Re: FIO creating empty latency log files

[Jens Axboe <axboe@kernel.dk>]

On 10/28/2015 09:20 AM, Jens Axboe wrote:
> On 10/28/2015 06:13 AM, Jens Axboe wrote:
>> On 10/28/2015 12:41 AM, Buttetsu Batou wrote:
>>> On 06/04/2015 05:09 PM, Reed Crowe wrote:
>>>> Using this version of FIO:
>>>> * io-threads
>>>>      master
>>>> rcrowe@rcrowe-desktop:~/fio_iothreads/fio$ ./fio --version
>>>> fio-2.2.7-14-g00b2
>>>>
>>>> I have a job file which looks like this:
>>>>
>>>> [global]
>>>> ioengine=libaio
>>>> direct=1
>>>> numjobs=1
>>>> bs=4k
>>>> iodepth=300
>>>> rw=randwrite
>>>> randrepeat=0
>>>> io_submit_mode=offload
>>>> rate_iops=8070
>>>> time_based=1
>>>> runtime=1m
>>>>
>>>> [slow_write]
>>>> filename=/mnt/test1/testfile.io
>>>> write_lat_log=/tmp/fio_lat
>>>>
>>>> And it produces empty files:
>>>> -rw-r--r-- 1 root   root         0 Jun  4 14:34 fio_lat_clat.1.log
>>>> -rw-r--r-- 1 root   root         0 Jun  4 14:34 fio_lat_lat.1.log
>>>> -rw-r--r-- 1 root   root         0 Jun  4 14:34 fio_lat_slat.1.log
>>>>
>>>> What am I missing?
>>>
>>> I am seeing this behavior in the latest release fio-2.2.10 and in
>>> master fio-2.2.10-29-g8a68 but I am not using offload.
>>>
>>> Is anyone currently able to successfully generate log output? If so
>>> could you share your fio version and your config?
>>>
>>> Here is the job file I am using at the moment.. have tried variations
>>> on these settings and different fio versions back to 2.0.15 but always
>>> the same empty output:
>>>
>>> [write]
>>> group_reporting=1
>>> numjobs=8
>>> ioengine=posixaio
>>> io_submit_mode=inline
>>> stonewall
>>> rw=write
>>> bs=32k
>>> time_based=1
>>> runtime=10
>>> filesize=64k
>>> directory=/tmp
>>> per_job_logs=1
>>> write_bw_log=write_bw
>>> write_iops_log=write_iops
>>>
>>> Updated the issue reported here:
>>>
>>> https://github.com/axboe/fio/issues/29
>>>
>>> Thank you for any ideas you may have!
>>
>> I think your issue is because you set filesize=64k. Your job basically
>> sets up 8 files of 64k each, then writes to them for 10 seconds. This
>> means that the fio main loop keeps re-entering for every 64k for each
>> thread. The bandwidth/iops is averaged over a window of 500msec by
>> default, and the main loop runs for much shorter than that since it only
>> does 64k each time. Hence you never get logged any entries.
>>
>> Now, it should work. If you comment out these two lines:
>>
>> memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
>> memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
>>
>> in backend.c around 1568/1569, then I suspect the logging should improve
>> for you. I'll have to see if we can safely do that without making other
>> changes, before committing a change like that.t
>
> Try the attached patch, we need to retain the stat block and byte counts
> too.

Just run current -git, I committed a similar patch.

-- 
Jens Axboe