testing thin provisioned block device

testing thin provisioned block device

[Tim Walker]

Hello-

I need to do random writes/reads to some thin-provisioned block
devices.  Writes are no problem, but I have to make sure the reads
come from blocks that have already been written (or else the device
synthesizes fill data).  Pre-filling the device is not desirable since
they are 8-12 TiB, plus that isn't the way they are really used.

IO at the block level is our customer's requirement, but I'd also be
interested in the same concept at the file level, where fio randomly
reads from files that it has randomly written.

I'm sure I'm not the first person who has come up against this, but
I've searched/Googled the best I can and have come up empty. Can
somebody point me to the correct switches to force reads to be
randomly selected from the blocks that have been written by that test
sequence?

Thanks in advance.

Best regards,
-Tim

-- 
Tim Walker
Product Design Systems Engineering, Seagate Technology

Re: testing thin provisioned block device

[Alireza Haghdoost]

On Mon, May 23, 2016 at 9:24 AM, Tim Walker <tim.t.walker@seagate.com> wrote:
>
> Hello-
>
> I need to do random writes/reads to some thin-provisioned block
> devices.  Writes are no problem, but I have to make sure the reads
> come from blocks that have already been written (or else the device
> synthesizes fill data).  Pre-filling the device is not desirable since
> they are 8-12 TiB, plus that isn't the way they are really used.
>
> IO at the block level is our customer's requirement, but I'd also be
> interested in the same concept at the file level, where fio randomly
> reads from files that it has randomly written.
>
> I'm sure I'm not the first person who has come up against this, but
> I've searched/Googled the best I can and have come up empty. Can
> somebody point me to the correct switches to force reads to be
> randomly selected from the blocks that have been written by that test
> sequence?
>
> Thanks in advance.
>
> Best regards,
> -Tim

Tim,

If you don't want to mix read and write workload, could you just write
randomly at your block device with one job first, then use the same
random seed in the next job to read those random blocks ?

--Alireza

Re: testing thin provisioned block device

[Alireza Haghdoost]

On Mon, May 23, 2016 at 9:40 AM, Alireza Haghdoost <alireza@cs.umn.edu> wrote:
> On Mon, May 23, 2016 at 9:24 AM, Tim Walker <tim.t.walker@seagate.com> wrote:
>>
>> Hello-
>>
>> I need to do random writes/reads to some thin-provisioned block
>> devices.  Writes are no problem, but I have to make sure the reads
>> come from blocks that have already been written (or else the device
>> synthesizes fill data).  Pre-filling the device is not desirable since
>> they are 8-12 TiB, plus that isn't the way they are really used.
>>
>> IO at the block level is our customer's requirement, but I'd also be
>> interested in the same concept at the file level, where fio randomly
>> reads from files that it has randomly written.
>>
>> I'm sure I'm not the first person who has come up against this, but
>> I've searched/Googled the best I can and have come up empty. Can
>> somebody point me to the correct switches to force reads to be
>> randomly selected from the blocks that have been written by that test
>> sequence?
>>
>> Thanks in advance.
>>
>> Best regards,
>> -Tim
>
> Tim,
>
> If you don't want to mix read and write workload, could you just write
> randomly at your block device with one job first, then use the same
> random seed in the next job to read those random blocks ?
>
> --Alireza

Or you could fire two Jobs simultaneously with the same random seed.
First job only writes with large queue depth and the second job only
reads with small queue depth. That would result in a mixed workload
where the writer writes on random blocks with a faster speed than the
reader job.

Re: testing thin provisioned block device

[Andrey Kuzmin]

On Mon, May 23, 2016 at 5:24 PM, Tim Walker <tim.t.walker@seagate.com> wrote:
> Hello-
>
> I need to do random writes/reads to some thin-provisioned block
> devices.  Writes are no problem, but I have to make sure the reads
> come from blocks that have already been written (or else the device
> synthesizes fill data).  Pre-filling the device is not desirable since
> they are 8-12 TiB, plus that isn't the way they are really used.
>
> IO at the block level is our customer's requirement, but I'd also be
> interested in the same concept at the file level, where fio randomly
> reads from files that it has randomly written.
>
> I'm sure I'm not the first person who has come up against this, but
> I've searched/Googled the best I can and have come up empty. Can
> somebody point me to the correct switches to force reads to be
> randomly selected from the blocks that have been written by that test
> sequence?

To my knowledge, there's no prebuilt configuration to read only those blocks
that have been written before.

At the same time, when randommap is on, FIO keeps track of the blocks written,
so it may be possible to limit the read generation to those by adding
a bit of code.

Regards,
Andrey

>
> Thanks in advance.
>
> Best regards,
> -Tim
>
> --
> Tim Walker
> Product Design Systems Engineering, Seagate Technology
> --
> To unsubscribe from this list: send the line "unsubscribe fio" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: testing thin provisioned block device

[Sitsofe Wheeler]

On 23 May 2016 at 16:02, Andrey Kuzmin <andrey.v.kuzmin@gmail.com> wrote:
>
> To my knowledge, there's no prebuilt configuration to read only those blocks
> that have been written before.
>
> At the same time, when randommap is on, FIO keeps track of the blocks written,
> so it may be possible to limit the read generation to those by adding
> a bit of code.

I thought this type of facility had to exist so that verification
triggers (https://github.com/axboe/fio/blob/fio-2.10/HOWTO#L2214 ) via
verify_state_load could work?

-- 
Sitsofe | http://sucs.org/~sits/

Re: testing thin provisioned block device

[Tim Walker]

This looks interesting.  Thanks!
I will report back...
-Tim

On Mon, May 23, 2016 at 11:41 AM, Sitsofe Wheeler <sitsofe@gmail.com> wrote:
> On 23 May 2016 at 16:02, Andrey Kuzmin <andrey.v.kuzmin@gmail.com> wrote:
>>
>> To my knowledge, there's no prebuilt configuration to read only those blocks
>> that have been written before.
>>
>> At the same time, when randommap is on, FIO keeps track of the blocks written,
>> so it may be possible to limit the read generation to those by adding
>> a bit of code.
>
> I thought this type of facility had to exist so that verification
> triggers (https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_axboe_fio_blob_fio-2D2.10_HOWTO-23L2214&d=CwIBaQ&c=IGDlg0lD0b-nebmJJ0Kp8A&r=NW1X0yRHNNEluZ8sOGXBxCbQJZPWcIkPT0Uy3ynVsFU&m=kMV_vJkDG44PlTPqwWzkRqZGd9WboeQTwskJijXbUB4&s=hocA_d46stzA2Fk0_dkI2uDHlaEs81U8qVRHKQhSfo0&e=  ) via
> verify_state_load could work?
>
> --
> Sitsofe | https://urldefense.proofpoint.com/v2/url?u=http-3A__sucs.org_-7Esits_&d=CwIBaQ&c=IGDlg0lD0b-nebmJJ0Kp8A&r=NW1X0yRHNNEluZ8sOGXBxCbQJZPWcIkPT0Uy3ynVsFU&m=kMV_vJkDG44PlTPqwWzkRqZGd9WboeQTwskJijXbUB4&s=j0EyneqqMvyjeXgg8I4VEb8qly5XZH14rVVpAtmzXK4&e=



-- 
Tim Walker
Product Design Systems Engineering, Seagate Technology
(303) 775-3770

Re: testing thin provisioned block device

[Sitsofe Wheeler]

On 24 May 2016 at 04:52, Tim Walker <tim.t.walker@seagate.com> wrote:
> This looks interesting.  Thanks!
> I will report back...

Other possibilities would be to use verify_backlog
(https://github.com/axboe/fio/blob/fio-2.10/HOWTO#L1467 ) so you're
re-reading periodically as you're going along or to use
io_limit/io_size
(https://github.com/axboe/fio/blob/fio-2.10/HOWTO#L495 ) with two
stonewalled jobs using the same seeds but where the first job was a
write and the second a read.

-- 
Sitsofe | http://sucs.org/~sits/

Re: testing thin provisioned block device

[Jens Axboe]

On 05/23/2016 11:41 AM, Sitsofe Wheeler wrote:
> On 23 May 2016 at 16:02, Andrey Kuzmin <andrey.v.kuzmin@gmail.com> wrote:
>>
>> To my knowledge, there's no prebuilt configuration to read only those blocks
>> that have been written before.
>>
>> At the same time, when randommap is on, FIO keeps track of the blocks written,
>> so it may be possible to limit the read generation to those by adding
>> a bit of code.
>
> I thought this type of facility had to exist so that verification
> triggers (https://github.com/axboe/fio/blob/fio-2.10/HOWTO#L2214 ) via
> verify_state_load could work?

Definitely. Similarly to what Alireza suggested, using the same random 
seed for a reader would read the exact same blocks that fio previously 
wrote. You don't strictly need the random map for that. If you seed 
things the same way, then the offsets/sizes/etc will be generated 
identically. The random map can help with verifies if you overwrite 
blocks or need to trim, but for a pure write-then-read setup, it's not 
needed.

For instance, if you do verifies, you can have a write job file that you 
run to completion. Then you take that exact same job file, and change 
the write (random or sequential) to the equivalent read, and run fio 
again. The same thing is of course possible without doing the 
verification, but just using that same mechanism to read only the parts 
that were written.

-- 
Jens Axboe