stat: Increase depth levels to log latency

[ikegami-t]

Previously latency log depth levels are 10 for nsec and usec and 12 for msec.
But to measure latency variation on a finer granularity required to increase.
So increase the depth leves as 20 for nsec and usec and 22 for msec.

Signed-off-by: Tokunori Ikegami [email protected]

[sitsofe]

I don't know we can accept this as-is - it's bound to break people depending on the previous behaviour who aren't checking the headings... Also wouldn't the documentation have to be updated (

fio/HOWTO

Lines 3884 to 3890 in 8bfe330

	I/O latencies microseconds::
	<=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000
	I/O latencies milliseconds::
	<=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000, 2000, >=2000

)?

[ikegami-t]

Hi,

I don't know we can accept this as-is - it's bound to break people depending on the previous behaviour who aren't checking the headings... Also wouldn't the documentation have to be updated (

fio/HOWTO

Lines 3884 to 3890 in 8bfe330

	I/O latencies microseconds::
	<=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000
	I/O latencies milliseconds::
	<=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000, 2000, >=2000

)?

Thanks for the comment.
Just update the document as mentioned.

Regards,
Ikegami

[ikegami-t]

Just fixed the commit message error as s/leves/levels/.

[sitsofe]

@axboe what are your thoughts on this one?

[sitsofe]

(Re-pinging @axboe for his thoughts on this)

[sitsofe]

I think this is dramatic enough that FIO_SERVER_VER will need ramping (see https://github.com/axboe/fio/blob/master/server.h#L51 ). I'm also unsure if all the tools that scoop up latency ranges will need re-testing:

• https://github.com/axboe/fio/blob/master/tools/hist/fio-histo-log-pctiles.py
• https://github.com/axboe/fio/blob/master/tools/hist/half-bins.py
• https://github.com/axboe/fio/blob/master/t/latency_percentiles.py
• https://github.com/axboe/fio/blob/master/tools/hist/fiologparser_hist.py

Just check if they work initially and report back.

[ikegami-t]

I think this is dramatic enough that FIO_SERVER_VER will need ramping (see https://github.com/axboe/fio/blob/master/server.h#L51 ). I'm also unsure if all the tools that scoop up latency ranges will need re-testing:

Just ramped up the version.

• https://github.com/axboe/fio/blob/master/tools/hist/fio-histo-log-pctiles.py
• https://github.com/axboe/fio/blob/master/tools/hist/half-bins.py
• https://github.com/axboe/fio/blob/master/t/latency_percentiles.py
• https://github.com/axboe/fio/blob/master/tools/hist/fiologparser_hist.py

Just check if they work initially and report back.

Also tested the tools with the changes and seems to work correctly as the same behavior with the original fio version.

[sitsofe]

@ikegami-t Will this impact the terse output https://fio.readthedocs.io/en/latest/fio_man.html#terse-output ?

[ikegami-t]

@ikegami-t Will this impact the terse output https://fio.readthedocs.io/en/latest/fio_man.html#terse-output ?

Yes the terse output will be increased the number of microsecond and millisecond latency outputs since those depth levels increased.

[sitsofe]

Hmm that's going to be a problem because the column headings are no longer going to match and CSV is not self describing...

[ikegami-t]

Hmm that's going to be a problem because the column headings are no longer going to match and CSV is not self describing...

Okay now I am thinking to add an option to change the log latency depth levels as not changed the terse output by default. Thank you.

[ikegami-t]

Hmm that's going to be a problem because the column headings are no longer going to match and CSV is not self describing...

Okay now I am thinking to add an option to change the log latency depth levels as not changed the terse output by default. Thank you.

To fix the issue changed to add an option to change the lat stats depth levels so please review the changes again.

[sitsofe]

@ikegami-t can you squash all these into one commit?

[ikegami-t]

@ikegami-t can you squash all these into one commit?

Yes just squashed the changes as mentioned. Thank you.

[sitsofe]

Where do the 2 - stat_lat_nr come from?

[ikegami-t]

This is for the millisecond latency ranges since it includes ranges "2000" and ">=2000".

[sitsofe]

LGTM @axboe this one looks ready for review/pulling.