Segmentation violation/abort when using h5::high_throughput with HDF5 1.12

[knoepfel]

Segmentation violation/abort when using h5::high_throughput with HDF5 1.12

[steven-varga]

The reason is HDF5 library mismatch. H5CPP supports upto 1.10.6, the upcoming release will support 1.12.x

The h5::high_throughput was designed to replace/supplement the HDF5 filter chain, which in some corner cases was measured to be sub optimal. This version of H5CPP is for libhdf5-1.10.6, in the 1.12 there are several internal changes which probably break things. Will mitigate these differences in the upcoming version.

Not passing the h5::high_throughput flag disables the code path, and uses the one provided by The HDFGroup.

If I may ask, what is the reason you are using 1.12.x HDF5 library?

[knoepfel]

Thanks, @steven-varga - I guess I should have figured that out. h5cpp and HDF5 are two of a large number of products the software I support requires. We try to stay on top of new versions of those packages as they become available, but we occasionally introduce library mismatches in the process.

I will bring up this library mismatch with my support team - we'll likely revert back to HDF5 1.10.

[lgarren]

@steven-varga You mention an upcoming release that will support hdf5 1.12.x. Is there an estimated release date?

[steven-varga]

Hi Lynn, @lgarren and Kyle @knoepfel would you interested in an online meet (and anyone else) perhaps next week Friday @5pm EDT, where I could could learn in what way you are using this library? It would accelerate it, if I know what features to focus on :)

[knoepfel]

Hi @steven-varga, yes a meeting next Friday (4/23 at 5p EDT) sounds good. @chissg, @sabasehrish, and @marcpaterno may also be interested in joining the meeting.

[steven-varga]

@knoepfel please check commit 8c3b5b0
if fixes #70 for lbihdf5 v1.12.0 with v112 api set. I use spack on my cluster with the following, linked against most recent but customised openMPI

spack install [email protected]~cxx~debug~fortran~hl~java~szip~threadsafe+mpi+shared+pic api=v112

The fix disables the custom pipeline code path for libhdf5 above v1.10.6; if this works, will re-enable it in the upcoming version; this requires no-change on your side (other then using this fix).

[knoepfel]

Thanks, @steven-varga. In our testing of the patch, we are encountering some build-time errors. I'm happy to fix those and create a PR. Is it acceptable to also address some other warnings that appear when we build our client code (e.g. sign-compare, unused variables)?