python multiprocessing 多进程并行计算的操作
作者:探索者v
python的multiprocessing包是标准库提供的多进程并行计算包,提供了和threading(多线程)相似的API函数,但是相比于threading,将任务分配到不同的CPU,避免了GIL(Global Interpreter Lock)的限制。
下面我们对multiprocessing中的Pool和Process类做介绍。
Pool
采用Pool进程池对任务并行处理更加方便,我们可以指定并行的CPU个数,然后 Pool 会自动把任务放到进程池中运行。 Pool 包含了多个并行函数。
apply apply_async
apply 要逐个执行任务,在python3中已经被弃用,而apply_async是apply的异步执行版本。并行计算一定要采用apply_async函数。
import multiprocessing import time from random import randint, seed def f(num): seed() rand_num = randint(0,10) # 每次都随机生成一个停顿时间 time.sleep(rand_num) return (num, rand_num) start_time = time.time() cores = multiprocessing.cpu_count() pool = multiprocessing.Pool(processes=cores) pool_list = [] result_list = [] start_time = time.time() for xx in xrange(10): pool_list.append(pool.apply_async(f, (xx, ))) # 这里不能 get, 会阻塞进程 result_list = [xx.get() for xx in pool_list] #在这里不免有人要疑问,为什么不直接在 for 循环中直接 result.get()呢?这是因为pool.apply_async之后的语句都是阻塞执行的,调用 result.get() 会等待上一个任务执行完之后才会分配下一个任务。事实上,获取返回值的过程最好放在进程池回收之后进行,避免阻塞后面的语句。 # 最后我们使用一下语句回收进程池: pool.close() pool.join() print result_list print '并行花费时间 %.2f' % (time.time() - start_time) print '串行花费时间 %.2f' % (sum([xx[1] for xx in result_list])) #[(0, 8), (1, 2), (2, 4), (3, 9), (4, 0), (5, 1), (6, 8), (7, 3), (8, 4), (9, 6)] #并行花费时间 14.11 #串行花费时间 45.00
map map_async
map_async 是 map的异步执行函数。
相比于 apply_async, map_async 只能接受一个参数。
import time from multiprocessing import Pool def run(fn): #fn: 函数参数是数据列表的一个元素 time.sleep(1) return fn*fn if __name__ == "__main__": testFL = [1,2,3,4,5,6] print '串行:' #顺序执行(也就是串行执行,单进程) s = time.time() for fn in testFL: run(fn) e1 = time.time() print "顺序执行时间:", int(e1 - s) print '并行:' #创建多个进程,并行执行 pool = Pool(4) #创建拥有5个进程数量的进程池 #testFL:要处理的数据列表,run:处理testFL列表中数据的函数 rl =pool.map(run, testFL) pool.close()#关闭进程池,不再接受新的进程 pool.join()#主进程阻塞等待子进程的退出 e2 = time.time() print "并行执行时间:", int(e2-e1) print rl # 串行: # 顺序执行时间: 6 # 并行: # 并行执行时间: 2 # [1, 4, 9, 16, 25, 36]
Process
采用Process必须注意的是,Process对象来创建进程,每一个进程占据一个CPU,所以要建立的进程必须 小于等于 CPU的个数。
如果启动进程数过多,特别是当遇到CPU密集型任务,会降低并行的效率。
#16.6.1.1. The Process class from multiprocessing import Process, cpu_count import os import time start_time = time.time() def info(title): # print(title) if hasattr(os, 'getppid'): # only available on Unix print 'parent process:', os.getppid() print 'process id:', os.getpid() time.sleep(3) def f(name): info('function f') print 'hello', name if __name__ == '__main__': # info('main line') p_list = [] # 保存Process新建的进程 cpu_num = cpu_count() for xx in xrange(cpu_num): p_list.append(Process(target=f, args=('xx_%s' % xx,))) for xx in p_list: xx.start() for xx in p_list: xx.join() print('spend time: %.2f' % (time.time() - start_time)) parent process: 11741 # parent process: 11741 # parent process: 11741 # process id: 12249 # process id: 12250 # parent process: 11741 # process id: 12251 # process id: 12252 # hello xx_1 # hello xx_0 # hello xx_2 # hello xx_3 # spend time: 3.04
进程间通信
Process和Pool均支持Queues 和 Pipes 两种类型的通信。
Queue 队列
队列遵循先进先出的原则,可以在各个进程间使用。
# 16.6.1.2. Exchanging objects between processes # Queues from multiprocessing import Process, Queue def f(q): q.put([42, None, 'hello']) if __name__ == '__main__': q = Queue() p = Process(target=f, args=(q,)) p.start() print q.get() # prints "[42, None, 'hello']" p.join()
pipe
from multiprocessing import Process, Pipe def f(conn): conn.send([42, None, 'hello']) conn.close() if __name__ == '__main__': parent_conn, child_conn = Pipe() p = Process(target=f, args=(child_conn,)) p.start() print parent_conn.recv() # prints "[42, None, 'hello']" p.join()
queue 与 pipe比较
Pipe() can only have two endpoints.
Queue() can have multiple producers and consumers.
When to use them
If you need more than two points to communicate, use a Queue().
If you need absolute performance, a Pipe() is much faster because Queue() is built on top of Pipe().
参考:
https://stackoverflow.com/questions/8463008/python-multiprocessing-pipe-vs-queue
共享资源
多进程应该避免共享资源。在多线程中,我们可以比较容易地共享资源,比如使用全局变量或者传递参数。
在多进程情况下,由于每个进程有自己独立的内存空间,以上方法并不合适。
此时我们可以通过共享内存和Manager的方法来共享资源。
但这样做提高了程序的复杂度,并因为同步的需要而降低了程序的效率。
共享内存
共享内存仅适用于 Process 类,不能用于进程池 Pool
# 16.6.1.4. Sharing state between processes # Shared memory from multiprocessing import Process, Value, Array def f(n, a): n.value = 3.1415927 for i in range(len(a)): a[i] = -a[i] if __name__ == '__main__': num = Value('d', 0.0) arr = Array('i', range(10)) p = Process(target=f, args=(num, arr)) p.start() p.join() print num.value print arr[:] # 3.1415927 # [0, -1, -2, -3, -4, -5, -6, -7, -8, -9]
Manager Class
Manager Class 既可以用于Process 也可以用于进程池 Pool。
from multiprocessing import Manager, Process def f(d, l, ii): d[ii] = ii l.append(ii) if __name__ == '__main__': manager = Manager() d = manager.dict() l = manager.list(range(10)) p_list = [] for xx in range(4): p_list.append(Process(target=f, args=(d, l, xx))) for xx in p_list: xx.start() for xx in p_list: xx.join() print d print l # {0: 0, 1: 1, 2: 2, 3: 3} # [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, 1, 2, 3]
补充:python程序多进程运行时间计算/多进程写数据/多进程读数据
import time time_start=time.time() time_end=time.time() print('time cost',time_end-time_start,'s')
单位为秒,也可以换算成其他单位输出
注意写测试的时候,函数名要以test开头,否则运行不了。
多线程中的问题:
1)多线程存数据:
def test_save_features_to_db(self): df1 = pd.read_csv('/home/sc/PycharmProjects/risk-model/xg_test/statis_data/shixin_company.csv') com_list = df1['company_name'].values.tolist() # com_list = com_list[400015:400019] # print 'test_save_features_to_db' # print(com_list) p_list = [] # 进程列表 i = 1 p_size = len(com_list) for company_name in com_list: # 创建进程 p = Process(target=self.__save_data_iter_method, args=[company_name]) # p.daemon = True p_list.append(p) # 间歇执行进程 if i % 20 == 0 or i == p_size: # 20页处理一次, 最后一页处理剩余 for p in p_list: p.start() for p in p_list: p.join() # 等待进程结束 p_list = [] # 清空进程列表 i += 1
总结:多进程写入的时候,不需要lock,也不需要返回值。
核心p = Process(target=self.__save_data_iter_method, args=[company_name]),其中target指向多进程的一次完整的迭代,arg则是该迭代的输入。
注意写法args=[company_name]才对,原来写成:args=company_name,args=(company_name)会报如下错:只需要1个参数,而给出了34个参数。
多进程外层循环则是由输入决定的,有多少个输入就为多少次循环,理解p.start和p.join;
def __save_data_iter_method(self, com): # time_start = time.time() # print(com) f_d_t = ShiXinFeaturesDealSvc() res = f_d_t.get_time_features(company_name=com) # 是否失信 shixin_label = res.shixin_label key1 = res.shixin_time if key1: public_at = res.shixin_time company_name = res.time_map_features[key1].company_name # print(company_name) established_years = res.time_map_features[key1].established_years industry_dx_rate = res.time_map_features[key1].industry_dx_rate regcap_change_cnt = res.time_map_features[key1].regcap_change_cnt share_change_cnt = res.time_map_features[key1].share_change_cnt industry_dx_cnt = res.time_map_features[key1].industry_dx_cnt address_change_cnt = res.time_map_features[key1].address_change_cnt fr_change_cnt = res.time_map_features[key1].fr_change_cnt judgedoc_cnt = res.time_map_features[key1].judgedoc_cnt bidding_cnt = res.time_map_features[key1].bidding_cnt trade_mark_cnt = res.time_map_features[key1].trade_mark_cnt network_share_cancel_cnt = res.time_map_features[key1].network_share_cancel_cnt cancel_cnt = res.time_map_features[key1].cancel_cnt industry_all_cnt = res.time_map_features[key1].industry_all_cnt network_share_zhixing_cnt = res.time_map_features[key1].network_share_zhixing_cnt network_share_judge_doc_cnt = res.time_map_features[key1].network_share_judge_doc_cnt net_judgedoc_defendant_cnt = res.time_map_features[key1].net_judgedoc_defendant_cnt judge_doc_cnt = res.time_map_features[key1].judge_doc_cnt f_d_do = ShixinFeaturesDto(company_name=company_name, established_years=established_years, industry_dx_rate=industry_dx_rate, regcap_change_cnt=regcap_change_cnt, share_change_cnt=share_change_cnt, industry_all_cnt=industry_all_cnt, industry_dx_cnt=industry_dx_cnt, address_change_cnt=address_change_cnt, fr_change_cnt=fr_change_cnt, judgedoc_cnt=judgedoc_cnt, bidding_cnt=bidding_cnt, trade_mark_cnt=trade_mark_cnt, network_share_cancel_cnt=network_share_cancel_cnt, cancel_cnt=cancel_cnt, network_share_zhixing_cnt=network_share_zhixing_cnt, network_share_judge_doc_cnt=network_share_judge_doc_cnt, net_judgedoc_defendant_cnt=net_judgedoc_defendant_cnt, judge_doc_cnt=judge_doc_cnt, public_at=public_at, shixin_label=shixin_label) # time_end = time.time() # print('totally cost', time_end - time_start) self.cfdbsvc.save_or_update_features(f_d_do) def save_or_update_features(self, shixin_features_dto): """ 添加或更新: 插入一行数据, 如果不存在则插入,存在则更新 """ self._pg_util = PgUtil() p_id = None if isinstance(shixin_features_dto, ShixinFeaturesDto): p_id = str(uuid.uuid1()) self._pg_util.execute_sql( self.s_b.insert_or_update_row( self.model.COMPANY_NAME, { self.model.ID: p_id, # 公司名 self.model.COMPANY_NAME: shixin_features_dto.company_name, # 失信时间 self.model.PUBLIC_AT: shixin_features_dto.public_at, self.model.SHIXIN_LABEL : shixin_features_dto.shixin_label, self.model.ESTABLISHED_YEARS: shixin_features_dto.established_years, self.model.INDUSTRY_DX_RATE: shixin_features_dto.industry_dx_rate, self.model.REGCAP_CHANGE_CNT: shixin_features_dto.regcap_change_cnt, self.model.SHARE_CHANGE_CNT: shixin_features_dto.share_change_cnt, self.model.INDUSTRY_ALL_CNT: shixin_features_dto.industry_all_cnt, self.model.INDUSTRY_DX_CNT: shixin_features_dto.industry_dx_cnt, self.model.ADDRESS_CHANGE_CNT: shixin_features_dto.address_change_cnt, self.model.NETWORK_SHARE_CANCEL_CNT: shixin_features_dto.network_share_cancel_cnt, self.model.CANCEL_CNT: shixin_features_dto.cancel_cnt, self.model.NETWORK_SHARE_ZHIXING_CNT: shixin_features_dto.network_share_zhixing_cnt, self.model.FR_CHANGE_CNT: shixin_features_dto.fr_change_cnt, self.model.JUDGEDOC_CNT: shixin_features_dto.judgedoc_cnt, self.model.NETWORK_SHARE_JUDGE_DOC_CNT: shixin_features_dto.network_share_judge_doc_cnt, self.model.BIDDING_CNT: shixin_features_dto.bidding_cnt, self.model.TRADE_MARK_CNT: shixin_features_dto.trade_mark_cnt, self.model.JUDGE_DOC_CNT: shixin_features_dto.judge_doc_cnt }, [self.model.ADDRESS_CHANGE_CNT,self.model.BIDDING_CNT,self.model.CANCEL_CNT, self.model.ESTABLISHED_YEARS,self.model.FR_CHANGE_CNT,self.model.INDUSTRY_ALL_CNT, self.model.INDUSTRY_DX_RATE,self.model.INDUSTRY_DX_CNT,self.model.JUDGE_DOC_CNT, self.model.JUDGEDOC_CNT,self.model.NETWORK_SHARE_CANCEL_CNT,self.model.NETWORK_SHARE_JUDGE_DOC_CNT, self.model.NETWORK_SHARE_ZHIXING_CNT,self.model.REGCAP_CHANGE_CNT,self.model.TRADE_MARK_CNT, self.model.SHARE_CHANGE_CNT,self.model.SHIXIN_LABEL,self.model.PUBLIC_AT] ) ) return p_id
函数中重新初始化了self._pg_util = PgUtil(),否则会报ssl error 和ssl decryption 的错误,背后原因有待研究!
**2)多进程取数据——(思考取数据为何要多进程)** def flush_process(self, lock): #需要传入lock; """ 运行待处理的方法队列 :type lock Lock :return 返回一个dict """ # process_pool = Pool(processes=20) # data_list = process_pool.map(one_process, self.__process_data_list) # # for (key, value) in data_list: # # 覆盖上期变量 self.__dct_share = self.__manager.Value('tmp', {}) # 进程共享变量 p_list = [] # 进程列表 i = 1 p_size = len(self.__process_data_list) for process_data in self.__process_data_list: **#循环遍历需要同时查找的公司列表!!!self.__process_data_list包含多个process_data,每个process_data包含三种属性?类对象也可以循环????** # 创建进程 p = Process(target=self.__one_process, args=(process_data, lock)) #参数需要lock # p.daemon = True p_list.append(p) # 间歇执行进程 if i % 20 == 0 or i == p_size: # 20页处理一次, 最后一页处理剩余 for p in p_list: p.start() for p in p_list: p.join() # 等待进程结束 p_list = [] # 清空进程列表 i += 1 # end for self.__process_data_list = [] # 清空订阅 return self.__dct_share.value def __one_process(self, process_data, lock): #迭代函数 """ 处理进程 :param process_data: 方法和参数集等 :param lock: 保护锁 """ fcn = process_data.fcn params = process_data.params data_key = process_data.data_key if isinstance(params, tuple): data = fcn(*params) #**注意:*params 与 params区别** else: data = fcn(params) with lock: temp_dct = dict(self.__dct_share.value) if data_key not in temp_dct: temp_dct[data_key] = [] temp_dct[data_key].append(data) self.__dct_share.value = temp_dct
主程序调用:
def exe_process(self, company_name, open_from, time_nodes): """ 多进程执行pre订阅的数据 :param company_name: 公司名 :return: """ mul_process_helper = MulProcessHelper() lock = Lock() self.__get_time_bidding_statistic(company_name, mul_process_helper) data = mul_process_helper.flush_process(lock) return data def __get_time_bidding_statistic(self, company_name, mul_process_helper): # 招投标信息 process_data = ProcessData(f_e_t_svc.get_bidding_statistic_time_node_api, company_name, self.__BIDDING_STATISTIC_TIME) **#此处怎么理解?ProcessData是一个类!!!** mul_process_helper.add_process_data_list(process_data) #同时调用多个api???将api方法当做迭代????用于同时查找多个公司???? def add_process_data_list(self, process_data): """ 添加用于进程处理的方法队列 :type process_data ProcessData :param process_data: :return: """ self.__process_data_list.append(process_data) class ProcessData(object): """ 用于进程处理的的数据 """ def __init__(self, fcn, params, data_key): self.fcn = fcn # 方法 self.params = params # 参数 self.data_key = data_key # 存储到进程共享变量中的名字
以上为个人经验,希望能给大家一个参考,也希望大家多多支持脚本之家。如有错误或未考虑完全的地方,望不吝赐教。