Hi Andreas,
On 06/06/2014 13:46, Andreas Joachim Peters wrote:> Hi Loic,
> the basic implementation looks very clean.
>
> I have few comments/ideas:
>
> - the reconstruction strategy using the three levels is certainly efficient enough for standard cases but does not guarantee always the minimum decoding (in cases where one layer is not enough to reconstruct) since your third algorithm is just brute-force to reconstruct everything through all layers until we have what we need ...
The third strategy is indeed brute force. Do you think it is worth changing to be minimal ? It would be nice to quantify the percent of cases it addresses. Do you know how to do that ? It looks like a very small percentage but there is no proof it is small ;-)
> - the whole LRC configuration actually does not describe the placement - it still looks disconnected from the placement strategy/crush rules ... wouldn't it make sense to have the crush rule implicit in the description or a function to derive it automatically based on the LRC configuration? Maybe you have this already done in another way and I didn't see it ...
Good catch.
What about this:
" [ \"_aAAA_aAA_\", \"set choose datacenter 2\","
" \"_aXXX_aXX_\" ],"
" [ \"b_BBB_____\", \"set choose host 5\","
" \"baXXX_____\" ],"
" [ \"_____cCCC_\", \"\","
" \"baXXXcaXX_\" ],"
" [ \"_____DDDDd\", \"\","
" \"baXXXcaXXd\" ],"
Which translates into
take root
set choose datacenter 2
set choose host 5
In other words, the ruleset is created by concatenating the strings from the description, without any kind of smart computation. It is up to the person who creates the description to add the ruleset near a description that makes sense. There is going to be minimal checking to make sure the ruleset can actually be used to get the required number of chunks.
It probably is very difficult and very confusing to automate the generation of the ruleset. If it is implicit rather than explicit as above, the operator will have to somehow understand and learn how it is computed to make sure it does what is desired. With an explicit set of crush rules loosely coupled to chunk mapping, the operator can read the crush documentation instead of guessing.
> - should the plug-in have the ability to select reconstruction on proximity or this should be up-to the higher layer to provide chunks in a way that reconstruction would select the 'closest' layer? The relevance of the question you will understand better in the next point ....
>
> - I remember we had this 3 data centre example with (8,4) where you can reconstruct every object if 2 data centres are up. Another appealing example avoiding remote access when reading an object is that you have 2 data centres having a replication of e.g. (4,2) encoded objects. Can you describe in your LRC configuration language to store the same chunk twice like __ABCCBA__ ?
Unless I'm mistaken that would require the caller of the plugin to support duplicate data chunks and provide a kind of proximity check. Since this is not currently supported by the OSD logic, it is difficult to figure out how an erasure code plugin could provide support for this use case.
Cheers
>
> Cheers Andreas.
>
>
> ________________________________________
> From: Loic Dachary [loic@xxxxxxxxxxx]
> Sent: 05 June 2014 16:05
> To: Andreas Joachim Peters
> Cc: Ceph Development
> Subject: Re: Locally repairable code description revisited (was Pyramid ...)
>
> Hi Andreas,
>
> Here is a preliminary implementation https://github.com/ceph/ceph/pull/1921 . It does not have tests yet, nor does it run but but it compiles and I think it makes sense. It is indeed simpler than the previous implementation.
>
> It would be great if you could quickly browse
>
> https://github.com/dachary/ceph/commit/5661c72864eca04c5dac9494a3f7e7a13515fca6
>
> and let me know if you have any concerns.
>
> Cheers
>
> On 31/05/2014 19:10, Loic Dachary wrote:
>> Hi Andreas,
>>
>> After a few weeks and a fresh eye, I revisited the way pyramid erasure code could be described by the system administrator. Here is a proposal that is hopefully more intuitive than the one from the last CDS ( http://pad.ceph.com/p/cdsgiant-pyramid-erasure-code ).
>>
>> These are the steps to create all coding chunks. The upper case letters are data chunks and the lower case letters are coding chunks.
>>
>> "__ABC__DE_" data chunks placement
>>
>> Step 1
>> "__ABC__DE_"
>> "_yVWX_zYZ_" K=5, M=2
>> "_aABC_bDE_"
>>
>> Step 2
>> "_aABC_bDE_"
>> "z_XYZ_____" K=3, M=1
>> "caABC_bDE_"
>>
>> Step 3
>> "caABC_bDE_"
>> "_____zXYZ_" K=3, M=1
>> "caABCdbDE_"
>>
>> Step 4
>> "caABCdbDE_"
>> "_____WXYZz" K=4, M=1
>> "caABCdbDEe"
>>
>> The interpretation of Step 3 is as follows:
>>
>> Given the output of the previous step ( "caABC_bDE_" ), the bDE chunks are considered to be data chunks at this stage and they are marked with XYZ. A K=3, M=1 coding chunk is calculated and placed in the chunk marked with z ( "_____zXYZ_" ). The output of this coding step is the previous step plus the coding chunk that was just calculated, named d ( "caABCdbDE_" ).
>>
>> This gives the flexibility of deciding wether or not a coding chunk from a previous step is used as data to compute the coding chunk of the next step. It also allows for unbalanced steps such as step 4.
>>
>> For decoding, the steps are walked from the bottom up. If E is missing, it can be reconstructed from dbD.e in step 4 and the other steps are skipped because it was the only missing chunk. If AB are missing, all steps that have not be used to encode it are ignored, up to step 2 that will fail to recover them because M=1 and yeild to step 1 that will use a..CbDE successfully because M=2.
>>
>> Giving up the recursion and favor iteration seems to simplify how it can be explained. And I suspect the implementation is also simpler. What do you think ?
>>
>> Cheers
>>
>
> --
> Loïc Dachary, Artisan Logiciel Libre
>
--
Loïc Dachary, Artisan Logiciel Libre
Attachment:
signature.asc
Description: OpenPGP digital signature
