diff options
Diffstat (limited to 'release_docs')
-rw-r--r-- | release_docs/RELEASE.txt | 308 |
1 files changed, 61 insertions, 247 deletions
diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index f8467be..37efa36 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -1,5 +1,8 @@ -HDF5 version 1.10.1-pre0 currently under development +HDF5 version 1.10.1-pre1 released on 2017-03-23 ================================================================================ +Note: this document is a work in progress. In particular not all bug fixes have +been documented here, and issues reported in the pre-release versions will appear in +"known problems" in the final 1.10.1 release. INTRODUCTION @@ -40,7 +43,8 @@ If you have any questions or comments, please send them to the HDF Help Desk: CONTENTS -- New Features +- Major New Features Introduced in HDF5 1.10.1 +- Other New Features - Support for new platforms and languages - Bug Fixes since HDF5-1.10.0-patch1 - Supported Platforms @@ -49,6 +53,59 @@ CONTENTS - Known Problems +Major New Features Introduced in HDF5 1.10.1 +============================================ + +For links to the RFCs and documentation in this section please view +https://support.hdfgroup.org/HDF5/docNewFeatures in a web browser. + +________________________________________ +Metadata Cache Image +________________________________________ + + HDF5 metadata is typically small, and scattered throughout the HDF5 file. + This can affect performance, particularly on large HPC systems. The Metadata + Cache Image feature can improve performance by writing the metadata cache in + a single block on file close, and then populating the cache with the contents + of this block on file open, thus avoiding the many small I/O operations that + would otherwise be required on file open and close. See the RFC for complete + details regarding this feature. Also, see the Fine Tuning the Metadata Cache + documentation. + +________________________________________ +Metadata Cache Evict on Close +________________________________________ + + The HDF5 library's metadata cache is fairly conservative about holding on to + HDF5 object metadata (object headers, chunk index structures, etc.), which can + cause the cache size to grow, resulting in memory pressure on an application + or system. The "evict on close" property will cause all metadata for an object + to be evicted from the cache as long as metadata is not referenced from any + other open object. See the Fine Tuning the Metadata Cache documentation + for information on the APIs. + +________________________________________ +Paged Aggregation +________________________________________ + + The current HDF5 file space allocation accumulates small pieces of metadata + and raw data in aggregator blocks which are not page aligned and vary widely + in sizes. The paged aggregation feature was implemented to provide efficient + paged access of these small pieces of metadata and raw data. See the RFC for + details. Also, see the File Space Management documentation. + +________________________________________ +Page Buffering +________________________________________ + + Small and random I/O accesses on parallel file systems result in poor + performance for applications. Page buffering in conjunction with paged + aggregation can improve performance by giving an application control of + minimizing HDF5 I/O requests to a specific granularity and alignment. + See the RFC for details. Also, see the Page Buffering documentation. + + + New Features ============ @@ -437,249 +494,6 @@ The following platforms are not supported but have been tested for this release. Known Problems ============== -* "make check" fails on CYGWIN when building shared lib files is enabled. The - default on Cygwin has been changed to disable shared. It can be enabled with - the --enable-shared configure option but is likely to fail "make check" - with GCC compilers. (LK -2015/04/16) - -* CLANG compiler with the options -fcatch-undefined-behavior and -ftrapv - catches some undefined behavior in the alignment algorithm of the macro DETECT_I - in H5detect.c (Issue 8147). Since the algorithm is trying to detect the alignment - of integers, ideally the flag -fcatch-undefined-behavior shouldn't to be used for - H5detect.c. In the future, we can separate flags for H5detect.c from the rest of - the library. (SLU - 2013/10/16) - -* The 5.9 C++ compiler on Sun failed to compile a C++ test ttypes.cpp. It - complains with this message: - "/home/hdf5/src/H5Vprivate.h", line 130: Error: __func__ is not defined. - - The reason is that __func__ is a predefined identifier in C99 standard. The - HDF5 C library uses it in H5private.h. The test ttypes.cpp includes - H5private.h (H5Tpkg.h<-H5Fprivate.h<-H5Vprivate.h<-H5private.h). Sun's 5.9 - C++ compiler doesn't support __func__, thus fails to compile the C++ test. - But 5.11 C++ compiler does. To check whether your Sun C++ compiler knows this - identifier, try to compile the following simple C++ program: - #include<stdio.h> - - int main(void) - { - printf("%s\n", __func__); - return 0; - } - (SLU - 2012/11/5) - -* The C++ and FORTRAN bindings are not currently working on FreeBSD with the - native release 8.2 compilers (4.2.1), but are working with gcc 4.6 from the - ports (and probably gcc releases after that). - (QAK - 2012/10/19) - -* The data conversion test dt_arith.c has failures (segmentation fault) from - "long double" to other datatypes during hard conversion when the library - is built with the default GCC 4.2.1 on Mac Lion system. It only happens - with optimization (-O3, -O2, and -O1). Some newer versions of GCC do not - have this problem. Users should disable optimization or try newer version - of GCC. (Issue 8017. SLU - 2012/6/12) - -* The data conversion test dt_arith.c fails in "long double" to integer - conversion on Ubuntu 11.10 (3.0.0.13 kernal) with GCC 4.6.1 if the library - is built with optimization -O3 or -O2. The older GCC (4.5) or newer kernal - (3.2.2 on Fedora) doesn't have the problem. Users should lower down the - optimization level (-O1 or -O0) by defining CFLAGS in the command line of - "configure" like: - - CFLAGS=-O1 ./configure - - It will overwrite the library's default optimization level. (Issue 7829. - SLU - 2012/2/7) - -* --with-mpe configure option does not work with Mpich2. AKC - 2011/03/10) - -* While working on the 1.8.6 release of HDF5, a bug was discovered that can - occur when reading from a dataset in parallel shortly after it has been - written to collectively. The issue was exposed by a new test in the parallel - HDF5 test suite, but had existed before that. We believe the problem lies with - certain MPI implementations and/or filesystems. - - We have provided a pure MPI test program, as well as a standalone HDF5 - program, that can be used to determine if this is an issue on your system. - They should be run across multiple nodes with a varying number of processes. - These programs can be found at: - http://www.hdfgroup.org/ftp/HDF5/examples/known_problems/ - -* Parallel mode in AIX will fail some of the testcheck_version.sh tests where - it treats "exit(134) the same as if process 0 had received an abort signal. - This is fixed and will be available in the next release. AKC - 2009/11/3 - -* The PathScale MPI implementation, accessing a Panasas file system, would - cause H5Fcreate() with H5F_ACC_EXCL to fail even when the file is not - existing. This is due to the MPI_File_open() call failing if the amode has - the MPI_MODE_EXCL bit set. (See bug 1468 for details.) AKC - 2009/8/11 - -* Parallel tests failed with 16 processes with data inconsistency at testphdf5 - / dataset_readAll. Parallel tests also failed with 32 and 64 processes with - collective abort of all ranks at t_posix_compliant / allwrite_allread_blocks - with MPI IO. (CMC - 2009/04/28) - -* For SNL, spirit/liberty/thunderbird: The serial tests pass but parallel - tests failed with MPI-IO file locking message. AKC - 2007/6/25. -* On Intel 64 Linux cluster (RH 4, Linux 2.6.9) with Intel 10.0 compilers use - -mp -O1 compilation flags to build the libraries. Higher level of optimization - causes failures in several HDF5 library tests. -* For HPUX 11.23 many tools tests failed for 64-bit version when linked to the - shared libraries (tested for 1.8.0-beta2) -* For SNL, Red Storm: only paralle HDF5 is supported. The serial tests pass - and the parallel tests also pass with lots of non-fatal error messages. -* on SUN 5.10 C++ test fails in the "Testing Shared Datatypes with Attributes" test -* configuring with --enable-debug=all produces compiler errors on most - platforms. Users who want to run HDF5 in debug mode should use - --enable-debug rather than --enable-debug=all to enable debugging - information on most modules. -* On Mac OS 10.4, test/dt_arith.c has some errors in conversion from long - double to (unsigned) long long and from (unsigned)long long to long double. -* On Altix SGI with Intel 9.0 testmeta.c would not compile with -O3 - optimization flag. -* On VAX, Scaleoffset filter isn't supported. The filter cannot be applied to - HDF5 data generated on VAX. Scaleoffset filter only supports IEEE standard - for floating-point data. -* On Cray X1, a lone colon on the command line of h5dump --xml (as in - the testh5dumpxml.sh script) is misinterpereted by the operating system - and causes an error. -* On mpich 1.2.5 and 1.2.6, we found that if more than two processes - contribute no IO and the application asks to do IO with collective, we found - that when using 4 processors, a simple collective write will be hung - sometimes. This can be verified with t_mpi test under testpar. -* The dataset created or rewritten with the v1.6.3 library or after can't - be read with the v1.6.2 library or before when Fletcher32 EDC(filter) is - enabled. There was a bug in the calculating code of the Fletcher32 - checksum in the library before v1.6.3. The checksum value wasn't consistent - between big-endian and little-endian systems. This bug was fixed in - Release 1.6.3. However, after fixing the bug, the checksum value is no - longer the same as before on little-endian system. The library release - after 1.6.4 can still read the dataset created or rewritten with the library - of v1.6.2 or before. SLU - 2005/6/30 -* For the version 6(6.02 and 6.04) of Portland Group compiler on AMD Opteron - processor, there's a bug in the compiler for optimization(-O2). The library - failed in several tests but all related to multi driver. The problem has - been reported to the vendor. -* On IBM AIX systems, parallel HDF5 mode will fail some tests with error - messages like "INFO: 0031-XXX ...". This is from the command poe. - Set the environment variable MP_INFOLEVEL to 0 to minimize the messages - and run the tests again. - The tests may fail with messages like "The socket name is already - in use". HDF5 does not use sockets (except for stream-VFD). This is - due to problems of the poe command trying to set up the debug socket. - Check if there are many old /tmp/s.pedb.* staying around. These are - sockets used by the poe command and left behind due to failed commands. - Ask your system administrator to clean them out. Lastly, request IBM - to provide a mean to run poe without the debug socket. - -* The C++ library's tests fails when compiling with PGI C++ compiler. The - workaround until the problem is correctly handled is to use the - flag "--instantiate=local" prior to the configure and build steps, as: - setenv CXX "pgCC --instantiate=local" for pgCC 5.02 and higher - - -* The stream-vfd test uses ip port 10007 for testing. If another - application is already using that port address, the test will hang - indefinitely and has to be terminated by the kill command. To try the - test again, change the port address in test/stream_test.c to one not - being used in the host. - -* The --enable-static-exec configure flag will only statically link libraries - if the static version of that library is present. If only the shared version - of a library exists (i.e., most system libraries on Solaris, AIX, and Mac, - for example, only have shared versions), the flag should still result in a - successful compilation, but note that the installed executables will not be - fully static. Thus, the only guarantee on these systems is that the - executable is statically linked with just the HDF5 library. - -* With the gcc 2.95.2 compiler, HDF 5 uses the `-ansi' flag during - compilation. The ANSI version of the compiler complains about not being - able to handle the `long long' datatype with the warning: - - warning: ANSI C does not support `long long' - - This warning is innocuous and can be safely ignored. - -* Certain platforms give false negatives when testing h5ls: - - Cray J90 and Cray T90IEEE give errors during testing when displaying - some floating-point values. These are benign differences due to - the different precision in the values displayed and h5ls appears to - be dumping floating-point numbers correctly. - -* Not all platforms behave correctly with szip's shared libraries. Szip is - disabled in these cases, and a message is relayed at configure time. Static - libraries should be working on all systems that support szip, and should be - used when shared libraries are unavailable. There is also a configure error - on Altix machines that incorrectly reports when a version of szip without - an encoder is being used. - -* On some platforms that use Intel and Absoft compilers to build HDF5 fortran library, - compilation may fail for fortranlib_test.f90, fflush1.f90 and fflush2.f90 - complaining about exit subroutine. Comment out the line - IF (total_error .ne. 0) CALL exit (total_error) - -* Information about building with PGI and Intel compilers is available in - INSTALL file sections 5.7 and 5.8 - -* On at least one system, (SDSC DataStar), the scheduler (in this case - LoadLeveler) sends job status updates to standard error when you run - any executable that was compiled with the parallel compilers. - - This causes problems when running "make check" on parallel builds, as - many of the tool tests function by saving the output from test runs, - and comparing it to an exemplar. - - The best solution is to reconfigure the target system so it no longer - inserts the extra text. However, this may not be practical. - - In such cases, one solution is to "setenv HDF5_Make_Ignore yes" prior to - the configure and build. This will cause "make check" to continue after - detecting errors in the tool tests. However, in the case of SDSC DataStar, - it also leaves you with some 150 "failed" tests to examine by hand. - - A second solution is to write a script to run serial tests and filter - out the text added by the scheduler. A sample script used on SDSC - DataStar is given below, but you will probably have to customize it - for your installation. - - Observe that the basic idea is to insert the script as the first item - on the command line which executes the the test. The script then - executes the test and filters out the offending text before passing - it on. - - #!/bin/csh - - set STDOUT_FILE=~/bin/serial_filter.stdout - set STDERR_FILE=~/bin/serial_filter.stderr - - rm -f $STDOUT_FILE $STDERR_FILE - - ($* > $STDOUT_FILE) >& $STDERR_FILE - - set RETURN_VALUE=$status - - cat $STDOUT_FILE - - tail +3 $STDERR_FILE - - exit $RETURN_VALUE - - You get the HDF make files and test scipts to execute your filter script - by setting the environment variable "RUNSERIAL" to the full path of the - script prior to running configure for parallel builds. Remember to - "unsetenv RUNSERIAL" before running configure for a serial build. - - Note that the RUNSERIAL environment variable exists so that we can - can prefix serial runs as necessary on the target system. On DataStar, - no prefix is necessary. However on an MPICH system, the prefix might - have to be set to something like "/usr/local/mpi/bin/mpirun -np 1" to - get the serial tests to run at all. - - In such cases, you will have to include the regular prefix in your - filter script. - -* H5Ocopy() does not copy reg_ref attributes correctly when shared-message - is turn on. The value of the reference in the destination attriubte is - wrong. This H5Ocopy problem will affect h5copy tool + This section will be populated with issues reported to help@hdfgroup.org + during pre-release testing. Please report any issues. |