为什么使用ddstream从管道中读取“ dd”要比我自己的程序快？

Question

I have two programs that pass data to each other via linux pipes (named or otherwise). I need to hit a transfer rate of ~2600 MB/s between the two programs, but am currently seeing a slower rate of about ~2200 MB/s. However, I found that if I replace my 2nd process with 'dd' instead, the transfer rate jumps to over 3000 MB/s. Is there something about the way my program is reading from the pipe that is less efficient than the way 'dd' does it? What can I do to improve this throughput? Is 'ifstream' inherently slower than other methods of reading binary data from pipe?

To summarize the two scenarios:

Scenario 1:

Program 1 -> [named pipe] -> Program 2

Yields ~2200 MB/s transfer rate

Scenario2:

Program 1 -> [named pipe] -> 'dd if=pipename of=/dev/null bs=8M'

Yields ~3000 MB/s transfer rate.

Here is the way my Program 2 currently reads from pipe:

ifstream inputFile;
inputFile.open(inputFileName.c_str(), ios::in | ios::binary);
while (keepLooping)
{
    inputFile.read(&buffer[0], 8*1024*1024);
    bytesRead = inputFile.gcount();
    //Do something with data
}

Update:

I have now tried using 'read(fd, &buffer[0], 8*1024*1024)' instead of istream, seemed to show a mild improvement (but not as much as dd)

I also tried using stream->rdbuf()->sgetn(&buffer[0], 8*1024*1024) instead of stream->read(), which did not help.

Answer 1

The difference appears to be due to using an array instead of std::vector, which I still have a hard time believing. My two sets of code are shown below for comparison. The first can ingest from Program 1 at a rate of about 2500 MB/s. The second can ingest at a rate of 3100 MB/s.

Program 1 (2500 MB/s)

int main(int argc, char **argv)
{
    int fd = open("/tmp/fifo2", O_RDONLY);

    std::vector<char> buf(8*1024*1024);

    while(1)
    {
       read(fd, &buf[0], 8*1024*1024);
    }
}

Program 2 (3100 MB/s)

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

    int fd = open("/tmp/fifo2", O_RDONLY);

    char buf[8*1024*1024];

    while(1)
    {
       read(fd, &buf[0], 8*1024*1024);
    }
}

Both are compiled with -O3 using gcc version 4.4.6. If anyone can explain the reason for this I'd be very interested (since I understand std::vector to basically be a wrapper around an array).

Edit : I just tested Program 3, below, that can uses ifstream and runs at 3000 MB/s. So it appears that using ifstream instead of 'read()' incurs a very slight performance degradation. Much less than the hit taken from using std::vector.

Program 3 (3000 MB/s)

int main(int argc, char **argv)
{
    ifstream file("/tmp/fifo2", ios::in | ios::binary);

    char buf[8*1024*1024];

    while(1)
    {
       file.read(&buf[0], 32*1024);
    }
}

Edit 2:

I modded Program 2's code to use malloc'd memory instead of memory on the stack and the performance dropped to match the vector performance. Thanks, ipc, for keying me onto this.

Answer 2

This code compiled with g++ -Ofast :

int main(int argc, char *argv[])
{
  if (argc != 2) return -1;
  std::ifstream in(argv[1]);
  std::vector<char> buf(8*1024*1024);
  in.rdbuf()->pubsetbuf(&buf[0], buf.size());
  std::ios_base::sync_with_stdio(false);
  std::cout << in.rdbuf();
}

does not perform that bad at all.

$ time <<this program>> <<big input file>> >/dev/null
0.20s user 3.50s system 9% cpu 40.548 total
$ time dd if=<<big input file>> bs=8M > /dev/null
0.01s user 3.84s system 9% cpu 40.786 total

You have to consider that std::cout shares a buffer with stdout which is really time consuming if not switched off. So call std::ios_base::sync_with_stdio(false); if you want speed and do not intend to use C's input output methods (which are slower anyway).

Also, for raw and fast input/output in C++, use the methods from streambuf , obtained by rdbuf() .