[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: netcdf Transition

On 02/08/2015 16:25, Sebastiaan Couwenberg wrote:
> On 02-08-15 17:17, Nico Schlömer wrote:
>>> also switches back to
>>> the HDF5 serial variant because we cannot provide only the MPI variant.
>> Dang, this was the most important thing for me. I have been under the
>> assumption that using MPI provides all the capabilities that the serial
>> variant provides, plus parallel computation. Were build failures related to
>> the fact that we used to ship MPI versions?
> Yes, the ncview & gmt FTBFS were directly caused by only having the MPI
> variant available. Most others were only caused by the non-standard HDF5
> library paths that were all fixed with the patched nc-config & netcdf.pc.
> After the GCC 5 transitions are completed we need to revisit this topic
> and take Gilles up on his offer to help netcdf support co-installable
> MPI variants as HDF5 does. To process large dataset this is an important
> feature to support.
There is a patch to do this; I wrote it pre-jessie but it wasn't ready
in time.

Part of the problem to look out for is that, beyond co-installability,
AFAIK the MPI variant of HDF5 is not thread-safe, but the 'serial'
variant is. (There is a long-term plan to merge the two; the APIs are
identical). We need to ensure that use of Openmp, etc. elsewhere in an
application  does not break a non-threadsafe HDF5
> Kind Regards,
> Bas

Alastair McKinstry, <alastair@sceal.ie>, <mckinstry@debian.org>, https://diaspora.sceal.ie/u/amckinstry
Misentropy: doubting that the Universe is becoming more disordered. 

Reply to: