[dm-devel] Device-mapper target driver for Write tracking

Zdenek Kabelac zkabelac at redhat.com
Wed Oct 19 14:48:20 UTC 2016


Dne 19.10.2016 v 16:20 Daniel Armbruster napsal(a):
> On 13:43 Wed 19 Oct     , Zdenek Kabelac wrote:
>> Dne 19.10.2016 v 12:53 Daniel Armbruster napsal(a):
>>> Hi List,
>>>
>>> I'm looking for a device-mapper target driver giving me the possibility to
>>> track/log WRITE requests. The idea is to track those WRITE requests within a
>>> Bitmap e.g. 1:LV-EXTEND (for LVM2). I read already the device-mapper
>>> documentation under
>>>
>>> https://www.kernel.org/doc/Documentation/device-mapper/
>>>
>>> but didn't find an appropriate module which would suite my needs.
>>>
>>> Is there any work happening on this to leverage? Any kind of help is
>>> appreciated.
>>>
>>> Many thanks in advance.
>>
>>
>> Hi
>>
>>
>> Not exactly sure what you are trying to do but 2 things to look at:
>>
>> Maybe  'dmstats' with it's region logic may cover your 'wanted' operation.
>>
>> The kernel dm target to check would be  'era' which keeps trace
>> of written blocks.
>>
>> Regards
>>
>> Zdenek
>>
>>
>>
> Hi Zdenek,
>
> thanks for your help.
>
> I'm trying to implement a synchronization tool which performs backups in an
> intelligent way. Since using a snapshot is only partly an option (the snapshot
> could fill up and the snapshot_autoextend_* practicality of LVM2 isn't stable
> enough under heavy load) I was thinking about the possibility of an additional
> device which keeps trace of written blocks. To improve performance I aim at
> reducing the data throughput. That's why I was thinking about a Bitmap e.g.
> 1:LV-EXTEND.
>

Hi

Old snapshot is OLD - switch to thin-provisioning - that's the real fix for
your problem.

Less nice suggest could be to simply use BIG enough snapshot you can't fulfill 
before it gets resized -  computable from disk write speed.
Clearly you can't win this fight with few MB sized snapshot -
few GB solves it...

> I thought already about the DM statistics module. Is there any performance
> impact when collecting statistics for a given volume (except of the
> additionally allocated memory used)? Is there a possibility to quantify it?

I'm afraid you will need to measure yourself - depends on huge number of 
settings - but should be still very small compared with rotational drives.

Regards

Zdenek





More information about the dm-devel mailing list