[Lustre-devel] rename in Changelog

[Lustre-devel] rename in Changelog

[Lai Siyao]

Rename record in Changelog is different from other operations, it's split
into
two records: RNMFRM and RNMTO. This makes Changelog analysis hard
because these two records may not be consecutive and several renames
may occur at the same time.

I'm not clear why it's designed to be so, except that RNMTO is needed
because for DNE (distributed namespace) the rename target may reside on
another MDS, and a separate RNMTO record is needed. But even with this
it's fine to store all information in a RENME record, but leave the
information of
whether rename removes the last hardlink of the target file (if it exists)
in
RNMTO record.

I tried to add a field spfid in struct changelog_rec to store source parent
fid,
and pack both source (if has) and target names into record. Normally the
record size if sizeof(fid) larger than before, and it can be differentiated
according to version. The test result looks good, but I want to know whether
anyone oppose to this? If not, I'll make the change and make it changelog
version 2.

Cheers,
- Lai
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20120425/e2f038d1/attachment.htm>

[Lustre-devel] rename in Changelog

[Nathan Rutman]

it wasn't done that way in the first place because the record size in an llog is fixed,
so any size increase is multiplied by the number of records, so fewer records
can be stored.  Splitting the rename into two was the unfortunate casualty of that goal.

On Apr 25, 2012, at 2:05 AM, Lai Siyao wrote:

> Rename record in Changelog is different from other operations, it's split into
> two records: RNMFRM and RNMTO. This makes Changelog analysis hard
> because these two records may not be consecutive and several renames
> may occur at the same time.
> 
> I'm not clear why it's designed to be so, except that RNMTO is needed
> because for DNE (distributed namespace) the rename target may reside on
> another MDS, and a separate RNMTO record is needed. But even with this
> it's fine to store all information in a RENME record, but leave the information of
> whether rename removes the last hardlink of the target file (if it exists) in
> RNMTO record.
> 
> I tried to add a field spfid in struct changelog_rec to store source parent fid,
> and pack both source (if has) and target names into record. Normally the
> record size if sizeof(fid) larger than before, and it can be differentiated
> according to version. The test result looks good, but I want to know whether
> anyone oppose to this? If not, I'll make the change and make it changelog
> version 2.
> 
> Cheers,
> - Lai
> 
> 
> _______________________________________________
> Lustre-devel mailing list
> Lustre-devel at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-devel

[Lustre-devel] rename in Changelog

[Mikhail Pershin]

On Wed, 25 Apr 2012 20:36:08 +0400, Nathan Rutman  
<Nathan_Rutman@xyratex.com> wrote:

> it wasn't done that way in the first place because the record size in an  
> llog is fixed,
> so any size increase is multiplied by the number of records, so fewer  
> records
> can be stored.  Splitting the rename into two was the unfortunate  
> casualty of that goal.

This is true if we are using the same record for all operations and Lai  
did so by adding extra field. But I'd note that fixed size is not  
mandatory thing but good to have. Moreover, changelog records are  
naturally not fixed size because contain name which size is not fixed. So  
I suppose rename was done in two parts just as simpler way to go because  
required no additional changes in processing, etc. Also, IIRC, CL_EXT  
appeared even later than CL_RENAME itself.

To avoid space consuming we can just introduce extended  
'changelog_ext_rec' for rename. It is not the problem to create such  
changelogs but requires processing tools to be aware about that, is that a  
big problem or acceptable?

>
> On Apr 25, 2012, at 2:05 AM, Lai Siyao wrote:
>
>> Rename record in Changelog is different from other operations, it's  
>> split into
>> two records: RNMFRM and RNMTO. This makes Changelog analysis hard
>> because these two records may not be consecutive and several renames
>> may occur at the same time.
>>
>> I'm not clear why it's designed to be so, except that RNMTO is needed
>> because for DNE (distributed namespace) the rename target may reside on
>> another MDS, and a separate RNMTO record is needed. But even with this
>> it's fine to store all information in a RENME record, but leave the  
>> information of
>> whether rename removes the last hardlink of the target file (if it  
>> exists) in
>> RNMTO record.
>>
>> I tried to add a field spfid in struct changelog_rec to store source  
>> parent fid,
>> and pack both source (if has) and target names into record. Normally the
>> record size if sizeof(fid) larger than before, and it can be  
>> differentiated
>> according to version. The test result looks good, but I want to know  
>> whether
>> anyone oppose to this? If not, I'll make the change and make it  
>> changelog
>> version 2.
>>
>> Cheers,
>> - Lai
>>
>>
>> _______________________________________________
>> Lustre-devel mailing list
>> Lustre-devel at lists.lustre.org
>> http://lists.lustre.org/mailman/listinfo/lustre-devel
> _______________________________________________
> Lustre-devel mailing list
> Lustre-devel at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-devel

[Lustre-devel] rename in Changelog

[Lai Siyao]

Mike, you're right, only rename needs to use the new version of changelog
format; and
it will be unpacked in liblustreapi, the user space tool will not even
notice it.

On Thu, Apr 26, 2012 at 2:06 AM, Mikhail Pershin <mike.tappro@gmail.com>wrote:

> On Wed, 25 Apr 2012 20:36:08 +0400, Nathan Rutman <
> Nathan_Rutman at xyratex.com> wrote:
>
>  it wasn't done that way in the first place because the record size in an
>> llog is fixed,
>> so any size increase is multiplied by the number of records, so fewer
>> records
>> can be stored.  Splitting the rename into two was the unfortunate
>> casualty of that goal.
>>
>
> This is true if we are using the same record for all operations and Lai
> did so by adding extra field. But I'd note that fixed size is not mandatory
> thing but good to have. Moreover, changelog records are naturally not fixed
> size because contain name which size is not fixed. So I suppose rename was
> done in two parts just as simpler way to go because required no additional
> changes in processing, etc. Also, IIRC, CL_EXT appeared even later than
> CL_RENAME itself.
>
> To avoid space consuming we can just introduce extended
> 'changelog_ext_rec' for rename. It is not the problem to create such
> changelogs but requires processing tools to be aware about that, is that a
> big problem or acceptable?
>
>
>
>> On Apr 25, 2012, at 2:05 AM, Lai Siyao wrote:
>>
>>  Rename record in Changelog is different from other operations, it's
>>> split into
>>> two records: RNMFRM and RNMTO. This makes Changelog analysis hard
>>> because these two records may not be consecutive and several renames
>>> may occur at the same time.
>>>
>>> I'm not clear why it's designed to be so, except that RNMTO is needed
>>> because for DNE (distributed namespace) the rename target may reside on
>>> another MDS, and a separate RNMTO record is needed. But even with this
>>> it's fine to store all information in a RENME record, but leave the
>>> information of
>>> whether rename removes the last hardlink of the target file (if it
>>> exists) in
>>> RNMTO record.
>>>
>>> I tried to add a field spfid in struct changelog_rec to store source
>>> parent fid,
>>> and pack both source (if has) and target names into record. Normally the
>>> record size if sizeof(fid) larger than before, and it can be
>>> differentiated
>>> according to version. The test result looks good, but I want to know
>>> whether
>>> anyone oppose to this? If not, I'll make the change and make it changelog
>>> version 2.
>>>
>>> Cheers,
>>> - Lai
>>>
>>>
>>> ______________________________**_________________
>>> Lustre-devel mailing list
>>> Lustre-devel at lists.lustre.org
>>> http://lists.lustre.org/**mailman/listinfo/lustre-devel<http://lists.lustre.org/mailman/listinfo/lustre-devel>
>>>
>> ______________________________**_________________
>> Lustre-devel mailing list
>> Lustre-devel at lists.lustre.org
>> http://lists.lustre.org/**mailman/listinfo/lustre-devel<http://lists.lustre.org/mailman/listinfo/lustre-devel>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20120426/693345da/attachment.htm>

[Lustre-devel] rename in Changelog

[Alexey Lyashkov]

but it's introduce the race. different mdt thread my write a one or more llog record between rename records.
so it's isn't always two next records.

On Apr 25, 2012, at 23:36, Nathan Rutman wrote:

> it wasn't done that way in the first place because the record size in an llog is fixed,
> so any size increase is multiplied by the number of records, so fewer records
> can be stored.  Splitting the rename into two was the unfortunate casualty of that goal.
> 
> On Apr 25, 2012, at 2:05 AM, Lai Siyao wrote:
> 
>> Rename record in Changelog is different from other operations, it's split into
>> two records: RNMFRM and RNMTO. This makes Changelog analysis hard
>> because these two records may not be consecutive and several renames
>> may occur at the same time.
>> 
>> I'm not clear why it's designed to be so, except that RNMTO is needed
>> because for DNE (distributed namespace) the rename target may reside on
>> another MDS, and a separate RNMTO record is needed. But even with this
>> it's fine to store all information in a RENME record, but leave the information of
>> whether rename removes the last hardlink of the target file (if it exists) in
>> RNMTO record.
>> 
>> I tried to add a field spfid in struct changelog_rec to store source parent fid,
>> and pack both source (if has) and target names into record. Normally the
>> record size if sizeof(fid) larger than before, and it can be differentiated
>> according to version. The test result looks good, but I want to know whether
>> anyone oppose to this? If not, I'll make the change and make it changelog
>> version 2.
>> 
>> Cheers,
>> - Lai
>> 
>> 
>> _______________________________________________
>> Lustre-devel mailing list
>> Lustre-devel at lists.lustre.org
>> http://lists.lustre.org/mailman/listinfo/lustre-devel
> _______________________________________________
> Lustre-devel mailing list
> Lustre-devel at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-devel