In my experience, the HDF5 C++ API is not very convincing. It was implemented based on HDF5 1.6.1 and not much progress had been made to keep track with more recent feature. For instance, it implements HDF5 C++ objects via reference counting, which is already done by HDF5 itself, but was not exposed in the API until 1.6.2. So HDF5 C++ does double work here. Also, it makes extensive usage of C++ exceptions, which C++ users complained about that exceptions would not be designed for such usage. In the end, most C++ users seem to use the C API and have their own C++ wrappers around HDF5. I haven't observed any significant work on HDF5 C++ in the last years, so don't have the impression there will happen much for 1.10 .
Werner
On Wed, 29 Sep 2010 12:42:51 -0500, Erik Schnetter schnetter@cct.lsu.edu wrote:
We're now building HDF5 ourselves with Cactus, since this ensures that we have up-to-date versions (with e.g. all Lustre performance improvements) that work with the particular compilers and MPI versions we use.
Unfortunately, we just decided to support the C++ interface for HDF5, and this is incompatible with the parallel HDF5 interface. I wonder why -- maybe that'll change in HDF5 1.10.
-erik
On Wed, Sep 29, 2010 at 10:11 AM, Werner Benger werner@cct.lsu.edu wrote:
I have not seen a concrete release date, I doubt it would already be part of 1.10. As of now they're looking for feedback on the envisioned API before starting to implement it, so very early state yet.
However, as far as I understand it, it will be based on parallel HDF5. Cactus doesn't use parallel HDF5, but serial HDF5. It might have been a decade ago since Thomas Radke tried parallel HDF5 with Cactus, so that would want to be re-animated, requiring MPI-IO to be installed on the respective machines as prerequisite. That apparently was an issue years ago, maybe it's no longer one.
WernerOn Wed, 29 Sep 2010 09:19:46 -0500, Erik Schnetter schnetter@cct.lsu.edu wrote:
I think this is only a proposal at the moment. I think HDF 1.10 is currently under development -- maybe it'll make it in there, maybe it'll be 1.12. I'm not following HDF5 development, but Werner Benger does, maybe he knows.
-erik
On Wed, Sep 29, 2010 at 2:28 AM, Michael Koppitz koppitz@aei.mpg.de wrote:
This is intriguing indeed! When will this new API be publicly available?
Michael
On Sep 29, 2010, at 3:24 AM, Erik Schnetter wrote:
The HDF5 groups is apparently contemplating adding a new high-level API for "HPC HDF5", based on parallel HDF5. I find the idea intriguing, as it would move much code that we have currently written in Carpet (and that we will have to write to address the "slow recovery" issue) into HDF5. In particular, recovering would consist essentially of a single call to H5HPCdataset_read_mult_all(), which would read multiple datasets from multiple processes.
-erik
-- Erik Schnetter schnetter@cct.lsu.edu http://www.cct.lsu.edu/~eschnett/
<H5HPC_RFC-2010-09-28.pdf>_______________________________________________ Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
-- ___________________________________________________________________________ Dr. Werner Benger Visualization Research Laboratory for Creative Arts and Technology (LCAT) Center for Computation & Technology at Louisiana State University (CCT/LSU) 211 Johnston Hall, Baton Rouge, Louisiana 70803 Tel.: +1 225 578 4809 Fax.: +1 225 578-5362