[lkml]   [2019]   [Jul]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: iowait v.s. idle accounting is "inconsistent" - iowait is too low
On 05/07/2019 12:38, Peter Zijlstra wrote:
> On Fri, Jul 05, 2019 at 12:25:46PM +0100, Alan Jenkins wrote:
>> Hi, scheduler experts!
>> My cpu "iowait" time appears to be reported incorrectly.  Do you know why
>> this could happen?
> Because iowait is a magic random number that has no sane meaning.
> Personally I'd prefer to just delete the whole thing, except ABI :/
> Also see the comment near nr_iowait():
> /*
> * IO-wait accounting, and how its mostly bollocks (on SMP).
> *
> * The idea behind IO-wait account is to account the idle time that we could
> * have spend running if it were not for IO. That is, if we were to improve the
> * storage performance, we'd have a proportional reduction in IO-wait time.
> *
> * This all works nicely on UP, where, when a task blocks on IO, we account
> * idle time as IO-wait, because if the storage were faster, it could've been
> * running and we'd not be idle.
> *
> * This has been extended to SMP, by doing the same for each CPU. This however
> * is broken.
> *
> * Imagine for instance the case where two tasks block on one CPU, only the one
> * CPU will have IO-wait accounted, while the other has regular idle. Even
> * though, if the storage were faster, both could've ran at the same time,
> * utilising both CPUs.
> *
> * This means, that when looking globally, the current IO-wait accounting on
> * SMP is a lower bound, by reason of under accounting.
> *
> * Worse, since the numbers are provided per CPU, they are sometimes
> * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
> * associated with any one particular CPU, it can wake to another CPU than it
> * blocked on. This means the per CPU IO-wait number is meaningless.
> *
> * Task CPU affinities can make all that even more 'interesting'.
> */

Thanks. I take those as being different problems, but you mean there is
not much demand (or point) to "fix" my issue.

> (2) Compare running "dd" with "taskset -c 1":
> %Cpu1  :  0.3 us,  3.0 sy,  0.0 ni, 83.7 id, 12.6 wa,  0.0 hi,  0.3 si,  0.0 st

^ non-zero idle time for Cpu1, despite the pinned IO hog.

The block layer recently decided they could break "disk busy%" reporting
for slow devices (mechanical HDD), in order to reduce overheads for fast
devices.  This means the summary view in "atop" now lacks any reliable

I suppose I need to look in "iotop".

The new /proc/pressure/io seems to have caveats related to the iowait
issues... it seems even more complex to interpret for this case, and it
does not seem to work how I think it does.[1]



 \ /
  Last update: 2019-07-05 15:38    [W:0.047 / U:0.412 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site