corresponding HDF5 operations.
re: github issue https://github.com/Unidata/netcdf-c/issues/908
also in reference to https://github.com/pydata/xarray/issues/2004
The netcdf-c library has implemented the nc_get_vars and nc_put_vars
operations as element at a time. This has resulted in very slow
operation.
This pr attempts to improve the situation for netcdf-4/hdf5 files
by using the slab operations provided by the hdf5 library. The new
implementation passes the get/put vars stride information down to
the hdf5 slab operations.
The result appears to improve performance significantly. Some simple
tests on large 2-D arrays shows speedups in excess of 150.
Misc. other changes:
1. fix bug in ncgen/semantics.c; using a list's allocated length
instead of actual length.
2. Added a temporary hook in the netcdf library plus a performance
test case (tst_varsperf.c) to estimate the speedup. After users
have had some experience with this, I will remove it, probably
after the 4.7 release.
The file docs/indexing.dox tries to provide design
information for the refactoring.
The primary change is to replace all walking of linked
lists with the use of the NCindex data structure.
Ncindex is a combination of a hash table (for name-based
lookup) and a vector (for walking the elements in the index).
Additionally, global vectors are added to NC_HDF5_FILE_INFO_T
to support direct mapping of an e.g. dimid to the NC_DIM_INFO_T
object. These global vectors exist for dimensions, types, and groups
because they have globally unique id numbers.
WARNING:
1. since libsrc4 and libsrchdf4 share code, there are also
changes in libsrchdf4.
2. Any outstanding pull requests that change libsrc4 or libhdf4
are likely to cause conflicts with this code.
3. The original reason for doing this was for performance improvements,
but as noted elsewhere, this may not be significant because
the meta-data read performance apparently is being dominated
by the hdf5 library because we do bulk meta-data reading rather
than lazy reading.
Fix#299
The conditions to make this error are the following:
* Two variables with different chunk sizes
* Both variables write on the same unlimited dimension
* The first variable has already written data when the second variable is created
This relies on the HDF5 capability to
dynamically load compression filters.
Note that a compression filter is just
a subcase of filters.
The primary user-visible changes are as follows:
1. Add a standard header "netcdf_filter.h" that defines
the necessary API extensions
2. Modify ncgen to support two new special attributes
"_Filter_ID" and "_Filter_Parameters" so that compression
can be turned on when creating a file using ncgen.
4. Add a detailed description of filtering support
to the user's guide; see the file filters.md
5. Add a test case directory for this: nc_test4/filter_test.
It is fragile and a ./configure flags (-enable-filter-test)
is defined (default disabled) to shut this off this test
to avoid spurious 'make check' failures.
Note that the HDF5 documentation is not up-to-date, so
much of what is encoded here comes from examining the
actual code in the file H5PL.c in the HDF5 source code.
Specific changes:
1. Add dap4 code: libdap4 and dap4_test.
Note that until the d4ts server problem is solved, dap4 is turned off.
2. Modify various files to support dap4 flags:
configure.ac, Makefile.am, CMakeLists.txt, etc.
3. Add nc_test/test_common.sh. This centralizes
the handling of the locations of various
things in the build tree: e.g. where is
ncgen.exe located. See nc_test/test_common.sh
for details.
4. Modify .sh files to use test_common.sh
5. Obsolete separate oc2 by moving it to be part of
netcdf-c. This means replacing code with netcdf-c
equivalents.
5. Add --with-testserver to configure.ac to allow
override of the servers to be used for --enable-dap-remote-tests.
6. There were multiple versions of nctypealignment code. Try to
centralize in libdispatch/doffset.c and include/ncoffsets.h
7. Add a unit test for the ncuri code because of its complexity.
8. Move the findserver code out of libdispatch and into
a separate, self contained program in ncdap_test and dap4_test.
9. Move the dispatch header files (nc{3,4}dispatch.h) to
.../include because they are now shared by modules.
10. Revamp the handling of TOPSRCDIR and TOPBUILDDIR for shell scripts.
11. Make use of MREMAP if available
12. Misc. minor changes e.g.
- #include <config.h> -> #include "config.h"
- Add some no-install headers to /include
- extern -> EXTERNL and vice versa as needed
- misc header cleanup
- clean up checking for misc. unix vs microsoft functions
13. Change copyright decls in some files to point to LICENSE file.
14. Add notes to RELEASENOTES.md
The nvars, ndims, and natts fields on the NC_HDF5_FILE_INFO struct are
never set. The nvars field is read, but since it is never written,
the value is always zero.
HDF5 does not permit a variable to have no_fill == TRUE if the variable type is variable length. This includes types NC_STRING and NC_VLEN.
The test above also excludes user-defined types which I'm not sure is needed or not.
In nc4 mode, the variables were ignoring the default "database" fill/no_fill mode set via a call to nc_set_fill(). This sets the no_fill mode on the variable to match the default database setting at the time the variable is defined.
This consists of a persistent attribute named
_NCProperties plus two computed attributes
_IsNetcdf4 and _SuperblockVersion.
See the 'Provenance Attributes' section
of docs/attribute_conventions.md for details.
The var struct has a 'dim' field which was not being used
Instead, the dimids field would always search for the dim
with the matching dimid. For db with large numbers of dims,
this could be a significant time sync.
Modified code to always set var-dim[i] when var->dimids[i] was
set (if the dim existed at that point). Then use the var->dim
field instead of var->dimids and search whenever requested.
All var->dim accesses are protected by asserts that verify
non-null and that the var->dim[]->dimid == var->dimids[].
In non-classic netcdf-4 models, it is allowable to have
large numbers of dims and vars. In many operations, the
entire list of dims or vars is searched for a dim/var matching
a specific name which results in *lots* of strncmp or strcmp
calls.
If we add a hash field to the var and dim structs similar to what
has already been done for the netcdf-3 formats, then we can hash the
name being searched for and numerically compare that value with
the var/dim hash value. If they match, then do a more expensive
strncmp call to ensure that the names truly match.
1. Added check to libsrc4/nc4var.nc_def_var_extra to
check that the no specified chunks size is greater than
the dimension size.
2. Added test to nc_test4/tst_chunks.c
re: e support ZCL-340681 and CPW-270700
HDF4 supports compression (and chunking)
but the chunking was not being recorded
for HDF4 files. So, I modified the necessary
files to support HDF4 chunking.
