Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Vijay Bellur]

On 06/07/2017 02:41 PM, Pranith Kumar Karampuri wrote:





Hi Yuan,



On Wed, Jun 7, 2017 at 1:14 PM, liuy...@cmss.chinamobile.com
 > wrote:

Hi Pranith,

First of all, we owe you guys big thanks for your excellent work.

I'm the lead of storage team in China Mobile, Xiubo is also in my
team :)


Ah! nice :-).



We have deployed more than 100 glusterfs nodes in producion(20 nodes
for the biggest single cluster), so far so good.  Use case ranging
from image/vedio, to small files for mailbox services.
For profile stuff, we'd like to enable it as always to collect perf
stat for monitoring.



Thank you for your feedback and details about deployment. It is always 
fascinating for us to learn about production usage of gluster :-).





Also, I have a question for NDMP support. I bumpped one thread
discussing it but not clear yet.
(http://lists.gluster.org/pipermail/gluster-devel/2016-August/050513.html
)

Do you have any plan to for a glusterfs-ndmp server from scratch? We
are new to the development of glusterfs, I think ndmp support would
be a good starting point for us to join the
development  of the community.  We'd like to see NDMP supported by
glusterfs because some of our customers want it.


I don't know enough about this. Vijay has idea about this, so will
respond later.


We have had some thoughts about implementing a native ndmp server but 
have not had a chance to implement that. If you have an idea or design 
in mind for glusterfs-ndmp server, please share that here. Happy to 
provide any assistance needed for implementing this feature.


Cheers,
Vijay


___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

On Wed, Jun 7, 2017 at 2:28 PM, Xie Changlong 
wrote:

> 在 6/7/2017 4:47 PM, Pranith Kumar Karampuri 写道:
>
>> Yes, they do. The patches I gave earlier provide the facility to this in
>> io-stats. Do let us know if you have any doubts.
>>
>
> Sorry for slowness, but i think i catch up with U now.
>

Don't worry. Do let us know if you have any doubts, we will be happy to
help.


>
>
>>
>>
> --
> Thanks
> -Xie
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

On Wed, Jun 7, 2017 at 1:14 PM, liuy...@cmss.chinamobile.com <
liuy...@cmss.chinamobile.com> wrote:

> Hi Pranith,
>
> First of all, we owe you guys big thanks for your excellent work.
>
> I'm the lead of storage team in China Mobile, Xiubo is also in my team :)
>

Ah! nice :-).


>
> We have deployed more than 100 glusterfs nodes in producion(20 nodes for
> the biggest single cluster), so far so good.  Use case ranging from
> image/vedio, to small files for mailbox services.
> For profile stuff, we'd like to enable it as always to collect perf stat
> for monitoring.
>

Yes, this is something FB guys do already and contributed the patches back
to community. So you should be able to do this too.


>
> We are planing to provide cloud file service by the combination of
> glusterfs and Openstack Manila this year. The enhancement of manila patches
> would be sent to upsteam later.
>

This is good news.


>
> Also, I have a question for NDMP support. I bumpped one thread discussing
> it but not clear yet. (http://lists.gluster.org/
> pipermail/gluster-devel/2016-August/050513.html)
>
> Do you have any plan to for a glusterfs-ndmp server from scratch? We are
> new to the development of glusterfs, I think ndmp support would be a good
> starting point for us to join the
> development  of the community.  We'd like to see NDMP supported by
> glusterfs because some of our customers want it.
>

I don't know enough about this. Vijay has idea about this, so will respond
later.


>
> Yuan
> --
>
>
> *From:* Pranith Kumar Karampuri <pkara...@redhat.com>
> *Date:* 2017-06-07 15:38
> *To:* Xie Changlong <xiechanglon...@gmail.com>
> *CC:* Gluster Devel <gluster-devel@gluster.org>; Hu Jianfei
> <hujian...@cmss.chinamobile.com>; Liu Yuan <liuy...@cmss.chinamobile.com>
> *Subject:* Re: [Gluster-devel] [Qusetions] How profile affect glusterfs
> performance?
>
>
> On Tue, Jun 6, 2017 at 8:14 AM, Xie Changlong <xiechanglon...@gmail.com>
> wrote:
>
>> 在 6/5/2017 6:30 PM, Pranith Kumar Karampuri 写道:
>>
>>> I meant what are you using gluster for? VM workload or image/video file
>>> creation or lots of small files etc etc.
>>>
>>
>> 1) We use glusterfs for general purpose, no limit to image/video file
>> creation or small files etc.
>>
> Okay, this is good. What is the cluster size?
> Is it replica 3 or replica 2 or arbiter or is it EC volume?
>
> Please don't mind me asking so many details. I am delighted to see you
> guys active in the community because I have seen Xiubo Li's work on
> tcmu-runner who is also from chinamobile and his work is pretty good :-).
>
>
>> 2) We just want a low performance affect way to calculate each brick's
>> iops/bandwidth for the upper layer management app. BTW, is there any other
>> way to get iops/bandwith for each brick besides profile?
>>
>
> At the moment we have these stats using profile commands. As per Facebook
> patches they enable json capture of io-stats on their volumes and measure
> these. They have it enabled always. If that is not enough for you, then we
> should probably look at enhancing things. Do send patches if you think
> something would make things better :-).
>
>
>>
>> --
>> Thanks
>> -Xie
>>
>
>
>
> --
> Pranith
>
>


-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/7/2017 4:47 PM, Pranith Kumar Karampuri 写道:

Yes, they do. The patches I gave earlier provide the facility to this in
io-stats. Do let us know if you have any doubts.


Sorry for slowness, but i think i catch up with U now.






--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

On Wed, Jun 7, 2017 at 1:55 PM, Xie Changlong 
wrote:

> 在 6/7/2017 3:08 PM, Pranith Kumar Karampuri 写道:
>
>> On Tue, Jun 6, 2017 at 8:14 AM, Xie Changlong 
>> wrote:
>>
>> 在 6/5/2017 6:30 PM, Pranith Kumar Karampuri 写道:
>>>
>>> I meant what are you using gluster for? VM workload or image/video file
 creation or lots of small files etc etc.


>>> 1) We use glusterfs for general purpose, no limit to image/video file
>>> creation or small files etc.
>>>
>>> Okay, this is good. What is the cluster size?
>> Is it replica 3 or replica 2 or arbiter or is it EC volume?
>>
>
> I use replica 2 on my test.
> But in the real world, "we have deployed more than 100 glusterfs nodes in
> producion(20 nodes for the biggest single cluster)" per Liu Yuan. Replica
> 2/3, or EC are optional.
>
>
>> Please don't mind me asking so many details. I am delighted to see you
>> guys
>>
>
> You are always welcome!
>
> active in the community because I have seen Xiubo Li's work on tcmu-runner
>> who is also from chinamobile and his work is pretty good :-).
>>
>
> Yeah, i would be very pleasant to convey your praise. :)
>
>
>>
>> 2) We just want a low performance affect way to calculate each brick's
>>> iops/bandwidth for the upper layer management app. BTW, is there any
>>> other
>>> way to get iops/bandwith for each brick besides profile?
>>>
>>>
>> At the moment we have these stats using profile commands. As per Facebook
>> patches they enable json capture of io-stats on their volumes and measure
>> these. They have it enabled always. If that is not enough for you, then we
>>
>
> I think you mean FB guys gather data based on io-stats with profile always
> on, am i right?
> BTW, are these patches open to everyone? so i can dig into it.
>

Yes, they do. The patches I gave earlier provide the facility to this in
io-stats. Do let us know if you have any doubts.


>
> should probably look at enhancing things. Do send patches if you think
>> something would make things better :-).
>>
>>
>>
>>> --
>>> Thanks
>>>  -Xie
>>>
>>>
>>
>>
>>
> --
> Thanks
> -Xie
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[liuy...@cmss.chinamobile.com]

Hi Pranith,

First of all, we owe you guys big thanks for your excellent work.

I'm the lead of storage team in China Mobile, Xiubo is also in my team :)

We have deployed more than 100 glusterfs nodes in producion(20 nodes for the 
biggest single cluster), so far so good.  Use case ranging from image/vedio, to 
small files for mailbox services.
For profile stuff, we'd like to enable it as always to collect perf stat for 
monitoring.

We are planing to provide cloud file service by the combination of glusterfs 
and Openstack Manila this year. The enhancement of manila patches would be sent 
to upsteam later.

Also, I have a question for NDMP support. I bumpped one thread discussing it 
but not clear yet. 
(http://lists.gluster.org/pipermail/gluster-devel/2016-August/050513.html)

Do you have any plan to for a glusterfs-ndmp server from scratch? We are new to 
the development of glusterfs, I think ndmp support would be a good starting 
point for us to join the
development  of the community.  We'd like to see NDMP supported by glusterfs 
because some of our customers want it.

Yuan


 
From: Pranith Kumar Karampuri
Date: 2017-06-07 15:38
To: Xie Changlong
CC: Gluster Devel; Hu Jianfei; Liu Yuan
Subject: Re: [Gluster-devel] [Qusetions] How profile affect glusterfs 
performance?


On Tue, Jun 6, 2017 at 8:14 AM, Xie Changlong <xiechanglon...@gmail.com> wrote:
在 6/5/2017 6:30 PM, Pranith Kumar Karampuri 写道:
I meant what are you using gluster for? VM workload or image/video file
creation or lots of small files etc etc.

1) We use glusterfs for general purpose, no limit to image/video file
creation or small files etc.
Okay, this is good. What is the cluster size?
Is it replica 3 or replica 2 or arbiter or is it EC volume?

Please don't mind me asking so many details. I am delighted to see you guys 
active in the community because I have seen Xiubo Li's work on tcmu-runner who 
is also from chinamobile and his work is pretty good :-).


2) We just want a low performance affect way to calculate each brick's 
iops/bandwidth for the upper layer management app. BTW, is there any other way 
to get iops/bandwith for each brick besides profile?

At the moment we have these stats using profile commands. As per Facebook 
patches they enable json capture of io-stats on their volumes and measure 
these. They have it enabled always. If that is not enough for you, then we 
should probably look at enhancing things. Do send patches if you think 
something would make things better :-).
 

-- 
Thanks
-Xie



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/7/2017 3:08 PM, Pranith Kumar Karampuri 写道:

On Tue, Jun 6, 2017 at 8:14 AM, Xie Changlong 
wrote:


在 6/5/2017 6:30 PM, Pranith Kumar Karampuri 写道:


I meant what are you using gluster for? VM workload or image/video file
creation or lots of small files etc etc.



1) We use glusterfs for general purpose, no limit to image/video file
creation or small files etc.


Okay, this is good. What is the cluster size?
Is it replica 3 or replica 2 or arbiter or is it EC volume?


I use replica 2 on my test.
But in the real world, "we have deployed more than 100 glusterfs nodes 
in producion(20 nodes for the biggest single cluster)" per Liu Yuan. 
Replica 2/3, or EC are optional.




Please don't mind me asking so many details. I am delighted to see you guys


You are always welcome!


active in the community because I have seen Xiubo Li's work on tcmu-runner
who is also from chinamobile and his work is pretty good :-).


Yeah, i would be very pleasant to convey your praise. :)





2) We just want a low performance affect way to calculate each brick's
iops/bandwidth for the upper layer management app. BTW, is there any other
way to get iops/bandwith for each brick besides profile?



At the moment we have these stats using profile commands. As per Facebook
patches they enable json capture of io-stats on their volumes and measure
these. They have it enabled always. If that is not enough for you, then we


I think you mean FB guys gather data based on io-stats with profile 
always on, am i right?

BTW, are these patches open to everyone? so i can dig into it.


should probably look at enhancing things. Do send patches if you think
something would make things better :-).




--
Thanks
 -Xie







--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

On Tue, Jun 6, 2017 at 8:14 AM, Xie Changlong 
wrote:

> 在 6/5/2017 6:30 PM, Pranith Kumar Karampuri 写道:
>
>> I meant what are you using gluster for? VM workload or image/video file
>> creation or lots of small files etc etc.
>>
>
> 1) We use glusterfs for general purpose, no limit to image/video file
> creation or small files etc.
>
Okay, this is good. What is the cluster size?
Is it replica 3 or replica 2 or arbiter or is it EC volume?

Please don't mind me asking so many details. I am delighted to see you guys
active in the community because I have seen Xiubo Li's work on tcmu-runner
who is also from chinamobile and his work is pretty good :-).


> 2) We just want a low performance affect way to calculate each brick's
> iops/bandwidth for the upper layer management app. BTW, is there any other
> way to get iops/bandwith for each brick besides profile?
>

At the moment we have these stats using profile commands. As per Facebook
patches they enable json capture of io-stats on their volumes and measure
these. They have it enabled always. If that is not enough for you, then we
should probably look at enhancing things. Do send patches if you think
something would make things better :-).


>
> --
> Thanks
> -Xie
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/5/2017 6:30 PM, Pranith Kumar Karampuri 写道:

I meant what are you using gluster for? VM workload or image/video file
creation or lots of small files etc etc.


1) We use glusterfs for general purpose, no limit to image/video file
creation or small files etc.

2) We just want a low performance affect way to calculate each brick's 
iops/bandwidth for the upper layer management app. BTW, is there any 
other way to get iops/bandwith for each brick besides profile?


--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

On Mon, Jun 5, 2017 at 12:40 PM, Xie Changlong 
wrote:

> 在 6/5/2017 2:34 PM, Pranith Kumar Karampuri 写道:
>
>> What is your usecase if I may ask?
>>
>>
> Only fio/iozone now, maybe you can give me better test sugguestion :)
>

I meant what are you using gluster for? VM workload or image/video file
creation or lots of small files etc etc.


>
> seq read
> # dd if=/dev/zero  of=test bs=4k count=524288
> #fio --filename=test -iodepth=64 -ioengine=libaio --direct=1 --rw=read
> --bs=1m --size=2g --numjobs=4 --runtime=10 --group_reporting
> --name=test-read
> seq write
> # fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=write
> -bs=1m -size=2g -numjobs=4 -runtime=20 -group_reporting -name=test-write
>
> random read
> #fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randread
> -bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting
> -name=test-rand-read
>
> random write
> #fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randwrite
> -bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting
> -name=test-rand-write
>
>
> iozone  1 thread
> # ./iozone -s 10g -r 1m -i 0 -i 1 -f /mnt/test-volume/test -Rb
> /tmp/test_iozone_1.xls
> iozone  2 multi-threads
> # ./iozone -s 10g -r 1m -i 0 -i 1 -f /mnt/test-volume/test -t 8 -Rb
> /tmp/test_iozone_2.xls
>
>
>
> On Mon, Jun 5, 2017 at 11:08 AM, Xie Changlong
>> wrote:
>>
>> 在 6/5/2017 1:22 PM, Pranith Kumar Karampuri 写道:
>>>
>>> On Mon, Jun 5, 2017 at 8:29 AM, Xie Changlong
 wrote:

 在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:

> I think there have been improvements here to use special instructions
> to
>

> --
> Thanks
> -Xie
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/5/2017 3:10 PM, Xie Changlong 写道:

在 6/5/2017 2:34 PM, Pranith Kumar Karampuri 写道:

What is your usecase if I may ask?



Only fio/iozone now, maybe you can give me better test sugguestion :)

seq read
# dd if=/dev/zero  of=test bs=4k count=524288
#fio --filename=test -iodepth=64 -ioengine=libaio --direct=1 --rw=read 
--bs=1m --size=2g --numjobs=4 --runtime=10 --group_reporting 
--name=test-read

seq write
# fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=write 
-bs=1m -size=2g -numjobs=4 -runtime=20 -group_reporting -name=test-write


random read
#fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randread 
-bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting 
-name=test-rand-read


random write
#fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randwrite 
-bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting 
-name=test-rand-write



iozone  1 thread
# ./iozone -s 10g -r 1m -i 0 -i 1 -f /mnt/test-volume/test -Rb 
/tmp/test_iozone_1.xls

iozone  2 multi-threads
# ./iozone -s 10g -r 1m -i 0 -i 1 -f /mnt/test-volume/test -t 8 -Rb 
/tmp/test_iozone_2.xls




Re-declare,
For fio, i compare "bw" and "iops" fileds.
For iozone, i compare "write" "rewrite" "read" "re-read" fileds.




On Mon, Jun 5, 2017 at 11:08 AM, Xie Changlong
wrote:


在 6/5/2017 1:22 PM, Pranith Kumar Karampuri 写道:


On Mon, Jun 5, 2017 at 8:29 AM, Xie Changlong
wrote:

在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:
I think there have been improvements here to use special 
instructions to




--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/5/2017 2:34 PM, Pranith Kumar Karampuri 写道:

What is your usecase if I may ask?



Only fio/iozone now, maybe you can give me better test sugguestion :)

seq read
# dd if=/dev/zero  of=test bs=4k count=524288
#fio --filename=test -iodepth=64 -ioengine=libaio --direct=1 --rw=read 
--bs=1m --size=2g --numjobs=4 --runtime=10 --group_reporting 
--name=test-read

seq write
# fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=write 
-bs=1m -size=2g -numjobs=4 -runtime=20 -group_reporting -name=test-write


random read
#fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randread 
-bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting 
-name=test-rand-read


random write
#fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randwrite 
-bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting 
-name=test-rand-write



iozone  1 thread
# ./iozone -s 10g -r 1m -i 0 -i 1 -f /mnt/test-volume/test -Rb 
/tmp/test_iozone_1.xls

iozone  2 multi-threads
# ./iozone -s 10g -r 1m -i 0 -i 1 -f /mnt/test-volume/test -t 8 -Rb 
/tmp/test_iozone_2.xls




On Mon, Jun 5, 2017 at 11:08 AM, Xie Changlong
wrote:


在 6/5/2017 1:22 PM, Pranith Kumar Karampuri 写道:


On Mon, Jun 5, 2017 at 8:29 AM, Xie Changlong
wrote:

在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:

I think there have been improvements here to use special instructions to


--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

What is your usecase if I may ask?

On Mon, Jun 5, 2017 at 11:08 AM, Xie Changlong 
wrote:

> 在 6/5/2017 1:22 PM, Pranith Kumar Karampuri 写道:
>
>> On Mon, Jun 5, 2017 at 8:29 AM, Xie Changlong 
>> wrote:
>>
>> 在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:
>>>
>>> I think there have been improvements here to use special instructions to


>>> if "special instructions" means profile command or more? Also can you
>>> point which commit?
>>>
>>>
>> I meant Intel CPU extentions
>> You should look at
>> http://review.gluster.org/12209
>> http://review.gluster.org/12210
>> https://review.gluster.org/16963
>> https://review.gluster.org/17009
>>
>
> Useful enough for me, thanks for your patience.
>
>
>>
>> do the increments instead of taking spin-locks and doing increments. So
>>>
 may be it doesn't affect performance as much anymore. I think if you
 don't
 see a difference, then the enhancements are doing a good job :-).

 Which version of gluster are you using?


>>> Our version is gluster 3.8.4
>>>
>>> --
>>> Thanks
>>>  -Xie
>>>
>>>
>>
>>
>>
> --
> Thanks
> -Xie
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/5/2017 1:22 PM, Pranith Kumar Karampuri 写道:

On Mon, Jun 5, 2017 at 8:29 AM, Xie Changlong 
wrote:


在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:


I think there have been improvements here to use special instructions to



if "special instructions" means profile command or more? Also can you
point which commit?



I meant Intel CPU extentions
You should look at
http://review.gluster.org/12209
http://review.gluster.org/12210
https://review.gluster.org/16963
https://review.gluster.org/17009


Useful enough for me, thanks for your patience.





do the increments instead of taking spin-locks and doing increments. So

may be it doesn't affect performance as much anymore. I think if you don't
see a difference, then the enhancements are doing a good job :-).

Which version of gluster are you using?



Our version is gluster 3.8.4

--
Thanks
 -Xie







--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

On Mon, Jun 5, 2017 at 8:29 AM, Xie Changlong 
wrote:

> 在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:
>
>> I think there have been improvements here to use special instructions to
>>
>
> if "special instructions" means profile command or more? Also can you
> point which commit?
>

I meant Intel CPU extentions
You should look at
http://review.gluster.org/12209
http://review.gluster.org/12210
https://review.gluster.org/16963
https://review.gluster.org/17009


> do the increments instead of taking spin-locks and doing increments. So
>> may be it doesn't affect performance as much anymore. I think if you don't
>> see a difference, then the enhancements are doing a good job :-).
>>
>> Which version of gluster are you using?
>>
>
> Our version is gluster 3.8.4
>
> --
> Thanks
> -Xie
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Xie Changlong]

在 6/5/2017 10:50 AM, Pranith Kumar Karampuri 写道:
I think there have been improvements here to use special instructions to 


if "special instructions" means profile command or more? Also can you 
point which commit?


do the increments instead of taking spin-locks and doing increments. So 
may be it doesn't affect performance as much anymore. I think if you 
don't see a difference, then the enhancements are doing a good job :-).


Which version of gluster are you using?


Our version is gluster 3.8.4

--
Thanks
-Xie
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel

Re: [Gluster-devel] [Qusetions] How profile affect glusterfs performance?

[Pranith Kumar Karampuri]

I think there have been improvements here to use special instructions to do
the increments instead of taking spin-locks and doing increments. So may be
it doesn't affect performance as much anymore. I think if you don't see a
difference, then the enhancements are doing a good job :-).

Which version of gluster are you using?

On Mon, Jun 5, 2017 at 8:09 AM, Xie Changlong 
wrote:

> Hi all
>
> It's said[1] that profile based on io-stats, if you enable this feature,
> it can affect system performance while the profile information is being
> collected.
>
> I do some tests on my two linux+vmware virtual machine with replica(lack
> of resources ). And the results shows no diffrence to me, following is the
> test case
> #dd if=/dev/zero  of=test bs=4k count=524288
> #fio --filename=test -iodepth=64 -ioengine=libaio --direct=1 --rw=read
> --bs=1m --size=2g --numjobs=4 --runtime=10 --group_reporting
> --name=test-read
> #fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=write
> -bs=1m -size=2g -numjobs=4 -runtime=20 -group_reporting -name=test-write
> #fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randread
> -bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting
> -name=test-rand-read
> #fio -filename=test -iodepth=64 -ioengine=libaio -direct=1 -rw=randwrite
> -bs=4k -size=2G -numjobs=64 -runtime=20 -group_reporting
> -name=test-rand-write
> It's said that fio is only for lagre files, also i suspect that the test
> infrastructure is too small. The question is that, if you guys have
> detailed data for how profile affect performance?
>
> More, we wanna gain the detail r/w iops/bandwidth data for each brick. It
> seems that only profile can provide relatived data to be calculated?if i'm
> wrong pls corrent me.
>
> If profile really affect peformance so much, would you mind a new command
> such as "gluster volume io [nfs]" to acquire brick r/w fops/data? Or just
> help us review it?
>
> [1] https://access.redhat.com/documentation/en-us/red_hat_gluste
> r_storage/3.2/html/administration_guide/chap-monitoring_red_
> hat_storage_workload#sect-Running_the_Volume_Profile_Command
> --
> Thanks
> -Xie
> ___
> Gluster-devel mailing list
> Gluster-devel@gluster.org
> http://lists.gluster.org/mailman/listinfo/gluster-devel
>



-- 
Pranith
___
Gluster-devel mailing list
Gluster-devel@gluster.org
http://lists.gluster.org/mailman/listinfo/gluster-devel