how to read extreme long lines from text file fast and safe in C++?

Question

There is a large text file of 6.53 GiB. Each line of it can be a data line or comment line. Comment lines are usually short, less than 80 characters, while a data line contains more than 2 million characters and is variable-length.

Considering each data line needs to be dealt with as a unit, is there a simple way to read lines safe and fast in C++?

safe (safe for variable-length data lines): The solution is as easy to use as std::getline(). Since the length is changing, it is hoped to avoid extra memory management.

fast: The solution can achieve as fast as readline() in python 3.6.0, or even as fast as fgets() of stdio.h.

A Pure C solution is welcomed. The interface for further processing is provided both in C and C++.

---

UPDATE 1: Thanks to short but invaluable comment from Basile Starynkevitch, the perfect solution comes up: POSIX getline(). Since further processing only involves converting from character to number and does not use many features of string class, a char array would be sufficient in this application.

---

UPDATE 2: Thanks to comments from Zulan and Galik, who both report comparable performance among std::getline(), fgets() and POSIX getline(), another possible solution is to use a better standard library implementation such as libstdc++. Moreover, here is a report claiming that the Visual C++ and libc++ implementations of std::getline is not well optimised.

Moving from libc++ to libstdc++ changes the results a lot. With libstdc++ 3.4.13 / Linux 2.6.32 on a different platform, POSIX getline(), std::getline() and fgets() show comparable performance. At the beginning, codes were run under the default settings of clang in Xcode 8.3.2 (8E2002), thus libc++ is used.

---

More details and some efforts (very long):

getline() of <string> can handle arbitrary long lines but is a bit slow. Is there an alternative in C++ for readline() in python?

// benchmark on Mac OS X with libc++ and SSD:
readline() of python                         ~550 MiB/s

fgets() of stdio.h, -O0 / -O2               ~1100 MiB/s

getline() of string, -O0                      ~27 MiB/s
getline() of string, -O2                     ~150 MiB/s
getline() of string + stack buffer, -O2      ~150 MiB/s

getline() of ifstream, -O0 / -O2             ~240 MiB/s
read() of ifstream, -O2                      ~340 MiB/s

wc -l                                        ~670 MiB/s

cat data.txt | ./read-cin-unsync              ~20 MiB/s

getline() of stdio.h (POSIX.1-2008), -O0    ~1300 MiB/s

• Speeds are rounded very roughly, only to show the magnitude, and all code blocks are run several times to assure that the values are representative.

• '-O0 / -O2' means the speeds are very similar for both optimization levels

• Codes are shown as follows.

---

readline() of python

# readline.py

import time
import os

t_start = time.perf_counter()

fname = 'data.txt'
fin = open(fname, 'rt')

count = 0

while True:
    l = fin.readline()
    length = len(l)
    if length == 0:     # EOF
        break
    if length > 80:     # data line
        count += 1

fin.close()

t_end = time.perf_counter()
time = t_end - t_start

fsize = os.path.getsize(fname)/1024/1024   # file size in MiB
print("speed: %d MiB/s" %(fsize/time))
print("reads %d data lines" %count)

# run as `python readline.py` with python 3.6.0

---

fgets() of stdio.h

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <string.h>

int main(int argc, char* argv[]){
  clock_t t_start = clock();

  if(argc != 2) {
    fprintf(stderr, "needs one input argument\n");
    return EXIT_FAILURE;
  }

  FILE* fp = fopen(argv[1], "r");
  if(fp == NULL) {
    perror("Failed to open file");
    return EXIT_FAILURE;
  }

  // maximum length of lines, determined previously by python
  const int SIZE = 1024*1024*3;
  char line[SIZE];

  int count = 0;
  while(fgets(line, SIZE, fp) == line) {
    if(strlen(line) > 80) {
      count += 1;
    }
  }

  clock_t t_end = clock();

  const double fsize = 6685;  // file size in MiB

  double time = (t_end-t_start) / (double)CLOCKS_PER_SEC;

  fprintf(stdout, "takes %.2f s\n", time);
  fprintf(stdout, "speed: %d MiB/s\n", (int)(fsize/time));
  fprintf(stdout, "reads %d data lines\n", count);

  return EXIT_SUCCESS;
}

---

getline() of <string>

// readline-string-getline.cpp
#include <string>
#include <fstream>
#include <iostream>
#include <ctime>
#include <cstdlib>

using namespace std;

int main(int argc, char* argv[]) {
  clock_t t_start = clock();

  if(argc != 2) {
    fprintf(stderr, "needs one input argument\n");
    return EXIT_FAILURE;
  }

  // manually set the buffer on stack
  const int BUFFERSIZE = 1024*1024*3;   // stack on my platform is 8 MiB
  char buffer[BUFFERSIZE];
  ifstream fin;
  fin.rdbuf()->pubsetbuf(buffer, BUFFERSIZE);
  fin.open(argv[1]);

  // default buffer setting
  // ifstream fin(argv[1]);

  if(!fin) {
    perror("Failed to open file");
    return EXIT_FAILURE;
  }

  // maximum length of lines, determined previously by python
  const int SIZE = 1024*1024*3;
  string line;
  line.reserve(SIZE);

  int count = 0;
  while(getline(fin, line)) {
    if(line.size() > 80) {
      count += 1;
    }
  }

  clock_t t_end = clock();

  const double fsize = 6685;  // file size in MiB

  double time = (t_end-t_start) / (double)CLOCKS_PER_SEC;

  fprintf(stdout, "takes %.2f s\n", time);
  fprintf(stdout, "speed: %d MiB/s\n", (int)(fsize/time));
  fprintf(stdout, "reads %d data lines\n", count);

  return EXIT_SUCCESS;
}

---

getline() of ifstream

// readline-ifstream-getline.cpp
#include <fstream>
#include <iostream>
#include <ctime>
#include <cstdlib>

using namespace std;

int main(int argc, char* argv[]) {
  clock_t t_start = clock();

  if(argc != 2) {
    fprintf(stderr, "needs one input argument\n");
    return EXIT_FAILURE;
  }

  ifstream fin(argv[1]);
  if(!fin) {
    perror("Failed to open file");
    return EXIT_FAILURE;
  }

  // maximum length of lines, determined previously by python
  const int SIZE = 1024*1024*3;
  char line[SIZE];

  int count = 0;
  while(fin.getline(line, SIZE)) {
    if(strlen(line) > 80) {
      count += 1;
    }
  }

  clock_t t_end = clock();

  const double fsize = 6685;  // file size in MiB

  double time = (t_end-t_start) / (double)CLOCKS_PER_SEC;

  fprintf(stdout, "takes %.2f s\n", time);
  fprintf(stdout, "speed: %d MiB/s\n", (int)(fsize/time));
  fprintf(stdout, "reads %d data lines\n", count);

  return EXIT_SUCCESS;
}

---

read() of ifstream

// seq-read-bin.cpp
// sequentially read the file to see the speed upper bound of
// ifstream

#include <iostream>
#include <fstream>
#include <ctime>

using namespace std;


int main(int argc, char* argv[]) {
  clock_t t_start = clock();

  if(argc != 2) {
    fprintf(stderr, "needs one input argument\n");
    return EXIT_FAILURE;
  }

  ifstream fin(argv[1], ios::binary);

  const int SIZE = 1024*1024*3;
  char str[SIZE];

  while(fin) {
    fin.read(str,SIZE);
  }

  clock_t t_end = clock();
  double time = (t_end-t_start) / (double)CLOCKS_PER_SEC;

  const double fsize = 6685;  // file size in MiB

  fprintf(stdout, "takes %.2f s\n", time);
  fprintf(stdout, "speed: %d MiB/s\n", (int)(fsize/time));

  return EXIT_SUCCESS;
}

---

use cat, then read from cin with cin.sync_with_stdio(false)

#include <iostream>
#include <ctime>
#include <cstdlib>

using namespace std;

int main(void) {
  clock_t t_start = clock();

  string input_line;

  cin.sync_with_stdio(false);

  while(cin) {
    getline(cin, input_line);
  }

  double time = (clock() - t_start) / (double)CLOCKS_PER_SEC;

  const double fsize = 6685;  // file size in MiB

  fprintf(stdout, "takes %.2f s\n", time);
  fprintf(stdout, "speed: %d MiB/s\n", (int)(fsize/time));

  return EXIT_SUCCESS;
}

---

POSIX getline()

// readline-c-getline.c
#include <stdio.h>
#include <stdlib.h>
#include <time.h>

int main(int argc, char *argv[]) {

  clock_t t_start = clock();

  char *line = NULL;
  size_t len = 0;
  ssize_t nread;

  if (argc != 2) {
    fprintf(stderr, "Usage: %s <file>\n", argv[1]);
    exit(EXIT_FAILURE);
  }

  FILE *stream = fopen(argv[1], "r");
  if (stream == NULL) {
    perror("fopen");
    exit(EXIT_FAILURE);
  }

  int length = -1;
  int count = 0;
  while ((nread = getline(&line, &len, stream)) != -1) {
    if (nread > 80) {
      count += 1;
    }
  }

  free(line);
  fclose(stream);

  double time = (clock() - t_start) / (double)CLOCKS_PER_SEC;
  const double fsize = 6685;  // file size in MiB
  fprintf(stdout, "takes %.2f s\n", time);
  fprintf(stdout, "speed: %d MiB/s\n", (int)(fsize/time));
  fprintf(stdout, "reads %d data lines.\n", count);
  // fprintf(stdout, "length of MSA: %d\n", length-1);

  exit(EXIT_SUCCESS);
}

Answer 1

Well, the C standard library is a subset of the C++ standard library. From n4296 draft from C++ 2014 standard:

17.2 The C standard library [library.c]

The C++ standard library also makes available the facilities of the C standard library, suitably adjusted to ensure static type safety.

So provided you explain in a comment that a performance bottleneck requires it, it is perfectly fine to use fgets in a C++ program - simply you should carefully encapsulate it in an utility class, in order to preserve the OO high level structures.

Answer 2

As I commented, on Linux & POSIX systems, you could consider using getline(3); I guess that the following could compile both as C and as C++ (assuming you do have some valid fopen-ed FILE*fil; ...)

char* linbuf = NULL; /// or nullptr in C++
size_t linsiz = 0;
ssize_t linlen = 0;

while((linlen=getline(&linbuf, &linsiz,fil))>=0) {
  // do something useful with linbuf; but no C++ exceptions
}
free(linbuf); linsiz=0;

I guess this might work (or be easily adapted) to C++. But then, beware of C++ exceptions, they should not go thru the while loop (or you should ensure that an appropriate destructor or catch is doing free(linbuf);).

Also getline could fail (e.g. if it calls a failing malloc) and you might need to handle that failure sensibly.

Answer 3

Yes, there's a faster way to read lines and create strings.

Query the file size, then load it into a buffer. Then iterate over the buffer replacing the newlines with nuls and storing the pointer to the next line.

It will be quite a bit faster if, as is likely, your platform has a call to load a file into memory.