Some questions about FIO

Some questions about FIO

[neto from Brazil]

Hi All,

This is neto from Brazil

How are you?

I would like to check if we have plans to have the following with FIO:

1) When FIO runs, it shows IOPS + MB/s. Do we have any plans to show also average latency? It would be really nice to have the latency (like Vdbench has) - LIVE
2) It is possible to have an option to show the queue depth used on the servers by FIO? It could help to identify potential bottlenecks (e.g. too many threads, etc) on the server side.
3) We don't have a curve option like vdbench. Do we have plans to have it on FIO? It would be really nice to have a parameter like curve (10-120) which means: try the MAX, after to get the average, does 10%, does 20%. It would help to identify the "knee" on the curve.

Thanks,

neto

Re: Some questions about FIO

[Sitsofe Wheeler]

Hey Neto!

On Thu, 9 Aug 2018 at 16:58, neto from Brazil
<Antonio.Jose.Rodrigues.Neto@netapp.com> wrote:
>
> Hi All,
>
> This is neto from Brazil
>
> How are you?
>
> I would like to check if we have plans to have the following with FIO:
>
> 1) When FIO runs, it shows IOPS + MB/s. Do we have any plans to show also average latency? It would be really nice to have the latency (like Vdbench has) - LIVE

This sounds similar to https://github.com/axboe/fio/issues/584 . I
don't see why it couldn't be done but someone is needed to code it
up...

> 2) It is possible to have an option to show the queue depth used on the servers by FIO? It could help to identify potential bottlenecks (e.g. too many threads, etc) on the server side.

Technically you get this in the IO depths summary data that fio
reports back (see
https://fio.readthedocs.io/en/latest/fio_doc.html#interpreting-the-output
) albeit from fio's perspective.

> 3) We don't have a curve option like vdbench. Do we have plans to have it on FIO? It would be really nice to have a parameter like curve (10-120) which means: try the MAX, after to get the average, does 10%, does 20%. It would help to identify the "knee" on the curve.

From the vdbench manual:

"When no skew is requested for any workload involved, determine the
maximum possible I/O rate by running 'iorate=max'. After this 'max'
run, the target skew for the requested workloads will be set to the
skew that was observed. Workloads of 10%, 50%, 70%, 80%, 90% and 100%
of the observed maximum I/O rate and skew will be run."

Sounds like a mix of Steady state
(https://fio.readthedocs.io/en/latest/fio_doc.html#steady-state ) and
diskplorer https://github.com/avikivity/diskplorer with more
flexibility. Like 1) if someone were to code it up I'm sure it would
be looked at...

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

Re: Some questions about FIO

[neto from Brazil]

Thanks my friend

> On Aug 31, 2018, at 2:36 AM, Sitsofe Wheeler <sitsofe@gmail.com> wrote:
> 
> Hey Neto!
> 
> On Thu, 9 Aug 2018 at 16:58, neto from Brazil
> <Antonio.Jose.Rodrigues.Neto@netapp.com> wrote:
>> 
>> Hi All,
>> 
>> This is neto from Brazil
>> 
>> How are you?
>> 
>> I would like to check if we have plans to have the following with FIO:
>> 
>> 1) When FIO runs, it shows IOPS + MB/s. Do we have any plans to show also average latency? It would be really nice to have the latency (like Vdbench has) - LIVE
> 
> This sounds similar to https://github.com/axboe/fio/issues/584 . I
> don't see why it couldn't be done but someone is needed to code it
> up...
> 
>> 2) It is possible to have an option to show the queue depth used on the servers by FIO? It could help to identify potential bottlenecks (e.g. too many threads, etc) on the server side.
> 
> Technically you get this in the IO depths summary data that fio
> reports back (see
> https://fio.readthedocs.io/en/latest/fio_doc.html#interpreting-the-output
> ) albeit from fio's perspective.
> 
>> 3) We don't have a curve option like vdbench. Do we have plans to have it on FIO? It would be really nice to have a parameter like curve (10-120) which means: try the MAX, after to get the average, does 10%, does 20%. It would help to identify the "knee" on the curve.
> 
> From the vdbench manual:
> 
> "When no skew is requested for any workload involved, determine the
> maximum possible I/O rate by running 'iorate=max'. After this 'max'
> run, the target skew for the requested workloads will be set to the
> skew that was observed. Workloads of 10%, 50%, 70%, 80%, 90% and 100%
> of the observed maximum I/O rate and skew will be run."
> 
> Sounds like a mix of Steady state
> (https://fio.readthedocs.io/en/latest/fio_doc.html#steady-state ) and
> diskplorer https://github.com/avikivity/diskplorer with more
> flexibility. Like 1) if someone were to code it up I'm sure it would
> be looked at...
> 
> -- 
> Sitsofe | http://sucs.org/~sits/