def file_len(full_path):
  """ Count number of lines in a file."""
  f = open(full_path)
  nr_of_lines = sum(1 for line in f)
  f.close()
  return nr_of_lines

下面是一个python程序，使用多处理库将行计数分布到不同的机器/核。使用8核windows 64服务器，我的测试将一个2000万行文件的计数从26秒提高到7秒。注意:不使用内存映射会使运行速度变慢。

import multiprocessing, sys, time, os, mmap
import logging, logging.handlers

def init_logger(pid):
    console_format = 'P{0} %(levelname)s %(message)s'.format(pid)
    logger = logging.getLogger()  # New logger at root level
    logger.setLevel( logging.INFO )
    logger.handlers.append( logging.StreamHandler() )
    logger.handlers[0].setFormatter( logging.Formatter( console_format, '%d/%m/%y %H:%M:%S' ) )

def getFileLineCount( queues, pid, processes, file1 ):
    init_logger(pid)
    logging.info( 'start' )

    physical_file = open(file1, "r")
    #  mmap.mmap(fileno, length[, tagname[, access[, offset]]]

    m1 = mmap.mmap( physical_file.fileno(), 0, access=mmap.ACCESS_READ )

    #work out file size to divide up line counting

    fSize = os.stat(file1).st_size
    chunk = (fSize / processes) + 1

    lines = 0

    #get where I start and stop
    _seedStart = chunk * (pid)
    _seekEnd = chunk * (pid+1)
    seekStart = int(_seedStart)
    seekEnd = int(_seekEnd)

    if seekEnd < int(_seekEnd + 1):
        seekEnd += 1

    if _seedStart < int(seekStart + 1):
        seekStart += 1

    if seekEnd > fSize:
        seekEnd = fSize

    #find where to start
    if pid > 0:
        m1.seek( seekStart )
        #read next line
        l1 = m1.readline()  # need to use readline with memory mapped files
        seekStart = m1.tell()

    #tell previous rank my seek start to make their seek end

    if pid > 0:
        queues[pid-1].put( seekStart )
    if pid < processes-1:
        seekEnd = queues[pid].get()

    m1.seek( seekStart )
    l1 = m1.readline()

    while len(l1) > 0:
        lines += 1
        l1 = m1.readline()
        if m1.tell() > seekEnd or len(l1) == 0:
            break

    logging.info( 'done' )
    # add up the results
    if pid == 0:
        for p in range(1,processes):
            lines += queues[0].get()
        queues[0].put(lines) # the total lines counted
    else:
        queues[0].put(lines)

    m1.close()
    physical_file.close()

if __name__ == '__main__':
    init_logger( 'main' )
    if len(sys.argv) > 1:
        file_name = sys.argv[1]
    else:
        logging.fatal( 'parameters required: file-name [processes]' )
        exit()

    t = time.time()
    processes = multiprocessing.cpu_count()
    if len(sys.argv) > 2:
        processes = int(sys.argv[2])
    queues=[] # a queue for each process
    for pid in range(processes):
        queues.append( multiprocessing.Queue() )
    jobs=[]
    prev_pipe = 0
    for pid in range(processes):
        p = multiprocessing.Process( target = getFileLineCount, args=(queues, pid, processes, file_name,) )
        p.start()
        jobs.append(p)

    jobs[0].join() #wait for counting to finish
    lines = queues[0].get()

    logging.info( 'finished {} Lines:{}'.format( time.time() - t, lines ) )

我得到了一个小(4-8%)的改进，这个版本重用了一个常量缓冲区，所以它应该避免任何内存或GC开销:

lines = 0
buffer = bytearray(2048)
with open(filename) as f:
  while f.readinto(buffer) > 0:
      lines += buffer.count('\n')

您可以调整缓冲区大小，可能会看到一些改进。

大文件的另一种选择是使用xreadlines():

count = 0
for line in open(thefilepath).xreadlines(  ): count += 1

对于Python 3，请参阅:在Python 3中什么替代xreadlines() ?

对我来说，这个变体是最快的:

#!/usr/bin/env python

def main():
    f = open('filename')                  
    lines = 0
    buf_size = 1024 * 1024
    read_f = f.read # loop optimization

    buf = read_f(buf_size)
    while buf:
        lines += buf.count('\n')
        buf = read_f(buf_size)

    print lines

if __name__ == '__main__':
    main()

原因:缓冲比逐行和逐字符串读取快。计数也非常快

使用Numba

我们可以使用Numba来JIT(及时)编译我们的函数到机器代码。Def numbacountparallel(fname)运行速度快2.8倍 然后从问题中定义file_len(fname)。

注:

在运行基准测试之前，操作系统已经将文件缓存到内存中，因为我在我的PC上没有看到太多的磁盘活动。 第一次读取文件时，时间会慢得多，因此使用Numba的时间优势并不显著。

第一次调用函数时，JIT编译需要额外的时间。

如果我们不只是计算行数，这个就很有用了。

Cython是另一个选择。

http://numba.pydata.org/

结论

因为计算行数是IO绑定的，所以使用问题中的def file_len(fname)，除非你想做的不仅仅是计算行数。

import timeit

from numba import jit, prange
import numpy as np

from itertools import (takewhile,repeat)

FILE = '../data/us_confirmed.csv' # 40.6MB, 371755 line file
CR = ord('\n')


# Copied from the question above. Used as a benchmark
def file_len(fname):
    with open(fname) as f:
        for i, l in enumerate(f):
            pass
    return i + 1


# Copied from another answer. Used as a benchmark
def rawincount(filename):
    f = open(filename, 'rb')
    bufgen = takewhile(lambda x: x, (f.read(1024*1024*10) for _ in repeat(None)))
    return sum( buf.count(b'\n') for buf in bufgen )


# Single thread
@jit(nopython=True)
def numbacountsingle_chunk(bs):

    c = 0
    for i in range(len(bs)):
        if bs[i] == CR:
            c += 1

    return c


def numbacountsingle(filename):
    f = open(filename, "rb")
    total = 0
    while True:
        chunk = f.read(1024*1024*10)
        lines = numbacountsingle_chunk(chunk)
        total += lines
        if not chunk:
            break

    return total


# Multi thread
@jit(nopython=True, parallel=True)
def numbacountparallel_chunk(bs):

    c = 0
    for i in prange(len(bs)):
        if bs[i] == CR:
            c += 1

    return c


def numbacountparallel(filename):
    f = open(filename, "rb")
    total = 0
    while True:
        chunk = f.read(1024*1024*10)
        lines = numbacountparallel_chunk(np.frombuffer(chunk, dtype=np.uint8))
        total += lines
        if not chunk:
            break

    return total

print('numbacountparallel')
print(numbacountparallel(FILE)) # This allows Numba to compile and cache the function without adding to the time.
print(timeit.Timer(lambda: numbacountparallel(FILE)).timeit(number=100))

print('\nnumbacountsingle')
print(numbacountsingle(FILE))
print(timeit.Timer(lambda: numbacountsingle(FILE)).timeit(number=100))

print('\nfile_len')
print(file_len(FILE))
print(timeit.Timer(lambda: rawincount(FILE)).timeit(number=100))

print('\nrawincount')
print(rawincount(FILE))
print(timeit.Timer(lambda: rawincount(FILE)).timeit(number=100))

每个函数调用100次的时间(以秒为单位)

numbacountparallel
371755
2.8007332000000003

numbacountsingle
371755
3.1508585999999994

file_len
371755
6.7945494

rawincount
371755
6.815438