Cin-Cout vs Scanf-Printf

Regular competitive programmers face common challenge when input is large and the task of reading such an input from stdin might prove to be a bottleneck. Such problem is accompanied with “Warning: large I/O data”. Let us create a dummy input file containing a line with 16 bytes followed by a newline and having 1000000 such lines, making a file of 17MB should be good enough.

// Creating a dummy file of size 17 MB to compare 
// performance of scanf() and cin()
$ yes 1111111111111111 | head -1000000 > tmp/dummy

Let us compare the time taken to read the file from stdin (get the file from disk to stdin using redirection) by using scanf() versus cin. 

CPP

// Filename : cin_test.cc to test the
// We redirect above created temp file
// of 17 MB to stdin when this program
// is run.
#include<iostream>
using namespace std;

int main()
{
    char buffer[256];
    while (cin >> buffer)
    {
    }
    return 0;
}

Output of above program when dummy file is redirected to stdin.

$ g++ cin_test.cc –o cin_test
$ time ./cin_test < /tmp/dummy
real     0m2.162s
user     0m1.696s
sys     0m0.332s

CPP

// Filename : scanf_test.c to see
// performance of scanf()
// We redirect above created temp file
// of 17 MB to stdin when this program
// is run.
#include<cstdlib>
#include<cstdio>
int main()
{
    char buffer[256];
    while (scanf("%s", buffer) != EOF)
    {
    }
    return 0;
}

Output of above program when dummy file is redirected to stdin.

$ g++ scanf_test.cc –o scanf_test
$ time ./scanf_test < /tmp/dummy
real     0m0.426s
user     0m0.248s
sys     0m0.084s

Well, the above results are consistent with our observations. Why is scanf faster than cin? On a high level both of them are wrappers over theread() system call, just syntactic sugar. The only visible difference is that scanf() has to explicitly declare the input type, whereas cin has the redirection operation overloaded using templates. This does not seem like a good enough reason for a performance hit of 5x. It turns out that iostream makes use of stdio‘s buffering system. So, cin wastes time synchronizing itself with the underlying C-library’s stdio buffer, so that calls to bothscanf()and cin can be interleaved. The good thing is that libstdc++ provides an option to turn off synchronization of all the iostream standard streams with their corresponding standard C streams using

std::ios_base::sync_with_stdio(false);

and cin becomes faster than scanf() as it should have been. A detailed article on Fast Input Output in Competitive Programming

Cpp

// Filename : cin_test_2.cc to see
// performance of cin() with stdio syc
// disabled using sync_with_stdio(false).
#include<iostream>
using namespace std;

int main()
{
    char buffer[256];
    ios_base::sync_with_stdio(false);

    while (cin >> buffer)
    {

    }
    return 0;
}

Running the program :

$ g++ cin_test_2.cc –o cin_test_2
$ time./cin_test_2 </tmp/dummy
real    0m0.380s
user   0m0.240s
sys    0m0.028s 

• Like with all things, there is a caveat here. With synchronization turned off, using cin and scanf() together will result in an undefined mess.
• With synchronization turned off, the above results indicate that cin is 8-10% faster than scanf(). This is probably because scanf() interprets the format arguments at runtime and makes use of variable number of arguments, whereas cin does this at compile time.

Now wondering, how fast can it be done?

// Redirecting contents of dummy file to null
// device (a special device that discards the
// information written to it) using command line.
$time cat /tmp/dummy > /dev/null
real   0m0.185s
user   0m0.000s
sys    0m0.092s

Wow! This is fast!!!