Re: [PATCH v3 1/4] btrfs: add read_policy latency

[Anand Jain <anand.jain@oracle.com>]

On 20/1/21 9:54 pm, Michal Rostecki wrote:
> On Wed, Jan 20, 2021 at 08:30:56PM +0800, Anand Jain wrote:
>>   I ran fio tests again, now with dstat in an another window. I don't
>>   notice any such stalls, the read numbers went continuous until fio
>>   finished. Could you please check with the below fio command, also
>>   could you please share your fio command options.
> 
> That's the fio config I used:
> 
> https://gitlab.com/vadorovsky/playground/-/blob/master/fio/btrfs-raid1-seqread.fio
> 
> The main differences seem to be:
> - the number of jobs (I used the number of CPU threads)
> - direct vs buffered I/O
> 
>>
>> fio \
>> --filename=/btrfs/largefile \
>> --directory=/btrfs \
>> --filesize=50G \
>> --size=50G \
>> --bs=64k \
>> --ioengine=libaio \
>> --rw=read \
>> --direct=1 \
>> --numjobs=1 \
>> --group_reporting \
>> --thread \
>> --name iops-test-job
>>
>>   It is system specific?
> 
> With this command, dstat output looks good:
> 
> https://paste.opensuse.org/view/simple/93159623
> 
> So I think it might be specific to whether we test direct of buffered
> I/O. Or to the number of jobs (single vs multiple jobs). Since the most
> of I/O on production environments is usually buffered, I think we should
> test with direct=0 too.
> 
  It explains the stall for now, so it might be reading from the cache
  so there was actually no IO to the device.

  Agreed it must be tested with the most common type of IO. But as the
  cache comes into play I was a bit reluctant.

Thanks, Anand


> Cheers,
> Michal
>