Python并行运算模块Parallel Python简介

一、概览

PP是一个python模块，提供在SMP（具有多个处理器或多核的系统）和集群（通过网络连接的计算机）上并行执行python代码的机制。它轻巧，易于安装和与其他python软件集成。PP是一个用纯Python编写的开源和跨平台模块。

二、特性

1. 在SMP和集群上并行执行python代码
2. 易于理解和实现基于Job的并行化技术（易于并行转换串行应用程序）
3. 自动检测最佳配置（默认情况下工作进程数设置为有效处理器数）
4. 动态处理器分配（工作进程数可以在运行时更改）
5. 具有相同功能的后续作业的低开销（实现透明高速缓存以减少开销）
6. 动态负载平衡（作业在运行时在处理器之间分布）
7. 容错（如果其中一个节点发生故障，任务在其他节点上重新调度）
8. 计算资源的自动发现
9. 计算资源的动态分配（自动发现和容错的结果）
10. 网络连接的基于SHA的认证
11. 跨平台可移植性和互操作性（Windows，Linux，Unix，Mac OS X）
12. 跨架构可移植性和互操作性（x86，x86-64等）
13. 开源

三、动机

现在，用python编写的软件应用在很多应用程序中，包括业务逻辑，数据分析和科学计算。这与市场上的SMP计算机（多处理器或多核）和集群（计算机通过网络连接）的广泛可用性一起创建了并行执行python代码的需求。
为SMP计算机编写并行应用程序的最简单和常见的方法是使用线程。虽然，如果应用程序是计算绑定使用线程或线程python模块将不允许并行运行python字节码。原因是python解释器使用GIL（全局解释器锁）进行内部记账。这个锁允许一次只执行一个python字节码指令，即使在SMP计算机上。
PP模块克服了这个限制，并提供了一种写并行python应用程序的简单方法。内部ppsmp使用进程和IPC（进程间通信）来组织并行计算。后者的所有细节和复杂性完全被照顾，应用程序只提交作业并检索其结果（写并行应用程序的最简单的方法）。
为了使事情更好，用PP编写的软件并行工作，即使在通过本地网络或Internet连接的许多计算机上。跨平台可移植性和动态负载平衡允许PP即使在异构和多平台集群上也能有效地并行计算。

四、安装

任何平台：下载模块存档并将其解压缩到本地目录。 运行安装脚本：python setup.py install
Windows：下载并执行Windows安装程序二进制文件。

五、例子

Example #1: sum_primes.py  

#!/usr/bin/python
# File: sum_primes.py
# Author: VItalii Vanovschi
# Desc: This program demonstrates parallel computations with pp module
# It calculates the sum of prime numbers below a given integer in parallel
# Parallel Python Software: http://www.parallelpython.com

import math, sys, time
import pp

def isprime(n):
    """Returns True if n is prime and False otherwise"""
    if not isinstance(n, int):
        raise TypeError("argument passed to is_prime is not of 'int' type")
    if n < 2:
        return False
    if n == 2:
        return True
    max = int(math.ceil(math.sqrt(n)))
    i = 2
    while i <= max:
        if n % i == 0:
            return False
        i += 1
    return True

def sum_primes(n):
    """Calculates sum of all primes below given integer n"""
    return sum([x for x in xrange(2,n) if isprime(x)])

print """Usage: python sum_primes.py [ncpus]
    [ncpus] - the number of workers to run in parallel, 
    if omitted it will be set to the number of processors in the system
"""

# tuple of all parallel python servers to connect with
ppservers = ()
#ppservers = ("10.0.0.1",)

if len(sys.argv) > 1:
    ncpus = int(sys.argv[1])
    # Creates jobserver with ncpus workers
    job_server = pp.Server(ncpus, ppservers=ppservers)
else:
    # Creates jobserver with automatically detected number of workers
    job_server = pp.Server(ppservers=ppservers)

print "Starting pp with", job_server.get_ncpus(), "workers"

# Submit a job of calulating sum_primes(100) for execution. 
# sum_primes - the function
# (100,) - tuple with arguments for sum_primes
# (isprime,) - tuple with functions on which function sum_primes depends
# ("math",) - tuple with module names which must be imported before sum_primes execution
# Execution starts as soon as one of the workers will become available
job1 = job_server.submit(sum_primes, (100,), (isprime,), ("math",))

# Retrieves the result calculated by job1
# The value of job1() is the same as sum_primes(100)
# If the job has not been finished yet, execution will wait here until result is available
result = job1()

print "Sum of primes below 100 is", result

start_time = time.time()

# The following submits 8 jobs and then retrieves the results
inputs = (100000, 100100, 100200, 100300, 100400, 100500, 100600, 100700)
jobs = [(input, job_server.submit(sum_primes,(input,), (isprime,), ("math",))) for input in inputs]
for input, job in jobs:
    print "Sum of primes below", input, "is", job()

print "Time elapsed: ", time.time() - start_time, "s"
job_server.print_stats()

# Parallel Python Software: http://www.parallelpython.com



Example #2: reverse_md5.py 

#!/usr/bin/python
# File: reverse_md5.py
# Author: Vitalii Vanovschi
# Desc: This program demonstrates parallel computations with pp module
# It tries to reverse an md5 hash in parallel
# Parallel Python Software: http://www.parallelpython.com

import math, sys, md5, time
import pp

def md5test(hash, start, end):
    """Calculates md5 of the integerss between 'start' and 'end' and compares it with 'hash'"""
    for x in xrange(start, end):
        if md5.new(str(x)).hexdigest() == hash:
            return x

print """Usage: python reverse_md5.py [ncpus]
    [ncpus] - the number of workers to run in parallel, 
    if omitted it will be set to the number of processors in the system
"""

# tuple of all parallel python servers to connect with
ppservers = ()
#ppservers = ("10.0.0.1",)

if len(sys.argv) > 1:
    ncpus = int(sys.argv[1])
    # Creates jobserver with ncpus workers
    job_server = pp.Server(ncpus, ppservers=ppservers)
else:
    # Creates jobserver with automatically detected number of workers
    job_server = pp.Server(ppservers=ppservers)

print "Starting pp with", job_server.get_ncpus(), "workers"

#Calculates md5 hash from the given number
hash = md5.new("1829182").hexdigest()
print "hash =", hash
#Now we will try to find the number with this hash value

start_time = time.time()
start = 1
end = 2000000

# Since jobs are not equal in the execution time, division of the problem 
# into a 100 of small subproblems leads to a better load balancing
parts = 128

step = (end - start) / parts + 1
jobs = []

for index in xrange(parts):
    starti = start+index*step
    endi = min(start+(index+1)*step, end)
    # Submit a job which will test if a number in the range starti-endi has given md5 hash
    # md5test - the function
    # (hash, starti, endi) - tuple with arguments for md5test
    # () - tuple with functions on which function md5test depends
    # ("md5",) - tuple with module names which must be imported before md5test execution
    jobs.append(job_server.submit(md5test, (hash, starti, endi), (), ("md5",)))

# Retrieve results of all submited jobs
for job in jobs:
    result = job()
    if result:
        break

# Print the results
if result:
    print "Reverse md5 for", hash, "is", result
else:
    print "Reverse md5 for", hash, "has not been found"

print "Time elapsed: ", time.time() - start_time, "s"
job_server.print_stats()

# Parallel Python Software: http://www.parallelpython.com
Example #3: dynamic_ncpus.py  

#!/usr/bin/python
# File: dynamic_ncpus.py
# Author: Vitalii Vanovschi
# Desc: This program demonstrates parallel computations with pp module 
# and dynamic cpu allocation feature.
# Program calculates the partial sum 1-1/2+1/3-1/4+1/5-1/6+... (in the limit it is ln(2))
# Parallel Python Software: http://www.parallelpython.com

import math, sys, md5, time
import pp

def part_sum(start, end):
    """Calculates partial sum"""
    sum = 0
    for x in xrange(start, end):
        if x % 2 == 0:
           sum -= 1.0 / x 
        else:
           sum += 1.0 / x 
    return sum

print """Usage: python dynamic_ncpus.py"""
print 

start = 1
end = 20000000

# Divide the task into 64 subtasks
parts = 64
step = (end - start) / parts + 1

# Create jobserver
job_server = pp.Server()

# Execute the same task with different amount of active workers and measure the time
for ncpus in (1, 2, 4, 8, 16, 1):
    job_server.set_ncpus(ncpus)
    jobs = []
    start_time = time.time()
    print "Starting ", job_server.get_ncpus(), " workers"
    for index in xrange(parts):
        starti = start+index*step
        endi = min(start+(index+1)*step, end)
        # Submit a job which will calculate partial sum 
        # part_sum - the function
        # (starti, endi) - tuple with arguments for part_sum
        # () - tuple with functions on which function part_sum depends
        # () - tuple with module names which must be imported before part_sum execution
        jobs.append(job_server.submit(part_sum, (starti, endi)))

    # Retrieve all the results and calculate their sum
    part_sum1 = sum([job() for job in jobs])
    # Print the partial sum
    print "Partial sum is", part_sum1, "| diff =", math.log(2) - part_sum1

    print "Time elapsed: ", time.time() - start_time, "s"
    print
job_server.print_stats()

# Parallel Python Software: http://www.parallelpython.com



Example #4: callback.py  

#!/usr/bin/python
# File: callback.py
# Author: Vitalii Vanovschi
# Desc: This program demonstrates parallel computations with pp module 
# using callbacks (available since pp 1.3).
# Program calculates the partial sum 1-1/2+1/3-1/4+1/5-1/6+... (in the limit it is ln(2))
# Parallel Python Software: http://www.parallelpython.com

import math, time, thread, sys
import pp

#class for callbacks
class Sum:
    def __init__(self):
        self.value = 0.0
        self.lock = thread.allocate_lock()
        self.count = 0

    #the callback function
    def add(self, value):
        # we must use lock here because += is not atomic
        self.count += 1
        self.lock.acquire()
        self.value += value
        self.lock.release()

def part_sum(start, end):
    """Calculates partial sum"""
    sum = 0
    for x in xrange(start, end):
        if x % 2 == 0:
           sum -= 1.0 / x 
        else:
           sum += 1.0 / x 
    return sum

print """Usage: python callback.py [ncpus]
    [ncpus] - the number of workers to run in parallel, 
    if omitted it will be set to the number of processors in the system
"""

start = 1
end = 20000000

# Divide the task into 128 subtasks
parts = 128
step = (end - start) / parts + 1

# tuple of all parallel python servers to connect with
ppservers = ()
#ppservers = ("localhost",)

if len(sys.argv) > 1:
    ncpus = int(sys.argv[1])
    # Creates jobserver with ncpus workers
    job_server = pp.Server(ncpus, ppservers=ppservers)
else:
    # Creates jobserver with automatically detected number of workers
    job_server = pp.Server(ppservers=ppservers)

print "Starting pp with", job_server.get_ncpus(), "workers"

# Create an instance of callback class
sum = Sum()

# Execute the same task with different amount of active workers and measure the time
start_time = time.time()
for index in xrange(parts):
    starti = start+index*step
    endi = min(start+(index+1)*step, end)
    # Submit a job which will calculate partial sum 
    # part_sum - the function
    # (starti, endi) - tuple with arguments for part_sum
    # callback=sum.add - callback function
    job_server.submit(part_sum, (starti, endi), callback=sum.add)

#wait for jobs in all groups to finish 
job_server.wait()

# Print the partial sum
print "Partial sum is", sum.value, "| diff =", math.log(2) - sum.value

job_server.print_stats()

# Parallel Python Software: http://www.parallelpython.com



Example #5: auto_diff.py  

#!/usr/bin/python
# File: auto_diff.py
# Author: Vitalii Vanovschi
# Desc: This program demonstrates parallel computations with pp module 
# using class methods as parallel functions (available since pp 1.4).
# Program calculates the partial sums of f(x) = x-x**2/2+x**3/3-x**4/4+... 
# and first derivatives f'(x) using automatic differentiation technique.
# In the limit f(x) = ln(x+1) and f'(x) = 1/(x+1).
# Parallel Python Software: http://www.parallelpython.com

import math, sys
import pp

# Partial implemenmtation of automatic differentiation class
class AD:
    def __init__(self, x, dx=0.0):
        self.x = float(x)
        self.dx = float(dx)

    def __pow__(self, val):
        if isinstance(val, int):
            p = self.x**val
            return AD(self.x**val, val*self.x**(val-1)*self.dx)
        else:
            raise TypeError("Second argumnet must be an integer")

    def __add__(self, val):
        if isinstance(val, AD):
            return AD(self.x+val.x, self.dx+val.dx) 
        else:
            return AD(self.x+val, self.dx) 

    def __radd__(self, val):
        return self+val

    def __mul__(self, val):
        if isinstance(val, AD):
            return AD(self.x*val.x, self.x*val.dx+val.x*self.dx) 
        else:
            return AD(self.x*val, val*self.dx) 

    def __rmul__(self, val):
        return self*val    

    def __div__(self, val):
        if isinstance(val, AD):
           return self*AD(1/val.x, -val.dx/val.x**2)
        else:
           return self*(1/float(val))

    def __rdiv__(self, val):
        return AD(val)/self

    def __sub__(self, val):
        if isinstance(val, AD):
            return AD(self.x-val.x, self.dx-val.dx) 
        else:
            return AD(self.x-val, self.dx) 

    def __repr__(self):
        return str((self.x, self.dx))

# This class contains methods which will be executed in parallel
class PartialSum:
    def __init__(self, n):
        self.n = n

    #truncated natural logarithm
    def t_log(self, x):
        return self.partial_sum(x-1)

    #partial sum for truncated natural logarithm
    def partial_sum(self, x):
        return sum([float(i%2 and 1 or -1)*x**i/i for i in xrange(1,self.n)])

print """Usage: python auto_diff.py [ncpus]
    [ncpus] - the number of workers to run in parallel, 
    if omitted it will be set to the number of processors in the system
"""

# tuple of all parallel python servers to connect with
ppservers = ()
#ppservers = ("10.0.0.1",)

if len(sys.argv) > 1:
    ncpus = int(sys.argv[1])
    # Creates jobserver with ncpus workers
    job_server = pp.Server(ncpus, ppservers=ppservers)
else:
    # Creates jobserver with automatically detected number of workers
    job_server = pp.Server(ppservers=ppservers)

print "Starting pp with", job_server.get_ncpus(), "workers"

proc = PartialSum(20000)

results = []
for i in range(32):
    # Creates an object with x = float(i)/32+1 and dx = 1.0
    ad_x = AD(float(i)/32+1, 1.0)
    # Submits a job of calulating proc.t_log(x). 
    f = job_server.submit(proc.t_log, (ad_x,))
    results.append((ad_x.x, f))

for x, f in results:
    # Retrieves the result of the calculation
    val = f()
    print "t_log(%lf) = %lf, t_log'(%lf) = %lf" % (x, val.x, x, val.dx)

# Print execution statistics
job_server.print_stats()

# Parallel Python Software: http://www.parallelpython.com