[Darshan-users] compatibility with HDF5 1.10.x

Phil Carns carns at mcs.anl.gov
Tue Jan 17 14:34:06 CST 2017


That's bad news :(

That means that we might get the wrapper prototypes wrong at link time 
no matter how you compile Darshan.  The user could always load or link 
to a different version of HDF5. There won't be any link time warning.

We can probably query the HDF library at runtime to find out what 
version it is within the H5 open() wrappers, but even so, the only thing 
we could do if we detected a mismatch would be to error out and print 
"Please recompile darshan, this version is not binary ABI compatible 
with HDF API FOO". I can't think of any way to silently recover or even 
proceed without instrumentation.  By the time we detect the mismatch we 
will have already jumbled the arguments.

We could also try to make software module conflicts (so that if you 
module load a newer version of HDF it automatically swaps the Darshan 
module or disables it), but that's not very clean either.

Does anyone on the list have an idea for how to handle this more 
transparently?

-Phil

On 01/17/2017 02:38 PM, Rob Latham wrote:
>
>
> On 01/16/2017 02:26 PM, Phil Carns wrote:
>> Hi Kalle,
>>
>> Sorry for the delay, just now catching up on this and doing some testing
>> so we can apply the patch.
>>
>> What kind of system did you trigger the problem on?  It seems like int
>> and int64_t should be interchangeable (lucky for us) on most systems, so
>> I'm trying to figure out what the regression test would be.
>
> As someone who not that long ago promoted  out all the 'ints' in MPICH 
> to "int64_t' (or its equivalent), the two are *not* interchangeable.
>
>  int is 32 bits on every common platform, and int64_t is, well, 64 bits.
>
> Longs and Pointers are 64 bits on x86_64
> ==rob
>
>>
>> thanks,
>> -Phil
>>
>> On 11/15/2016 10:41 AM, Phil Carns wrote:
>>> Thank you very much for the patch, Kalle!
>>>
>>> The team is pretty busy this week with SC16 activities, but we will
>>> have a look at your patch soon.  Thanks for catching that.
>>>
>>> I agree that it would be great if we could have one Darshan build that
>>> works with both versions of the HDF API, but I don't see a way to do
>>> it either...
>>>
>>> -Phil
>>>
>>> On 11/13/2016 11:21 AM, Karl-Ulrich Bamberg wrote:
>>>> Dear Darshan-developers,
>>>>
>>>> Thanks for this great tool, it helps us a lot.
>>>>
>>>> We switched to HDF5 1.10 and found that DARSHAN (Git/master) is not
>>>> compatible due to the typedef of hid_t which changed in HDF5 1.10.
>>>> There is a patch attached, that provides a configure-switch for HDF5
>>>> 1.10.
>>>> Unfortunately 1.10 breaks compatibility tim 1.8 so I do not (yet)
>>>> see, if there is any way to use both with the same DARSHAN.
>>>>
>>>> Kind regards and keep up with the good work :-)
>>>>
>>>> Kalle
>>>>
>>>>
>>>> Mit freundlichen Grüßen
>>>> -- 
>>>> Karl-Ulrich Bamberg
>>>>
>>>> Ludwig-Maximilians-Universität München
>>>> Arnold-Sommerfeld-Center (ASC)
>>>> Computational & Plasma Physics
>>>> Theresienstr. 37, D-80333 München
>>>>
>>>> phone: +49 (0)89 2180 4577
>>>> fax: +49 (0)89 2180 99 4577
>>>> e-mail: Karl-Ulrich.Bamberg at physik.uni-muenchen.de
>>>> <mailto:Karl-Ulrich.Bamberg at physik.uni-muenchen.de>
>>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> Darshan-users mailing list
>>> Darshan-users at lists.mcs.anl.gov
>>> https://lists.mcs.anl.gov/mailman/listinfo/darshan-users
>>
>>
>>
>>
>> _______________________________________________
>> Darshan-users mailing list
>> Darshan-users at lists.mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/darshan-users
>>



More information about the Darshan-users mailing list