[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jensen]

Hi all,
   I found ocfs1.0 use the Distributed lock manager by disk. on ocfs2 why give up? because performance or other reason?

   Btw:
     on the ocfs1.0 why not use scsi-2/3 reservation to protect to update the dlm sector data?

Thanks,
Jensen.

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jensen]

Hi Mark, sunil, jeff an Joel,
   Do you know why ? Thanks.

Jensen.
2013-7-2

On 2013/6/29 11:27, Jensen wrote:

> Hi all,
>    I found ocfs1.0 use the Distributed lock manager by disk. on ocfs2 why give up? because performance or other reason?
> 
>    Btw:
>      on the ocfs1.0 why not use scsi-2/3 reservation to protect to update the dlm sector data?
> 
> Thanks,
> Jensen.
> 
> 
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
> 
> 

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Sunil Mushran]

A general purpose file system requires one to manage over a million locks
concurrently. So performance is the main reason.


On Mon, Jul 1, 2013 at 6:22 PM, Jensen <shencanquan@huawei.com> wrote:

> Hi Mark, sunil, jeff an Joel,
>    Do you know why ? Thanks.
>
> Jensen.
> 2013-7-2
>
> On 2013/6/29 11:27, Jensen wrote:
>
> > Hi all,
> >    I found ocfs1.0 use the Distributed lock manager by disk. on ocfs2
> why give up? because performance or other reason?
> >
> >    Btw:
> >      on the ocfs1.0 why not use scsi-2/3 reservation to protect to
> update the dlm sector data?
> >
> > Thanks,
> > Jensen.
> >
> >
> > _______________________________________________
> > Ocfs2-devel mailing list
> > Ocfs2-devel at oss.oracle.com
> > https://oss.oracle.com/mailman/listinfo/ocfs2-devel
> >
> >
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20130701/0df7a53b/attachment.html 

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jensen]

On 2013/7/2 9:35, Sunil Mushran wrote:

> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
> 

Thanks for your comments.

Has Oracle compared the performance between ocfs2 and ocfs1?

> 
> On Mon, Jul 1, 2013 at 6:22 PM, Jensen <shencanquan at huawei.com <mailto:shencanquan@huawei.com>> wrote:
> 
>     Hi Mark, sunil, jeff an Joel,
>        Do you know why ? Thanks.
> 
>     Jensen.
>     2013-7-2
> 
>     On 2013/6/29 11:27, Jensen wrote:
> 
>     > Hi all,
>     >    I found ocfs1.0 use the Distributed lock manager by disk. on ocfs2 why give up? because performance or other reason?
>     >
>     >    Btw:
>     >      on the ocfs1.0 why not use scsi-2/3 reservation to protect to update the dlm sector data?
>     >
>     > Thanks,
>     > Jensen.
>     >
>     >
>     > _______________________________________________
>     > Ocfs2-devel mailing list
>     > Ocfs2-devel at oss.oracle.com <mailto:Ocfs2-devel@oss.oracle.com>
>     > https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>     >
>     >
> 
> 
> 
> 

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Mark Fasheh]

On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
> On 2013/7/2 9:35, Sunil Mushran wrote:
> 
> > A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
> > 
> 
> Thanks for your comments.
> 
> Has Oracle compared the performance between ocfs2 and ocfs1?

Firstly, that's implied in the answer you just got. Also, who wouldn't
compare performance from one version of a file system to the next?

Can you please cut to the chase and either ask what you really want to know
or make the statement you're trying to make so we can move on?

Thanks,
	--Mark

--
Mark Fasheh

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jensen]

On 2013/7/3 1:20, Mark Fasheh wrote:

> On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
>> On 2013/7/2 9:35, Sunil Mushran wrote:
>>
>>> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
>>>
>>
>> Thanks for your comments.
>>
>> Has Oracle compared the performance between ocfs2 and ocfs1?
> 
> Firstly, that's implied in the answer you just got. Also, who wouldn't
> compare performance from one version of a file system to the next?
> 
> Can you please cut to the chase and either ask what you really want to know
> or make the statement you're trying to make so we can move on?
> 


Thanks for your answer.
we want to use compare and write scsi command to replace DLM module. it is similar
with vmware vmfs.

why we want to replace dlm in ocfs2? because:
1. The stability of ocfs2 dlm is very poor, we found 100+ bug.
2. The Reliability of ocfs2 dlm is very poor, especially in network split, the
   worse case split two domain, the half of node must be reboot.
3. the maximum number of mounted machine is 32, we want to support more.

currently we worry about two thing:
1. The performance lock and unlock. because it use the scsi command and it is similar with IO read and write.
2. The change is very very large. because it maybe modify the disk layout of ocfs2.

so anyone interested with this?

> Thanks,
> 	--Mark
> 
> --
> Mark Fasheh
> 
> 

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jeff Liu]

On 07/03/2013 09:27 AM, Jensen wrote:

> On 2013/7/3 1:20, Mark Fasheh wrote:
> 
>> On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
>>> On 2013/7/2 9:35, Sunil Mushran wrote:
>>>
>>>> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
>>>>
>>>
>>> Thanks for your comments.
>>>
>>> Has Oracle compared the performance between ocfs2 and ocfs1?
>>
>> Firstly, that's implied in the answer you just got. Also, who wouldn't
>> compare performance from one version of a file system to the next?
>>
>> Can you please cut to the chase and either ask what you really want to know
>> or make the statement you're trying to make so we can move on?
>>
> 
> 
> Thanks for your answer.
> we want to use compare and write scsi command to replace DLM module. it is similar
> with vmware vmfs.
> 

I'm not trying to answer this question.

I knew that OCFS2 is deployed in HuaWei.com in a large-scale cluster up
to 128 nodes, so I'm not very much surprised at something you mentioned
below, but...

> why we want to replace dlm in ocfs2? because:
> 1. The stability of ocfs2 dlm is very poor, we found 100+ bug.

That sounds interesting, how to classify those problems?
- Fatal error, panic
- Result in an interruption in service
- Wrong results, but can work around?
- Trivial

Reporting bugs to bugzilla/OCFS2 would be useful to keep track of them:
https://oss.oracle.com/bugzilla/

Thanks,
-Jeff

> 2. The Reliability of ocfs2 dlm is very poor, especially in network split, the
>    worse case split two domain, the half of node must be reboot.
> 3. the maximum number of mounted machine is 32, we want to support more.
> 
> currently we worry about two thing:
> 1. The performance lock and unlock. because it use the scsi command and it is similar with IO read and write.
> 2. The change is very very large. because it maybe modify the disk layout of ocfs2.
> 
> so anyone interested with this?
> 
>> Thanks,
>> 	--Mark
>>
>> --
>> Mark Fasheh
>>
>>
> 
> 
> 
> 
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jensen]

On 2013/7/3 13:06, Jeff Liu wrote:

> On 07/03/2013 09:27 AM, Jensen wrote:
> 
>> On 2013/7/3 1:20, Mark Fasheh wrote:
>>
>>> On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
>>>> On 2013/7/2 9:35, Sunil Mushran wrote:
>>>>
>>>>> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
>>>>>
>>>>
>>>> Thanks for your comments.
>>>>
>>>> Has Oracle compared the performance between ocfs2 and ocfs1?
>>>
>>> Firstly, that's implied in the answer you just got. Also, who wouldn't
>>> compare performance from one version of a file system to the next?
>>>
>>> Can you please cut to the chase and either ask what you really want to know
>>> or make the statement you're trying to make so we can move on?
>>>
>>
>>
>> Thanks for your answer.
>> we want to use compare and write scsi command to replace DLM module. it is similar
>> with vmware vmfs.
>>
> 
> I'm not trying to answer this question.
> 
> I knew that OCFS2 is deployed in HuaWei.com in a large-scale cluster up
> to 128 nodes, so I'm not very much surprised at something you mentioned
> below, but...
> 
>> why we want to replace dlm in ocfs2? because:
>> 1. The stability of ocfs2 dlm is very poor, we found 100+ bug.
> 
> That sounds interesting, how to classify those problems?
> - Fatal error, panic
> - Result in an interruption in service
> - Wrong results, but can work around?
> - Trivial
> 
> Reporting bugs to bugzilla/OCFS2 would be useful to keep track of them:
> https://oss.oracle.com/bugzilla/
> 


Recently, Huawei has send many bug to open source community. which bug exist in
open source. other change or bug belong to enforce ocfs2 function ,for example, when disk
timeout, open source code will reboot the machine, we modify it, we set the ocfs2
invalid(can't read and write), open source may not interesting with those modify.

> Thanks,
> -Jeff
> 
>> 2. The Reliability of ocfs2 dlm is very poor, especially in network split, the
>>    worse case split two domain, the half of node must be reboot.
>> 3. the maximum number of mounted machine is 32, we want to support more.
>>
>> currently we worry about two thing:
>> 1. The performance lock and unlock. because it use the scsi command and it is similar with IO read and write.
>> 2. The change is very very large. because it maybe modify the disk layout of ocfs2.
>>
>> so anyone interested with this?
>>
>>> Thanks,
>>> 	--Mark
>>>
>>> --
>>> Mark Fasheh
>>>
>>>
>>
>>
>>
>>
>> _______________________________________________
>> Ocfs2-devel mailing list
>> Ocfs2-devel at oss.oracle.com
>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
> 
> 
> 
> 

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jeff Liu]

On 07/03/2013 04:28 PM, Jensen wrote:

> On 2013/7/3 13:06, Jeff Liu wrote:
> 
>> On 07/03/2013 09:27 AM, Jensen wrote:
>>
>>> On 2013/7/3 1:20, Mark Fasheh wrote:
>>>
>>>> On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
>>>>> On 2013/7/2 9:35, Sunil Mushran wrote:
>>>>>
>>>>>> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
>>>>>>
>>>>>
>>>>> Thanks for your comments.
>>>>>
>>>>> Has Oracle compared the performance between ocfs2 and ocfs1?
>>>>
>>>> Firstly, that's implied in the answer you just got. Also, who wouldn't
>>>> compare performance from one version of a file system to the next?
>>>>
>>>> Can you please cut to the chase and either ask what you really want to know
>>>> or make the statement you're trying to make so we can move on?
>>>>
>>>
>>>
>>> Thanks for your answer.
>>> we want to use compare and write scsi command to replace DLM module. it is similar
>>> with vmware vmfs.
>>>
>>
>> I'm not trying to answer this question.
>>
>> I knew that OCFS2 is deployed in HuaWei.com in a large-scale cluster up
>> to 128 nodes, so I'm not very much surprised at something you mentioned
>> below, but...
>>
>>> why we want to replace dlm in ocfs2? because:
>>> 1. The stability of ocfs2 dlm is very poor, we found 100+ bug.
>>
>> That sounds interesting, how to classify those problems?
>> - Fatal error, panic
>> - Result in an interruption in service
>> - Wrong results, but can work around?
>> - Trivial
>>
>> Reporting bugs to bugzilla/OCFS2 would be useful to keep track of them:
>> https://oss.oracle.com/bugzilla/
>>
> 
> 
> Recently, Huawei has send many bug to open source community. which bug exist in
> open source. 

We always appreciating the contributions.

> other change or bug belong to enforce ocfs2 function ,for example, when disk
> timeout, open source code will reboot the machine, we modify it, we set the ocfs2
> invalid(can't read and write), open source may not interesting with those modify.

Why not give a try if those changes are fair enough?

Look at what you mentioned and what I asked above:
">>> 1. The stability of ocfs2 dlm is very poor, we found 100+ bug."

If take away around 10 DLM related patches from HuaWei, so the left several DLM
bugs(might be) are totally can not be fixed in mainline?

This sort of replies can provide nothing useful information to the community and
it is in fact a waste of our(yours and my) time.

I understand that we have a language barrier for the communication to the open
source community, but some of your questions/feedbacks are deserve to spend a few
days to think it over before sending it out to the public, so that someone who is
capable to answer your questions can well understand your opinions. :)

> 
>> Thanks,
>> -Jeff
>>
>>> 2. The Reliability of ocfs2 dlm is very poor, especially in network split, the
>>>    worse case split two domain, the half of node must be reboot.
>>> 3. the maximum number of mounted machine is 32, we want to support more.
>>>
>>> currently we worry about two thing:
>>> 1. The performance lock and unlock. because it use the scsi command and it is similar with IO read and write.
>>> 2. The change is very very large. because it maybe modify the disk layout of ocfs2.
>>>
>>> so anyone interested with this?
>>>
>>>> Thanks,
>>>> 	--Mark
>>>>
>>>> --
>>>> Mark Fasheh
>>>>
>>>>
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> Ocfs2-devel mailing list
>>> Ocfs2-devel at oss.oracle.com
>>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>>
>>
>>
>>
> 
> 
> 

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Srinivas Eeda]

On 07/03/2013 01:28 AM, Jensen wrote:
> On 2013/7/3 13:06, Jeff Liu wrote:
>
>> On 07/03/2013 09:27 AM, Jensen wrote:
>>
>>> On 2013/7/3 1:20, Mark Fasheh wrote:
>>>
>>>> On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
>>>>> On 2013/7/2 9:35, Sunil Mushran wrote:
>>>>>
>>>>>> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
>>>>>>
>>>>> Thanks for your comments.
>>>>>
>>>>> Has Oracle compared the performance between ocfs2 and ocfs1?
>>>> Firstly, that's implied in the answer you just got. Also, who wouldn't
>>>> compare performance from one version of a file system to the next?
>>>>
>>>> Can you please cut to the chase and either ask what you really want to know
>>>> or make the statement you're trying to make so we can move on?
>>>>
>>>
>>> Thanks for your answer.
>>> we want to use compare and write scsi command to replace DLM module. it is similar
>>> with vmware vmfs.
>>>
>> I'm not trying to answer this question.
>>
>> I knew that OCFS2 is deployed in HuaWei.com in a large-scale cluster up
>> to 128 nodes, so I'm not very much surprised at something you mentioned
>> below, but...
>>
>>> why we want to replace dlm in ocfs2? because:
>>> 1. The stability of ocfs2 dlm is very poor, we found 100+ bug.
>> That sounds interesting, how to classify those problems?
>> - Fatal error, panic
>> - Result in an interruption in service
>> - Wrong results, but can work around?
>> - Trivial
>>
>> Reporting bugs to bugzilla/OCFS2 would be useful to keep track of them:
>> https://oss.oracle.com/bugzilla/
>>
>
> Recently, Huawei has send many bug to open source community.
Yes and we appreciate your contributions :)
>   which bug exist in
> open source. other change or bug belong to enforce ocfs2 function ,for example, when disk
> timeout, open source code will reboot the machine, we modify it, we set the ocfs2
> invalid(can't read and write), open source may not interesting with those modify.
Can you please point me to these patches, I would be very interested to 
look at. If the patches are safe for all kind of workloads and guarantee 
to stop in-flight i/o's then it makes sense to add them to mainline ocfs2.
>
>> Thanks,
>> -Jeff
>>
>>> 2. The Reliability of ocfs2 dlm is very poor, especially in network split, the
>>>     worse case split two domain, the half of node must be reboot.
Currently we are looking into this. It appears that most of the network 
issues we are seeing may not really caused by network layer. They are 
mostly false errors because o2net thread got busy. Once this happens, in 
some scenarios reconnect always fails which is a bug. If you are 
currently running into these problems please share so we can analyse them.
>>> 3. the maximum number of mounted machine is 32, we want to support more.
>>>
>>> currently we worry about two thing:
>>> 1. The performance lock and unlock. because it use the scsi command and it is similar with IO read and write.
>>> 2. The change is very very large. because it maybe modify the disk layout of ocfs2.
I am just curious, have you already made these changes? Are the changes 
scalable?
>>>
>>> so anyone interested with this?
>>>
>>>> Thanks,
>>>> 	--Mark
>>>>
>>>> --
>>>> Mark Fasheh
>>>>
>>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> Ocfs2-devel mailing list
>>> Ocfs2-devel at oss.oracle.com
>>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>>
>>
>>
>
>
>
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel

[Ocfs2-devel] why oracle give up dlm by disk on ocfs2? because performance?

[Jensen]

On 2013/7/4 0:42, Srinivas Eeda wrote:

> On 07/03/2013 01:28 AM, Jensen wrote:
>> On 2013/7/3 13:06, Jeff Liu wrote:
>>
>>> On 07/03/2013 09:27 AM, Jensen wrote:
>>>
>>>> On 2013/7/3 1:20, Mark Fasheh wrote:
>>>>
>>>>> On Tue, Jul 02, 2013 at 10:07:52AM +0800, Jensen wrote:
>>>>>> On 2013/7/2 9:35, Sunil Mushran wrote:
>>>>>>
>>>>>>> A general purpose file system requires one to manage over a million locks concurrently. So performance is the main reason.
>>>>>>>
>>>>>> Thanks for your comments.
>>>>>>
>>>>>> Has Oracle compared the performance between ocfs2 and ocfs1?
>>>>> Firstly, that's implied in the answer you just got. Also, who wouldn't
>>>>> compare performance from one version of a file system to the next?
>>>>>
>>>>> Can you please cut to the chase and either ask what you really want to know
>>>>> or make the statement you're trying to make so we can move on?
>>>>>
>>>>
>>>> Thanks for your answer.
>>>> we want to use compare and write scsi command to replace DLM module. it is similar
>>>> with vmware vmfs.
>>>>
>>> I'm not trying to answer this question.
>>>
>>> I knew that OCFS2 is deployed in HuaWei.com in a large-scale cluster up
>>> to 128 nodes, so I'm not very much surprised at something you mentioned
>>> below, but...
>>>
>>>> why we want to replace dlm in ocfs2? because:
>>>> 1. The stability of ocfs2 dlm is very poor, we found 100+ bug.
>>> That sounds interesting, how to classify those problems?
>>> - Fatal error, panic
>>> - Result in an interruption in service
>>> - Wrong results, but can work around?
>>> - Trivial
>>>
>>> Reporting bugs to bugzilla/OCFS2 would be useful to keep track of them:
>>> https://oss.oracle.com/bugzilla/
>>>
>>
>> Recently, Huawei has send many bug to open source community.
> Yes and we appreciate your contributions :)
>>   which bug exist in
>> open source. other change or bug belong to enforce ocfs2 function ,for example, when disk
>> timeout, open source code will reboot the machine, we modify it, we set the ocfs2
>> invalid(can't read and write), open source may not interesting with those modify.
> Can you please point me to these patches, I would be very interested to 
> look at. If the patches are safe for all kind of workloads and guarantee 
> to stop in-flight i/o's then it makes sense to add them to mainline ocfs2.
>>
>>> Thanks,
>>> -Jeff
>>>
>>>> 2. The Reliability of ocfs2 dlm is very poor, especially in network split, the
>>>>     worse case split two domain, the half of node must be reboot.
> Currently we are looking into this. It appears that most of the network 
> issues we are seeing may not really caused by network layer. They are 
> mostly false errors because o2net thread got busy. Once this happens, in 
> some scenarios reconnect always fails which is a bug. If you are 
> currently running into these problems please share so we can analyse them.
>>>> 3. the maximum number of mounted machine is 32, we want to support more.
>>>>
>>>> currently we worry about two thing:
>>>> 1. The performance lock and unlock. because it use the scsi command and it is similar with IO read and write.
>>>> 2. The change is very very large. because it maybe modify the disk layout of ocfs2.
> I am just curious, have you already made these changes? Are the changes 
> scalable?

we had not made these changes. we plan to change this.

>>>>
>>>> so anyone interested with this?
>>>>
>>>>> Thanks,
>>>>> 	--Mark
>>>>>
>>>>> --
>>>>> Mark Fasheh
>>>>>
>>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Ocfs2-devel mailing list
>>>> Ocfs2-devel at oss.oracle.com
>>>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>>>
>>>
>>>
>>
>>
>>
>> _______________________________________________
>> Ocfs2-devel mailing list
>> Ocfs2-devel at oss.oracle.com
>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
> 
> 
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
> 
>