = comm.scatter(data, root=0)
local_data = local_data * 2

print(‘rank %d, got and do:‘ % comm_rank)
print(local_data)

all_sum = comm.reduce(local_data, root=0,op=MPI.SUM)

if comm_rank == 0:
    print(‘sum is:%d‘ % all_sum)