dd-tests.md

dd as an I/O benchmarking tool

dd does have different performance on different systems. See below for code that generates an in-memory 1G file not touching the file system. On one system the dd performance is 2.5x faster than the other.

dd workflow:

1. allocate buffers (+memset in some cases)
2. read data into input buffer from input file
3. copy data into output buffer
4. write data to output file

In general dd should never be used for I/O performance testing because what is being measured is the cost of the following operations:

1. malloc
2. mempcpy
3. reading from input file
4. iterating over the number of blocks
5. ...
6. writing into output file

It can however be used to spot performance differences on the same system over time.

Have a close look at the dd_copy function in the dd source file.

E.g. from dd.c:

//READ into memory

/* Wrapper around iread function to accumulate full blocks.  */
static ssize_t
iread_fullblock (int fd, char *buf, size_t size)
{
  ssize_t nread = 0;

  while (0 < size)
  {
    ssize_t ncurr = iread(fd, buf, size); //invokes read()
    if (ncurr < 0)
      return ncurr;
    if (ncurr == 0)
      break;
    nread += ncurr;
    buf   += ncurr;
    size  -= ncurr;
  }

  return nread;
}

//COPY and WRITE

static void
copy_simple (char const *buf, size_t nread)
{
  const char *start = buf;	/* First uncopied char in BUF.  */

  do
    {
      size_t nfree = MIN (nread, output_blocksize - oc);

      memcpy (obuf + oc, start, nfree);

      nread -= nfree;		/* Update the number of bytes left to copy. */
      start += nfree;
      oc += nfree;
      if (oc >= output_blocksize)
        write_output ();
    }
  while (nread != 0);
}

A minimal test of the actual performance of dd on two separate systems follows. 1 GiB of data is read from /dev/zero and written into RAM (/dev/shm). As mentioned above the measured performance is very different.

#!/usr/bin/env bash

mkdir /dev/shm/ugo
dd if=/dev/zero of=/dev/shm/ugo/test bs=1G count=1
rm -rf /dev/shm/ugo

1 (zeus.pawsey.org.au)

uvaretto@zeus-1:~/projects/scratch> srun ./dd.sh 
srun: job 4880258 queued and waiting for resources
srun: job 4880258 has been allocated resources
1+0 records in
1+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 0.672728 s, 1.6 GB/s

2 (magnus.pawsey.org.au)

uvaretto@nid00672:~/projects/scratch> srun ./dd.sh 
1+0 records in
1+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 1.54844 s, 693 MB/s