#### H5PART 1.99.10 ########################################################### Bugfixes in the malloc wrappers Converter from vtk to H5hut added. Bugfixes in the autotools files. #### H5PART 1.99.9 ############################################################ Bugfix in option parsing in tools/h5PartDcToVtk and tools/h5PartSurfaceToVtk tools/h5PartDcToVtk and tools/h5PartSurfaceToVtk removed from list of tools to be compiled. These tools should be moved to another repository anyway. #### H5PART 1.99.8 ############################################################ Bugfixes #### H5PART 1.99.7 ############################################################ Linking problem on Linux fixed #### H5PART 1.99.6 ############################################################ #### H5PART 1.99.5 ############################################################ #### H5PART 1.99.4 ############################################################ #### H5PART 1.99.3 ############################################################ #### H5PART 1.99.2 ############################################################ #### H5PART 1.99.1 ############################################################ #### H5PART 1.99 ############################################################ Renamed type: H5PartFile -> h5_file_t h5part_int64_t -> h5_int64_t Renamed functions: #### H5PART 1.6.6 ############################################################ Fixed typo (Write instead of Read) in the h5bl_3d_read_scalar_field* Fortran interface. Added missing type normalization for int32 type. Rewrite of h5pAttrib tool. #### H5PART 1.6.5 ############################################################ Fixed several build errors reported by Iuri Prilepov on Ubuntu 11.04. #### H5PART 1.6.4 ############################################################ Fixed bug where H5PartSetViewIndices was not setting an empty view when the number of elements is 0. Set a threshold on the HDF5 alignment parameter so that small metadata writes are not aligned, which causes large gaps in the file. #### H5PART 1.6.3 ############################################################ New build system uses libtool and can build shared libraries. Fixed a bug that incorrectly identifies the number of selected points in a view when using H5PartSetViewIndices. Fixed bug in Fortran test reported by several people, as well as several incorrect views set in the C test that were causing segfaults. Fixed name mismatches in the Fortran interface, and an off-by-one indexing problem. #### H5PART 1.6.2 ############################################################ Removed H5PartSetViewEmpty -------------------------- An empty view can now be selected with: H5PartSetNumParticles(file, 0); Bug Fixes to Attribute Calls in Fortran API ------------------------------------------- Fixed a problem where attribute values were reverting to zero. #### H5PART 1.6.1 ############################################################ Chunking in the H5Part API -------------------------- There is now an H5Part equivalent H5PartSetChunkSize to the existing H5BlockDefine3DChunkDims call. Both of these calls enable the chunking mechanism in the underlying HDF5 layer. Chunking is used in combination with the aligned open calls to pad datasets to alignment multiples on disk. All Steps Available on Write ---------------------------- Previously, existing steps in a file were only accessible in read-only mode. Now, all steps are available in all modes, including write-truncate and write-append. Thus, it is now possible to overwrite existing data in write mode. To help alert the user to this possibility, a warning is issued every time an existing dataset is written to. Warnings can be printed by setting the verbosity to H5PART_VERB_WARN or higher. New Throttling Routine ---------------------- Previously, a throttle factor of N meant that for P processors, the total number of P writes were divided into N batches of P/N writes. To better accommodate round-robin lustre striping, the new policy is to execute P/N batches of N writes. Thus, matching N to the number of stripes results in the desirable 1-1 matching of writers to stripes. Other Fixes ----------- * The configure script should correctly detect PGI compilers now. * There was a small memory leak in the Fortran file open calls. * There was a compile error with HDF5 1.6.x due to a missing assignment. #### H5PART 1.6 ############################################################## Updated Documentation --------------------- The Fortran API has been added to the doxygen documentation. Most entries have a reference to their respective C API call. Regression Test Suite --------------------- The test subdirectory has been reorganized to include a systematic series of regression tests of common use cases for the API. The goal is to have both complete coverage of the API (every call is exercised) and testing of some unlikely or unsuspected values and cases. Currently, the C test covers the entire H5Part API and the Fortran test is still in development. Benchmark Utilities ------------------- Two benchmarking utilities, H5PartBench and H5BlockBench, have been added to the tools collection. Both require the parallel library and have functionality and syntax similar to the IOR benchmark: http://sourceforge.net/projects/ior-sio/ File Mode Flags --------------- Previously, the only flags used were H5PART_READ, H5PART_WRITE, and H5PART_APPEND, and these were mutually exclusive. Additional flags have been added to select the underlying virtual file driver (VFD) that HDF5 uses in parallel mode and to accommodate the lustre filesystem: H5PART_VFD_MPIPOSIX (use the MPI-POSIX driver, which bypasses MPI-IO) H5PART_VFD_MPIIO_IND (use MPI-IO in independent mode) H5PART_FS_LUSTRE (activate H5P tunings for the lustre stripe size) These flags can be specified using the typical bitwise OR method, e.g. char flag = H5PART_WRITE | H5PART_VFD_MPIPOSIX | H5PART_FS_LUSTRE; will select write mode with these two additional options. The parallel library now defaults to using the MPI-IO collective mode driver (previously it used independent mode). The collective buffering algorithm in the Cray XT4/5 environment has been substantially improved in the last year, and defaulting to collective mode allows us to take advantage of these improvements automatically. There are also new open calls H5PartOpenAlign and H5PartOpenParallelAlign that take an additional 'alignment' value. This value is passed to HDF5 and used to pad out objects so that they align to filesystem boundaries. For instance, it would make sense to use the stripe size as the alignment value when writing to a lustre filesystem. Unified 'View' Model -------------------- Views are now supported in both read and write mode (previously only supported in read mode). It is possible to specify non-sensical views: you could for instance specify a read view that is larger than the dataset on disk. Or you could create a write view that is larger than the data in memory. In those cases, you will encounter an error when you try to perform a read or write operation on a dataset. There are two new methods for selecting views. H5PartSetNumParticlesStrided selects a view with $n$ particles per processor, but such that the data in memory is expected to have a 'stride' factor. For instance, if you have particle data with fields $x$ and $y$ and a single array with entries $x1,y1,x2,y2...$, then the stride factor is 2. Striding works for both reads and writes and only affects the view of memory: individual fields are still stored as individual arrays on disk. The second new method is H5PartSetViewIndices, which allows for point selections of datasets. You can pass a list of indices for the points you want to select. This is useful, for instance, when interfacing with FastBit to perform queries that select a small subset of non-contiguous particles. Because views are now supported on write, it is possible to write a dataset using multiple 'passes' or to leave some values unwritten. Also, all views are now *inclusive*, so that a view of (0,9) corresponds to the 10 items 0, 1, 2, ... 9. Internal Handling of Dataspaces ------------------------------- H5PartSetNumParticles and H5PartSetNumParticlesStrided now share the same HDF5 dataspace state in the H5PartFile struct with the calls that modify the view (H5PartSetView etc.). Previously, the memory and disk dataspaces were regenerated from the view on every read access, which was unnecessary. Autogeneration of API Calls --------------------------- Python scripts have been added to automatically generate read/write call variants for different data types. Renamed Fortran Include ----------------------- The automake system no longer greps the *F90.inc files to generate an include called 'H5Part.inc'. Instead, the *F90.inc files have been renamed to *.f90, and awk is used to generate an 'H5PartF.h' file (the same naming convention as in MPI's 'mpif.h'). The change to *.f90 was necessary to facilitate integration of the Fortran API into doxygen. Note that the *.f90 files will not compile, nor can they be included directly. Simply use include 'H5PartF.h' in your Fortran code to include all the definitions for Fortran H5Part/H5Block calls. 64 Char Limit on Dataset Names ------------------------------ Previously, a user could overrun internal buffers for dataset names. Now, a fixed limit of 64 chars is imposed. Dataset names that are longer than this are truncated and a warning is printed. We expect that most users are using short canonical names like x, px, id, etc. Changes to Existing API ----------------------- The H5BlockGetFieldInfo and H5BlockGetFieldInfoByName calls now both include a parameter for the 'type' of the field. #### H5PART 1.4 ############################################################## Error Handling -------------- Now all functions, in which an error could occure, are returning a value. This value is either a 64bit integer or a pointer. A negative 64bit integer as result indicates an error. Values >= 0 indicates successfull execution. This is similiar to the convention used in UNIX system calls, but we do not return -1 but the error number (which is always negative). For functions returning a pointer the NULL-pointer is used to indicate an error. You can call the function H5PartGetErrno() to get the error number. For the time being there is no strerror(3) or perror(3) equivalent. API Changes ----------- There are several changes in the API of H5Part. The biggest change is a side-effect of the implementation of strict error handling. The impact to the existing C/C++ code is very small. But now you *can* and *should* implement some kind of error handling. The simplest error handling is to set an error handler which aborts the program as soon as an error occured. The next "biggest" change is the use of H5Part types instead of "long long" and "double". The replacement of "long long" is "h5part_int64_t" and for "double" "h5part_float64_t". This change has now effect in Fortran. In some functions the argument type changed from "int" to "h5part_int64_t". Since on most systems "int" is a 32bit integer, calls to these functions must be adapted to the new API. In C/C++ the compiler will complain about it. But in Fortran you will *not* get an error message.So, you must check your Fortran code carefully. Changes in detail (not listed are functions where only the return type changed from any to "h5part_int64_t"): New API Changes to old API h5part_int64_t void H5PartSetNumParticles ( H5PartFile *f, h5part_int64_t nparticles long long ); h5part_int64_t int H5PartWriteDataFloat64 ( H5PartFile *f, char *name, h5part_float64_t *dta double ); h5part_int64_t int H5PartWriteDataInt64 ( H5PartFile *f, char *name, h5part_int64_t *dta long long ); h5part_int64_t void H5PartSetStep ( H5PartFile *f, h5part_int64_t step int ); h5part_int64_t void H5PartSetView ( H5PartFile *f, h5part_int64_t start, long long h5part_int64_t end long long ); h5part_int64_t int H5PartGetView ( H5PartFile *f, h5part_int64_t *start, long long h5part_int64_t *end long long ); h5part_int64_t int H5PartReadDataFloat64 ( H5PartFile *f, char *name, h5part_float64_t *dta double ); h5part_int64_t int H5PartReadDataInt64 ( H5PartFile *f, char *name, h5part_int64_t *dta long long ); h5part_int64_t void H5PartReadParticleStep ( H5PartFile *f, h5part_int64_t step, int h5part_float64_t *x, double h5part_float64_t *y, double h5part_float64_t *z, double h5part_float64_t *px, double h5part_float64_t *py, double h5part_float64_t *pz, double h5part_int64_t *id long long ); New functions: h5part_int64_t H5PartSetVerbosityLevel ( unsigned int ); h5part_int64_t H5PartSetErrorHandler ( h5part_error_handler handler ); h5part_int64_t H5PartGetErrno ( void ); h5part_int64_t H5PartDefaultErrorHandler ( const char *funcname, const h5part_int64_t eno, const char *fmt, ... ); h5part_int64_t H5PartAbortErrorHandler ( const char *funcname, const h5part_int64_t eno, const char *fmt, ... ); Removed functions: int H5PartFileIsValid ( H5PartFile *f );