2.1.2 released

2.1.2 released

[Jens Axboe]

Hi Folks,

I tagged the release of 2.1.2 this morning. The basic changes since
2.1.1 are:

- The big graphing rework from Erwan.
- A fix for a potential verification failure, courtesy of Josef Bacik.
- Support for randomness of reads and writes independently.
- Support for setting block sizes (or ranges) for random of sequential
  IO.
- Support for multicast network IO in the net ioengine, from Shawn
  Bohrer.
- RDMA io engine should now compile and work again, was broken with the
  recent rewrites. From Yufei Ren.

65 commits in total, of which I only did 8 of them. I believe that is a
new record! Thanks a lot for your contributions, keep them coming.

-- 
Jens Axboe

Re: 2.1.2 released

[Erwan Velu]

On 06/08/2013 17:58, Jens Axboe wrote:
> Hi Folks,
>
> I tagged the release of 2.1.2 this morning. The basic changes since
> 2.1.1 are:
>
> - The big graphing rework from Erwan.
> - A fix for a potential verification failure, courtesy of Josef Bacik.
> - Support for randomness of reads and writes independently.
> - Support for setting block sizes (or ranges) for random of sequential
>    IO.
> - Support for multicast network IO in the net ioengine, from Shawn
>    Bohrer.
> - RDMA io engine should now compile and work again, was broken with the
>    recent rewrites. From Yufei Ren.
>
> 65 commits in total, of which I only did 8 of them. I believe that is a
> new record! Thanks a lot for your contributions, keep them coming.
>
I'd like to add also the genfio tool which is very useful to describe 
complex and repetitive fio's jobs.

I'm also adding an url where people can see the kind of graphing 
fio2gnuplot does : 
http://www.flickr.com/photos/ennael/sets/72157634249027122/

Cheers,
Erwan

Re: 2.1.2 released

[Jens Axboe]

On 08/06/2013 10:24 AM, Erwan Velu wrote:
> On 06/08/2013 17:58, Jens Axboe wrote:
>> Hi Folks,
>>
>> I tagged the release of 2.1.2 this morning. The basic changes since
>> 2.1.1 are:
>>
>> - The big graphing rework from Erwan.
>> - A fix for a potential verification failure, courtesy of Josef Bacik.
>> - Support for randomness of reads and writes independently.
>> - Support for setting block sizes (or ranges) for random of sequential
>>    IO.
>> - Support for multicast network IO in the net ioengine, from Shawn
>>    Bohrer.
>> - RDMA io engine should now compile and work again, was broken with the
>>    recent rewrites. From Yufei Ren.
>>
>> 65 commits in total, of which I only did 8 of them. I believe that is a
>> new record! Thanks a lot for your contributions, keep them coming.
>>
> I'd like to add also the genfio tool which is very useful to describe
> complex and repetitive fio's jobs.

Indeed, I completely forgot about that one. Sorry about that.

> I'm also adding an url where people can see the kind of graphing
> fio2gnuplot does :
> http://www.flickr.com/photos/ennael/sets/72157634249027122/

Great, always a good thing to have samples to point people at.

-- 
Jens Axboe

Re: 2.1.2 released

[Dennis Jacobfeuerborn]

On 06.08.2013 18:24, Erwan Velu wrote:
> On 06/08/2013 17:58, Jens Axboe wrote:
>> Hi Folks,
>>
>> I tagged the release of 2.1.2 this morning. The basic changes since
>> 2.1.1 are:
>>
>> - The big graphing rework from Erwan.
>> - A fix for a potential verification failure, courtesy of Josef Bacik.
>> - Support for randomness of reads and writes independently.
>> - Support for setting block sizes (or ranges) for random of sequential
>>    IO.
>> - Support for multicast network IO in the net ioengine, from Shawn
>>    Bohrer.
>> - RDMA io engine should now compile and work again, was broken with the
>>    recent rewrites. From Yufei Ren.
>>
>> 65 commits in total, of which I only did 8 of them. I believe that is a
>> new record! Thanks a lot for your contributions, keep them coming.
>>
> I'd like to add also the genfio tool which is very useful to describe
> complex and repetitive fio's jobs.
>
> I'm also adding an url where people can see the kind of graphing
> fio2gnuplot does :
> http://www.flickr.com/photos/ennael/sets/72157634249027122/

How does the fio2gnuplot tool work? There is no man page or example 
invocation that explains how to generate the plots. I tried "fio2gnuplot 
--iops" but that only dumps the usage options without any hint of what 
is wrong with that call.

Regards,
   Dennis

Re: 2.1.2 released

[Jens Axboe]

On Tue, Aug 06 2013, Dennis Jacobfeuerborn wrote:
> On 06.08.2013 18:24, Erwan Velu wrote:
> >On 06/08/2013 17:58, Jens Axboe wrote:
> >>Hi Folks,
> >>
> >>I tagged the release of 2.1.2 this morning. The basic changes since
> >>2.1.1 are:
> >>
> >>- The big graphing rework from Erwan.
> >>- A fix for a potential verification failure, courtesy of Josef Bacik.
> >>- Support for randomness of reads and writes independently.
> >>- Support for setting block sizes (or ranges) for random of sequential
> >>   IO.
> >>- Support for multicast network IO in the net ioengine, from Shawn
> >>   Bohrer.
> >>- RDMA io engine should now compile and work again, was broken with the
> >>   recent rewrites. From Yufei Ren.
> >>
> >>65 commits in total, of which I only did 8 of them. I believe that is a
> >>new record! Thanks a lot for your contributions, keep them coming.
> >>
> >I'd like to add also the genfio tool which is very useful to describe
> >complex and repetitive fio's jobs.
> >
> >I'm also adding an url where people can see the kind of graphing
> >fio2gnuplot does :
> >http://www.flickr.com/photos/ennael/sets/72157634249027122/
> 
> How does the fio2gnuplot tool work? There is no man page or example
> invocation that explains how to generate the plots. I tried "fio2gnuplot
> --iops" but that only dumps the usage options without any hint of what is
> wrong with that call.

Erwan, care to provide some documentation with the tool? That's the
first place I would look, too.

-- 
Jens Axboe

Re: 2.1.2 released

[Erwan Velu]

On 07/08/2013 03:06, Jens Axboe wrote:
> [..]
> Erwan, care to provide some documentation with the tool? That's the
> first place I would look, too.
I will.

Erwan,

Re: 2.1.2 released

[Martin Steigerwald]

Hi Erwan,

Am Mittwoch, 7. August 2013, 10:57:53 schrieb Erwan Velu:
> On 07/08/2013 03:06, Jens Axboe wrote:
> > [..]
> > Erwan, care to provide some documentation with the tool? That's the
> > first place I would look, too.
> 
> I will.

Great would be a manpage. As its required for each binary as per Debian 
packaging policy. If its printing a --help usage in some standard format, 
help2man could help with that.

I can help as well.

Thanks,
-- 
Martin 'Helios' Steigerwald - http://www.Lichtvoll.de
GPG: 03B0 0D6C 0040 0710 4AFA  B82F 991B EAAC A599 84C7

Re: 2.1.2 released

[Erwan Velu]

On 07/08/2013 11:46, Martin Steigerwald wrote:
> Hi Erwan,
>
> [...]
> Great would be a manpage. As its required for each binary as per Debian
> packaging policy. If its printing a --help usage in some standard format,
> help2man could help with that.
>
> I can help as well.
>
> Thanks,
I'm writting the man page. You are pefectly right it's important and I 
shouldn't have postponed that ...
That shall be ready by the day.

Cheers,

Re: 2.1.2 released

[Erwan Velu]

On 07/08/2013 03:06, Jens Axboe wrote:
[...]
> Erwan, care to provide some documentation with the tool? That's the 
> first place I would look, too. 
I've been pushing the first man page and fixes requested by Denis here : 
https://github.com/enovance/fio/commits/erwan/next

English is not my mother tongue, so if some want to review it, please 
consider the fio2gnuplot.manpage which is much more easier to hack.

The makefile does the rest of converting.

Hope this makes things clearer & easier to use.

Cheers,
Erwan

Re: 2.1.2 released

[Erwan Velu]

On 06/08/2013 23:59, Dennis Jacobfeuerborn wrote:
> [..]
> How does the fio2gnuplot tool work? There is no man page or example 
> invocation that explains how to generate the plots. I tried 
> "fio2gnuplot --iops" but that only dumps the usage options without any 
> hint of what is wrong with that call.
Dear Dennis,

I'm sorry the inline help of fio2gnuplot isn't sufficient yet to get 
understood easily, I'll work on fixing this.

Anyway, here come how it works :

While running fio jobs, you are producing a set of log files located in 
a directory. When considering a good amount of logs, logs file are 
usually having different filename to remember what log match a 
particular run. i.e genfio does works this way ;o)

Then, when you want to plot some fio's output, you have to call 
fio2gnuplot with a 'pattern' (-p) option.
The pattern is a kind of regexp that allow you to choose what file you 
want to plot. Mostly, I do use the '*' for this purpose.

Let's say you have four files named like :
host1-read-4k_iops.log
host2-read-4k_iops.log
host3-read-4k_iops.log
host1-write-4k_iops.log

If you want to plot all the traces related to the reading you'll do :
fio2gnuplot -p 'host*-read-4k_iops.log'

Note the single quote to avoid the shell interpreting the star (*), it's 
up to fio2gnuplot to handle this.

Then, if you want to generate the rendering, you can use the '-g' flag 
to request the plotting (maybe this could be a default option).

Then, you can choose in which dir you want plot all this stuff, let's 
use the '-d' option. fio2gnuplot generates a _lot_ of traces, so that's 
a recommanded option.

So a typical cmdline would be :

fio2gnuplot -p 'host*-read-4k_iops.log' -g -d read-4k

About the --iops output you choose, it's just a predefined pattern that 
match all the "iops" files from the current directory. In fact, I didn't 
used them a lot .... a you spotted two nice bugs I shoudn't have commit.

First, it does print the help while the cmdline wasn't correct. It 
should have told you something was wrong on the cmdline.

Then, the --iops wasn't properly managed and as a result didn't got 
supported... .What a shame... Please make a try with '-i' on this 
version of the tool.

To fix this two bugs, I've been pushing two fixes in my dev tree that 
I'll ask to merge a little bit later.
If you want to test them, that's available here : 
https://github.com/enovance/fio/commits/erwan/next

I do agree that I should had a man page and much more example.
Thanks for testing my work & reporting.

Cheers,
Erwan,

Re: 2.1.2 released

[Dennis Jacobfeuerborn]

[-- Attachment #1: Type: text/plain, Size: 4243 bytes --]

On 07.08.2013 10:57, Erwan Velu wrote:
> On 06/08/2013 23:59, Dennis Jacobfeuerborn wrote:
>> [..]
>> How does the fio2gnuplot tool work? There is no man page or example
>> invocation that explains how to generate the plots. I tried
>> "fio2gnuplot --iops" but that only dumps the usage options without any
>> hint of what is wrong with that call.
> Dear Dennis,
>
> I'm sorry the inline help of fio2gnuplot isn't sufficient yet to get
> understood easily, I'll work on fixing this.
>
> Anyway, here come how it works :
>
> While running fio jobs, you are producing a set of log files located in
> a directory. When considering a good amount of logs, logs file are
> usually having different filename to remember what log match a
> particular run. i.e genfio does works this way ;o)
>
> Then, when you want to plot some fio's output, you have to call
> fio2gnuplot with a 'pattern' (-p) option.
> The pattern is a kind of regexp that allow you to choose what file you
> want to plot. Mostly, I do use the '*' for this purpose.
>
> Let's say you have four files named like :
> host1-read-4k_iops.log
> host2-read-4k_iops.log
> host3-read-4k_iops.log
> host1-write-4k_iops.log
>
> If you want to plot all the traces related to the reading you'll do :
> fio2gnuplot -p 'host*-read-4k_iops.log'
>
> Note the single quote to avoid the shell interpreting the star (*), it's
> up to fio2gnuplot to handle this.
>
> Then, if you want to generate the rendering, you can use the '-g' flag
> to request the plotting (maybe this could be a default option).
>
> Then, you can choose in which dir you want plot all this stuff, let's
> use the '-d' option. fio2gnuplot generates a _lot_ of traces, so that's
> a recommanded option.
>
> So a typical cmdline would be :
>
> fio2gnuplot -p 'host*-read-4k_iops.log' -g -d read-4k
>
> About the --iops output you choose, it's just a predefined pattern that
> match all the "iops" files from the current directory. In fact, I didn't
> used them a lot .... a you spotted two nice bugs I shoudn't have commit.
>
> First, it does print the help while the cmdline wasn't correct. It
> should have told you something was wrong on the cmdline.
>
> Then, the --iops wasn't properly managed and as a result didn't got
> supported... .What a shame... Please make a try with '-i' on this
> version of the tool.
>
> To fix this two bugs, I've been pushing two fixes in my dev tree that
> I'll ask to merge a little bit later.
> If you want to test them, that's available here :
> https://github.com/enovance/fio/commits/erwan/next
>
> I do agree that I should had a man page and much more example.
> Thanks for testing my work & reporting.

Thanks for providing the script. I'm currently looking at using fio to 
create a few baseline test to get a quick glimpse of how fast a given 
system can do certain I/O patterns and being able to graph this would 
make it easier to compare the results.

I tested the version from the erwan/next branch and the --iops option 
works now and if I don't provide it I get a proper error. Although when 
I provide a nonsense option like "--fnsofno" I get the error "Error: One 
of the option passed to the cmdline was supported". That is probably 
supposed to say "was not supported".

The graphing seems to have issues too. I attached a job file and the 
graphs generated using "./fio2gnuplot.py -d t -g -i". As you can see in 
the compare-2Draw graph I get multiple lines for the test4k and test8k 
jobs but not for the test16k job and in the compare-2Dtrend graph I get 
strange loops for the same two out of three jobs. The 3D graph looks 
strange too and shows "disks" which doesn't really apply in this 
particular case.

Lastly the tool does output a lot of things by default which are 
probably not useful to 99.9% of the people out there. I think the 
default should be to only output the 2Draw and the 3D graph and delete 
the data files (mygraph, mymath, etc.) and allow the user to keep these 
files and create the additional graphs using command line options.
In fact in the case of multiple input files it would probably be best to 
only end up with the compate-2Draw graph as the graphs for the 
individual jobs are pretty much redundant in that case.

Regards,
   Dennis




[-- Attachment #2: compare-result-2Draw.png --]
[-- Type: image/png, Size: 13898 bytes --]

[-- Attachment #3: compare-result-2Dtrend.png --]
[-- Type: image/png, Size: 11147 bytes --]

[-- Attachment #4: result-3D.png --]
[-- Type: image/png, Size: 14446 bytes --]

[-- Attachment #5: randread.fio --]
[-- Type: text/plain, Size: 291 bytes --]

[global]
rw=randread
direct=1
filename=/dev/sda
numjobs=1
group_reporting=1
runtime=5

[randread]
bs=4k
write_iops_log=test4k
log_avg_msec=200
stonewall

[randread2]
bs=8k
write_iops_log=test8k
log_avg_msec=200
stonewall

[randread3]
bs=16k
write_iops_log=test16k
log_avg_msec=200
stonewall

Re: 2.1.2 released

[Erwan Velu]

On 07/08/2013 13:14, Dennis Jacobfeuerborn wrote:

[...]
> I tested the version from the erwan/next branch and the --iops option 
> works now and if I don't provide it I get a proper error. Although 
> when I provide a nonsense option like "--fnsofno" I get the error 
> "Error: One of the option passed to the cmdline was supported". That 
> is probably supposed to say "was not supported".
This is now fixed in my branch.... silly typo.
>
> The graphing seems to have issues too. I attached a job file and the 
> graphs generated using "./fio2gnuplot.py -d t -g -i". As you can see 
> in the compare-2Draw graph I get multiple lines for the test4k and 
> test8k jobs but not for the test16k job and in the compare-2Dtrend 
> graph I get strange loops for the same two out of three jobs. The 3D 
> graph looks strange too and shows "disks" which doesn't really apply 
> in this particular case.
Waow... I ran this test on my local setup with fio 2.1.2 and this 
doesn't occurs.... I do have the 3 lines perfectly printed.
That's maybe a gnuplot effect.
Can you provide the version of gnuplot you use and the log files ?
I'm running 4.6 patchlevel 1

> Lastly the tool does output a lot of things by default which are 
> probably not useful to 99.9% of the people out there. 
This is now fixed in my branch.

> I think the default should be to only output the 2Draw and the 3D 
> graph and delete the data files (mygraph, mymath, etc.) and allow the 
> user to keep these files and create the additional graphs using 
> command line options.
Fixed in my branch. Default is to kill temp files if gnuplot success 
unless keep them. If user ask to keep them using -k option, let's always 
keep them.
That shall solve your issue.

> In fact in the case of multiple input files it would probably be best 
> to only end up with the compate-2Draw graph as the graphs for the 
> individual jobs are pretty much redundant in that case.
That really depends on people's usage. Some could be interested at 
publishing the compare-* graphs but need to understand some traces if 
weird effects occurs. I did develop all thoses graphs with a "pick the 
one you need" approach.

Yes I do plot too much but this avoids having too much options to pass 
to the tool. People have just to pick the one they need.

Re: 2.1.2 released

[Dennis Jacobfeuerborn]

[-- Attachment #1: Type: text/plain, Size: 3219 bytes --]

On 07.08.2013 14:26, Erwan Velu wrote:
> On 07/08/2013 13:14, Dennis Jacobfeuerborn wrote:
>
> [...]
>> I tested the version from the erwan/next branch and the --iops option
>> works now and if I don't provide it I get a proper error. Although
>> when I provide a nonsense option like "--fnsofno" I get the error
>> "Error: One of the option passed to the cmdline was supported". That
>> is probably supposed to say "was not supported".
> This is now fixed in my branch.... silly typo.
>>
>> The graphing seems to have issues too. I attached a job file and the
>> graphs generated using "./fio2gnuplot.py -d t -g -i". As you can see
>> in the compare-2Draw graph I get multiple lines for the test4k and
>> test8k jobs but not for the test16k job and in the compare-2Dtrend
>> graph I get strange loops for the same two out of three jobs. The 3D
>> graph looks strange too and shows "disks" which doesn't really apply
>> in this particular case.
> Waow... I ran this test on my local setup with fio 2.1.2 and this
> doesn't occurs.... I do have the 3 lines perfectly printed.
> That's maybe a gnuplot effect.
> Can you provide the version of gnuplot you use and the log files ?
> I'm running 4.6 patchlevel 1

I attached the iops logs used and the Gnuplot is version 4.6.1 on Fedora 
18 (Package name "gnuplot-4.6.1-4.fc18.x86_64").

>> Lastly the tool does output a lot of things by default which are
>> probably not useful to 99.9% of the people out there.
> This is now fixed in my branch.
>
>> I think the default should be to only output the 2Draw and the 3D
>> graph and delete the data files (mygraph, mymath, etc.) and allow the
>> user to keep these files and create the additional graphs using
>> command line options.
> Fixed in my branch. Default is to kill temp files if gnuplot success
> unless keep them. If user ask to keep them using -k option, let's always
> keep them.
> That shall solve your issue.

I just tested this and for me it works as advertized. Thanks!

>> In fact in the case of multiple input files it would probably be best
>> to only end up with the compate-2Draw graph as the graphs for the
>> individual jobs are pretty much redundant in that case.
> That really depends on people's usage. Some could be interested at
> publishing the compare-* graphs but need to understand some traces if
> weird effects occurs. I did develop all thoses graphs with a "pick the
> one you need" approach.
>
> Yes I do plot too much but this avoids having too much options to pass
> to the tool. People have just to pick the one they need.

I have thought about this a bit and wonder if it would be useful to to 
use a job file similar to the fio job files to describe which graphs to 
render and which parameters to use. Example:

[global]
output_dir=test
min_time=0
max_time=-1

[graph1]
type=individual
interpolation=raw
title="raw iops"
pattern="*_iops.log"

[graph2]
type=grouped2d
interpolation=smooth
title="smooth iops summary"
pattern="*_iops.log"

Doing it this way would allow fio2gnuplot to be invoked like 
"fio2gnuplot -c/--config file.fiograph" so no complex command line 
arguments would be necessary and the user could easily customize which 
graphs are generated.

Regards,
   Dennis

[-- Attachment #2: test4k_iops.log --]
[-- Type: text/x-log, Size: 456 bytes --]

500, 8716, 0, 0
1000, 8760, 0, 0
1500, 8730, 0, 0
2000, 8692, 0, 0
2500, 8868, 0, 0
3000, 8686, 0, 0
3500, 9002, 0, 0
4000, 8838, 0, 0
4500, 8966, 0, 0
500, 8998, 0, 0
1000, 9002, 0, 0
1500, 9076, 0, 0
2000, 9050, 0, 0
2500, 8966, 0, 0
3000, 9008, 0, 0
3500, 9078, 0, 0
4000, 9018, 0, 0
4500, 9022, 0, 0
500, 8628, 0, 0
1000, 8842, 0, 0
1500, 8880, 0, 0
2000, 8828, 0, 0
2500, 8748, 0, 0
3000, 8780, 0, 0
3500, 8948, 0, 0
4000, 8764, 0, 0
4500, 8856, 0, 0

[-- Attachment #3: test8k_iops.log --]
[-- Type: text/x-log, Size: 456 bytes --]

500, 7198, 0, 0
1000, 7222, 0, 0
1500, 7300, 0, 0
2000, 7154, 0, 0
2500, 7096, 0, 0
3000, 7268, 0, 0
3500, 7372, 0, 0
4000, 7228, 0, 0
4500, 7268, 0, 0
500, 7678, 0, 0
1000, 7812, 0, 0
1500, 7774, 0, 0
2000, 7728, 0, 0
2500, 7572, 0, 0
3000, 7598, 0, 0
3500, 7254, 0, 0
4000, 7144, 0, 0
4500, 7298, 0, 0
500, 7142, 0, 0
1000, 6960, 0, 0
1500, 7430, 0, 0
2000, 7258, 0, 0
2500, 7376, 0, 0
3000, 7468, 0, 0
3500, 7362, 0, 0
4000, 7100, 0, 0
4500, 7192, 0, 0

[-- Attachment #4: test16k_iops.log --]
[-- Type: text/x-log, Size: 152 bytes --]

500, 5536, 0, 0
1000, 5524, 0, 0
1500, 5416, 0, 0
2000, 5212, 0, 0
2500, 5384, 0, 0
3000, 5288, 0, 0
3500, 5116, 0, 0
4001, 5418, 0, 0
4501, 5380, 0, 0

Re: 2.1.2 released

[Erwan Velu]

On 07/08/2013 18:44, Dennis Jacobfeuerborn wrote:
>    I attached the iops logs used and the Gnuplot is version 4.6.1 on 
> Fedora 18 (Package name "gnuplot-4.6.1-4.fc18.x86_64").

In fact, your logs are pretty surprising for me.
Can anyone, maybe jens, could explains why the time is getting backward 
in test4k_iops.log ?
It goes from 500 to 4500 and restart to 500 to 4500 and one more time.
My plotting considers that time is only getting forward never backward.

[...]
> I just tested this and for me it works as advertized. Thanks!

Glad it works !

[...]
> Doing it this way would allow fio2gnuplot to be invoked like 
> "fio2gnuplot -c/--config file.fiograph" so no complex command line 
> arguments would be necessary and the user could easily customize which 
> graphs are generated.
Interesting approach, I'll think about it.