My disk often is utilized, but top (and htop, a custom replacement) show nothing suspicious.

Is there a way to sort processes by I/O (more specific: disk) utilization?

Found out using iotop that those strange processes are flush-8:16 and jbd2/sdb3-7. Seems to have to do with usual filesystem operations.


Have you tried iotop ?

You may need to install it before. Also, it depends on a kernel feature that may or may not enabled in your specific distribution.

  • 1
    Works on Mac OS X, too. Not sure that it is going to help with the problem at hand, though. Mar 2 '10 at 16:27
  • 4
    wow. That's fantastic. 15+ years of admin work and I've never run into that. Thanks!
    – skarface
    Mar 2 '10 at 17:06
  • 3
    Root access needed, FYI. Jan 24 '15 at 17:27

You might want to give atop a try. It seems to do a good job of letting you know what is going on.

  • 2
    Seems to be better than iotop because it also shows IO busy percent.
    – sorin
    Mar 18 '14 at 13:36
  • 2
    Also no root or suid needed.
    – Rolf
    Feb 25 '18 at 21:08

iostat is still the king of detailed I/O information, the kind of info that can confirm if your SSD is living up to expectations, for example.

$ iostat -xht 5 nvme0n1
Linux 4.15.0-43-generic (myhost)    06/02/2021  _x86_64_    (8 CPU)

06/02/2021 08:59:20 AM
avg-cpu:  %user   %nice %system %iowait  %steal   %idle
           2.6%    0.1%    0.8%    0.1%    0.0%   96.4%

Device            r/s     w/s     rkB/s     wkB/s   rrqm/s   wrqm/s  %rrqm  %wrqm r_await w_await aqu-sz rareq-sz wareq-sz  svctm  %util
                95.03   20.80      1.8M    300.1k     0.00    13.15   0.0%  38.7%    0.24    0.57   0.03    19.2k    14.4k   0.02   0.3%

Every five seconds, this will print detailed IO stats for the NVMe drive. The most important ones are usually not actually read/write bandwidth -- rKB/s and wKB/s -- but rather the reads and writes per second (r/s and w/s, aka IOPS) and, perhaps most important, the average time a process waited for each read and write -- r_await and w_await (in milliseconds).

All of these values feed into each other using the surprisingly useful Little's law formula, translated a bit to latency = queue_size / IOPS, or await = aqu_sz / (r/s + w/s).

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.