Thank you Becky and Elaine for your replies.

Actually, I have no problem with this particular behavior. I just noticed
that it contrasted with what is available in the documentation.

Best,

On Wed, Sep 7, 2016 at 7:34 PM Elaine Quarles <[email protected]> wrote:

> Eduardo,
>
> We have confirmed that if file stuffing is turned off, servers to store
> the data are allocated as you expect in a round robin fashion. If file
> stuffing is turned on, then the order of servers is not guaranteed.
>
> If your use case depends on round robin, you can turn off stuffing by
> modifying the config file as shown below and restarting your servers.
>
> <Filesystem>
>     ...
>     FileStuffing no
>    ...
> </Filesystem>
>
> Hope that helps!
> Elaine
>
>
> On Tue, Sep 6, 2016 at 5:49 PM, Becky Ligon <[email protected]> wrote:
>
>> Eduardo:
>>
>> Looks like you have uncovered a bug.  So far, we have found the problem
>> as far back as 2.8.4.  We are still investigating and will keep you posted.
>>
>> thanks!
>> Becky Ligon
>>
>> On Tue, Sep 6, 2016 at 3:34 PM, Eduardo CAMILO Inacio <
>> [email protected]> wrote:
>>
>>> Dear PVFS2 users,
>>>
>>> After performing some experiments on OrangeFS 2.8.8, I am a little
>>> confused about the default distribution layout. Systems manual, change log,
>>> and some published papers claim that it should be ROUND_ROBIN. By that, I
>>> understand that a server is randomly chosen to store the first 64 KB of the
>>> file, and the second 64 KB block would be stored in the next server
>>> according to the order in the configuration file, and so on.
>>>
>>> In my experiments, what I am observing is more close to the behavior I
>>> would expect with a RANDOM layout. For instance, when two out of four
>>> servers are chosen to store objects from a file, independently if the
>>> "user.pvfs2.num_dfiles" is set to 2 or if the file size is just 2x the
>>> stripe size, it seems like the two servers are being chosen randomly.
>>>
>>> It is worth noting that no change was performed in the generated config
>>> file; the number of data files is changed through setfattr commands before
>>> experiments. POSIX requests are being generated using the IOR benchmark
>>> v3.0.1 with MPICH 3.2, which, I am assuming, is not changing the layout of
>>> the file system.
>>>
>>> Attached, you will find some output that illustrate the aforementioned
>>> behavior.
>>>
>>> Any insight on this subject would be helpful.
>>>
>>> Best,
>>> --
>>> Eduardo Camilo Inacio
>>>
>>> _______________________________________________
>>> Pvfs2-users mailing list
>>> [email protected]
>>> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
>>>
>>>
>>
>> _______________________________________________
>> Pvfs2-users mailing list
>> [email protected]
>> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
>>
>>
> --
Eduardo Camilo Inacio
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users

Reply via email to