Re: Running a separate fio process for each disk?

[Jens Axboe <axboe@kernel.dk>]

And finally, there's a potential fix, if you run commit
99afcdb53dc3 or later. So please do try that as well, and
see if that behaves any better for you.


On 11/20/2015 05:03 PM, Jens Axboe wrote:
> Hi,
>
> OK, I see. Can you pull the latest -git, and then run fio
> --cpuclock-test on one of the boxes where you see the issue? It should
> have commit 5896d827e1e2 or later.
>
>
> On Fri, Nov 20, 2015 at 3:20 PM, Akash Verma <akashv@google.com
> <mailto:akashv@google.com>> wrote:
>
>     Hi Jens,
>     The issue is not seen with non-cpu clock sources, or when using a
>     single process (with individual threads, the only config I tried). We
>     only see the issue when using multiple processes and the cpu clock
>     source.
>
>     On Fri, Nov 20, 2015 at 11:50 AM, Jens Axboe <axboe@kernel.dk
>     <mailto:axboe@kernel.dk>> wrote:
>      > On 11/20/2015 12:37 PM, Caio Villela wrote:
>      >>
>      >> Hello Allen and Jens,
>      >>
>      >> Sorry for the long output, this is just in case you want the
>     details.
>      >> Here is a simple explanation for the problem. I want to run a 15
>     minute
>      >> random write, using 1 Meg requests, and measure throughput and
>     latency.
>      >> What seems to be the problem is that if the test system has a large
>      >> number of drives - the system that I am testing here has 28 drives -
>      >> then the time accounting seems to go bad for some of the processes.
>      >> What you see below is that during the 15 minutes from start, all
>     disks
>      >> are getting hit the same, as they should. Then, after 15
>     minutes, there
>      >> are 15 drives that are still running.... after 5 minutes over the
>      >> specified 15 minutes, there is still one drive running. Then
>     looking at
>      >> the amount of IOs sent to each drive, the ones that ran on that
>     excess
>      >> time have much more IOs. FIO still reports that all drives ran
>     for 15
>      >> minutes, although some ran for more than 20 minutes.
>      >>
>      >> We will attempt to run a single process instead of 28 instances
>     of FIO
>      >> to see if this goes away.
>      >
>      >
>      > Could you also check if adding clocksource=gettimeofday makes any
>      > difference? This sounds very odd.
>      >
>      > Assuming this was run with fio -git?
>      >
>      >
>      > --
>      > Jens Axboe
>      >
>     > --
>     > To unsubscribe from this list: send the line "unsubscribe fio" in
>     > the body of a message tomajordomo@vger.kernel.org <mailto:majordomo@vger.kernel.org>
>     > More majordomo info athttp://vger.kernel.org/majordomo-info.html
>
>


-- 
Jens Axboe