How to use 'dd' to benchmark your disk or CPU?

Storage benchmark (hard disk or SSD)

This part is written on 2010-11-29 to address sometimes incorrect usage of “dd” to measure disk write performance on a VPS by some visitors of the lowendbox.com website, and is originally based on this question and my answer to it.

Ways in which you can invoke 'dd' to test the write speed:

  1. dd bs=1M count=256 if=/dev/zero of=test
  2. dd bs=1M count=256 if=/dev/zero of=test; sync
  3. dd bs=1M count=256 if=/dev/zero of=test conv=fdatasync
  4. dd bs=1M count=256 if=/dev/zero of=test oflag=dsync

What is the difference between those?

  1. dd bs=1M count=256 if=/dev/zero of=test
    The default behaviour of dd is to not “sync” (i.e. not ask the OS to completely write the data to disk before dd exiting). The above command will just commit your 128 MB of data into a RAM buffer (write cache) – this will be really fast and it will show you the hugely inflated benchmark result right away. However, the server in the background is still busy, continuing to write out data from the RAM cache to disk.
  2. dd bs=1M count=256 if=/dev/zero of=test; sync
    Absolutely identical to the previous case, as anyone who understands how *nix shell works should surely know that adding a ; sync does not affect the operation of previous command in any way, because it is executed independently, after the first command completes. So your (inflated) MB/sec value is already printed on screen while that sync is only preparing to be executed.
  3. dd bs=1M count=256 if=/dev/zero of=test conv=fdatasync
    This tells dd to require a complete “sync” once, right before it exits. So it commits the whole 256 MB of data, then tells the operating system: “OK, now ensure this is completely on disk”, only then measures the total time it took to do all that and calculates the benchmark result.
  4. dd bs=1M count=256 if=/dev/zero of=test oflag=dsync
    Here dd will ask for completely synchronous output to disk, i.e. ensure that its write requests don’t even return until the submitted data is on disk. In the above example, this will mean sync'ing once per megabyte, or 256 times in total. It will be the slowest mode, as the write cache is basically unused at all in this case.

Which one do you recommend to use?

This behaviour is perhaps the closest to the way real-world tasks behave:

  • dd bs=1M count=256 if=/dev/zero of=test conv=fdatasync

If your server or VPS is really fast and the above test completes in a second or less, try increasing the count= number to 1024 or so, to get a more accurate averaged result.

CPU benchmark

Perhaps not many people use this, but 'dd' in conjunction with any stream-processing CPU-intensive program can also be used as a simple CPU benchmark! It may be not very accurate, but the huge advantage is that it doesn't require installing any additional software whatsoever, and typically you can run this “out of the box” on any GNU/Linux system. The usage is as follows:

  • dd if=/dev/zero bs=1M count=1024 | md5sum

In this case I used the md5sum program, which calculates the MD5 hash of data that is fed to it. In effect, 'dd' here fetches 1 GB of zeroes from the Linux kernel, feeds that to md5sum, and then prints how fast in MB/sec that was processed. So for example on a modern 3.6 GHz AMD Phenom II CPU:

$ dd if=/dev/zero bs=1M count=1024 | md5sum
1024+0 records in
1024+0 records out
1073741824 bytes (1.1 GB) copied, 2.28735 s, 469 MB/s
cd573cfaace07e7949bc0c46028904ff  -

And on a VPS with a 2.1 GHz Opteron:

# dd if=/dev/zero bs=1M count=1024 | md5sum
1024+0 records in
1024+0 records out
1073741824 bytes (1.1 GB) copied, 5.14561 s, 209 MB/s
cd573cfaace07e7949bc0c46028904ff  -

Generally on any modern 2.0+ GHz CPU you should be looking at a result of 200 MB/sec or more in this test. If you see very low results, like 20-40-50-100 MB/sec, it's a sure sign that whatever system you're running this on is either overloaded CPU-wise, or is hard-limiting your CPU allowance to only some portion of a full CPU core.


dd-benchmark.txt · Last modified: 2014-04-24 05:38 UTC by rm