IPython并行计算工具
纸上得来终觉浅,绝知此事要躬行。

解决并行计算和分布式计算的问题
- 运行解释说明
- 一直以来Python的并发问题都会被大家所诟病,正是因为全局解释锁的存在,导致其不能够真正的做到并发的执行。所以,我们就需要ipyparallel的存在来帮助我们处理并发计算的问题了。
- 在ipyparallel中,可以利用多个engine同时运行一个任务来加快处理的速度。集群被抽象为view,包括direct_view和balanced_view。其中,direct_view是所有的engine的抽象,当然也可以自行指定由哪些engine构成,而balanced_view是多个engine经过负载均衡之后,抽象出来的由“单一”engine构成的view。利用ipyparallel并行化的基本思路是将要处理的数据首先进行切分,然后分布到每一个engine上,然后将最终的处理结果合并,得到最终的结果,其思路和mapreduce类似。
- 并行计算分类
- ipcluster - 单机并行计算
- ipyparallel - 分布式计算
- 相关连接地址
- 安装方式
# 使用pip安装 $ pip install ipyparallel
- 配置并行环境
# 命令可以简单的创建一个通用的并行环境profile配置文件 $ ipython profile create --parallel --profile=myprofile
1. 并行计算示例
做一次wordcount的计算测试。
- 数据来源地址
# 使用wget下载 $ wget http://www.gutenberg.org/files/27287/27287-0.txt
- 不并行的版本
In [1]: import re
In [2]: import io
In [3]: from collections import defaultdict
In [4]: non_word = re.compile(r'[Wd]+', re.UNICODE)
In [5]: common_words = {
...: 'the','of','and','in','to','a','is','it','that','which','as','on','by',
...: 'be','this','with','are','from','will','at','you','not','for','no','have',
...: 'i','or','if','his','its','they','but','their','one','all','he','when',
...: 'than','so','these','them','may','see','other','was','has','an','there',
...: 'more','we','footnote', 'who', 'had', 'been', 'she', 'do', 'what',
...: 'her', 'him', 'my', 'me', 'would', 'could', 'said', 'am', 'were', 'very',
...: 'your', 'did', 'not',
...: }
In [6]: def yield_words(filename):
...: import io
...: with io.open(filename, encoding='latin-1') as f:
...: for line in f:
...: for word in line.split():
...: word = non_word.sub('', word.lower())
...: if word and word not in common_words:
...: yield word
...:
In [7]: def word_count(filename):
...: word_iterator = yield_words(filename)
...: counts = {}
...: counts = defaultdict(int)
...: while True:
...: try:
...: word = next(word_iterator)
...: except StopIteration:
...: break
...: else:
...: counts[word] += 1
...: return counts
...:
In [8]: %time counts = word_count(filename)
CPU times: user 3.32 ms, sys: 1.4 ms, total: 4.72 ms
Wall time: 10.9 ms
- 用 IPython 来跑一下
# 在terminal输入如下命令,然后在ipython中就都可使用并行计算 # 指定两个核心来执行 [[email protected] ~]$ ipcluster start -n 2
- 先讲下 IPython 并行计算的用法
# import之后才能用%px*的magic In [1]: from IPython.parallel import Client In [2]: rc = Client() # 因为我启动了2个进程 In [3]: rc.ids Out[3]: [0, 1] # 如果不自动每句都需要: `%px xxx` In [4]: %autopx %autopx enabled # 这里没autopx的话需要: `%px import os` In [5]: import os # 2个进程的pid In [6]: print os.getpid() [stdout:0] 62638 [stdout:1] 62636 # 在autopx下这个magic不可用 In [7]: %pxconfig --targets 1 [stderr:0] ERROR: Line magic function `%pxconfig` not found. [stderr:1] ERROR: Line magic function `%pxconfig` not found. # 再执行一次就会关闭autopx In [8]: %autopx %autopx disabled # 指定目标对象, 这样下面执行的代码就会只在第2个进程下运行 In [10]: %pxconfig --targets 1 # 其实就是执行一段非阻塞的代码 In [11]: %%px --noblock ....: import time ....: time.sleep(1) ....: os.getpid() ....: Out[11]: <AsyncResult: execute> # 看只返回了第二个进程的pid In [12]: %pxresult Out[1:21]: 62636 # 使用全部的进程, ipython可以细粒度的控制那个engine执行的内容 In [13]: v = rc[:] # 每个进程都导入time模块 In [14]: with v.sync_imports(): ....: import time ....: importing time on engine(s) In [15]: def f(x): ....: time.sleep(1) ....: return x * x ....: # 同步的执行 In [16]: v.map_sync(f, range(10)) Out[16]: [0, 1, 4, 9, 16, 25, 36, 49, 64, 81] # 异步的执行 In [17]: r = v.map(f, range(10)) # celery的用法 In [18]: r.ready(), r.elapsed Out[18]: (True, 5.87735) # 获得执行的结果 In [19]: r.get() Out[19]: [0, 1, 4, 9, 16, 25, 36, 49, 64, 81]
- 并行版本
In [20]: def split_text(filename):
....: text = open(filename).read()
....: lines = text.splitlines()
....: nlines = len(lines)
....: n = 10
....: block = nlines//n
....: for i in range(n):
....: chunk = lines[i*block:(i+1)*(block)]
....: with open('count_file%i.txt' % i, 'w') as f:
....: f.write('n'.join(chunk))
....: cwd = os.path.abspath(os.getcwd())
....: # 不用glob是为了精准
....: fnames = [ os.path.join(cwd, 'count_file%i.txt' % i) for i in range(n)]
....: return fnames
In [21]: from IPython import parallel
In [22]: rc = parallel.Client()
In [23]: view = rc.load_balanced_view()
In [24]: v = rc[:]
In [25]: v.push(dict(
....: non_word=non_word,
....: yield_words=yield_words,
....: common_words=common_words
....: ))
Out[25]: <AsyncResult: _push>
In [26]: fnames = split_text(filename)
In [27]: def count_parallel():
.....: pcounts = view.map(word_count, fnames)
.....: counts = defaultdict(int)
.....: for pcount in pcounts.get():
.....: for k, v in pcount.iteritems():
.....: counts[k] += v
.....: return counts, pcounts
.....:
# 这个时间包含了我再聚合的时间
In [28]: %time counts, pcounts = count_parallel()
# 是不是比直接运行少了很多时间
CPU times: user 50.6 ms, sys: 8.82 ms, total: 59.4 ms
# 这个时间是
Wall time: 99.6 ms
In [29]: pcounts.elapsed, pcounts.serial_time, pcounts.wall_time
Out[29]: (0.104384, 0.13980499999999998, 0.104384)
可以看出cpu时间上确实减少了,几乎一半,但真实时间上却反而增加到了164ms,用%timeit 查看,发现实际使用时间反而多出了20ms这是因为cpu计算完后还要聚合结果。这个过程也得耗时,也就是说,并行是有额外开销的。
2. 最简单的应用
并行就是多个核心同时执行任务了,最简单的就是执行重复任务,将函数提交到引擎中。
c = Client() a = lambda :"hi~"
# 并行计算 %time c[:].apply_sync(a) CPU times: user 22.6 ms, sys: 5.05 ms, total: 27.7 ms Wall time: 35.4 ms ['hi~', 'hi~', 'hi~', 'hi~']
# 使用列表生成器 %time [a() for i in range(2)] CPU times: user 10 µs, sys: 6 µs, total: 16 µs Wall time: 17.9 µs ['hi~', 'hi~']
看得出,cpython还是相当给力的,在这种小规模计算上并行反而比用列表生成器慢很多。
3. 直接调用 ipyparallel
我们可以通过DirectView直接在ipython中通过Client对象直接的操作多个engine。
from ipyparallel import Client rc = Client() # 查看有多少个engine rc.ids [0, 1, 2, 3] # 使用全部engine dview = rc[:]
%time map(lambda x:x**2,range(32)) CPU times: user 21 µs, sys: 5 µs, total: 26 µs Wall time: 26.9 µs [0, 1, 4, 9, ..., 900, 961]
# 并行的map工具 %time dview.map_sync(lambda x:x**2,range(32)) CPU times: user 31.3 ms, sys: 5.12 ms, total: 36.4 ms Wall time: 41.4 ms [0, 1, 4, 9, ..., 900, 961]
看来还是单进程给力哇!
4. 负载均衡 view
并行的一大难题便是负载均衡,直接使用DirectView并没有这方面优化,可以使用LoadBalancedView来使用负载均衡的view。
lview = rc.load_balanced_view()
%time lview.map_sync(lambda x:x**2,range(32)) CPU times: user 230 ms, sys: 47.3 ms, total: 277 ms Wall time: 305 ms [0, 1, 4, 9, ..., 900, 961]