still need to add option in H5VL_cdf_dataset_get to return the variable type, but code/test is otherwise working nicely on my Mac. There may be issues on Theta/etc, but we are in a good place for a decent documentation draft.
now reading record variables correctly -- now only need to address cases with multiple record variables (stride between records). This may only require changes to the construction of the dimension-multiplier array (mm). The multiplier for the first dimension can just be set to the record stride distance.
woriking on record variable support with hyperslab selections -- may be able to adjust test to use correct process mapping, then just modify the dimension multiplyer used to calculate the position in source, etc..
adding record-variable reading for ALL selections in hdf5 -- still need to add capability to hyperslab selection. Should also add optimized pathway for simple hyperslab mappings (current approach is more general than it needs to be in many cases).
adding ability to open and read an attribute BY NAME ONLY.Still need to add functionality for user to inspect file (determine the attributes and variables in the file).
hyperslab selection is improved (and allows for MPI and POSIX-based IO).Still need to clean up and add functionality to read attributes and query metadata.
Replaced hyperslab selection with code from the ADIOS VOL - The code should be more flexible now, but the assignToMemSpace call seems to be wrong - need to check if the benchmark is actually the part that is wrong (is the memory hyperslab selection correct?).
Progress - Metadata is read through a cache that is read by rank 0 and then broadcasted to the other processes.The current version of the code should scale reasonably well (but the H5VL_CDF_CACHE_SIZE parameter will determine the performance).
CDF VOL will now parse typical file header correctly - Just need to clean up the implementation and tackle serial data access next (then parallel data access)
Quincey Koziol
Huihuo Zheng
Richard James Zamora