Enhanced collective chunk IO supports
Description:
1. When using collective IO with chunking storage without any tuning, performance may become worse under some circumstances.
2. Current HDF5 handles raw-data IO per chunk. So for many small chunks, many small IOs will be passed into MPI-IO. That may cause bad performance.
3. For one IO per chunk case, sometimes performance with collective is worse than performance with independent. An obvious case is when only one process is doing IO and all other process are not doing IO, the collective IO will only add overheads for communication. We want to avoid this case. Some management inside our library needs to be done.
Solution:
- Added managements of collective IO supports for chunking storage
inside parallel HDF5
1) Implemented One IO with collective mode for all chunks in the
application by building one MPI derived datatype accross all
chunks.
2) Implemented the decision-making support to do collective IO inside
MPI-IO per chunk.
3) Added the decision-making support to do one IO accross all chunks
or to do multiple IOs with each IO per chunk.
4) Added the support to handle the case some processes won't do any IOs in
collectively.
5) Some MPI-IO package(mpich 1.2.6 or lower, e.g.) cannot handle
collective IO correctly for the case when some processes have no
contributions to IOs, a special macro is added to change
collective IO mode to independent IO mode inside HDF5 library.
Platforms tested:
Parallel:
IBM AIX 5.2(copper)
Linux (heping) mpich-1.2.6
SDSC Teragrid mpich-1.2.5
Linux(Tungsten) mpich-1.2.6
Altix(NCSA cobalt)
Seq:
Linux(heping)
Misc. update:
Bug fix & new feature
Description:
Support variable-length datatypes in compact data storage and chunked
data storage, along with attributes.
Bug fix on the H5T_vlen_set_loc to allow for changing the file on a
variable-length datatype on disk.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Linux 2.4
Can't h5committest right now, due to missing cache files.
1) Add parallel test for metadata cache
2) Split serial test for metadata cache into two parts
3) Fix bug in which cache was flushed needlessly when the
cache wasn't full.
4) Performance improvements
5) Update API for parallel cache coherency bug fix.
Description:
See above.
Solution:
See above.
Platforms tested:
h5committest
Misc. update:
Code cleanup
Description:
Check in some of the code cleanups from working on the external link
support. (This doesn't include any of the external link features)
Platforms tested:
FreeBSD 4.11 (sleipnir)
Mac OSX.4 (amazon)
Linux 2.4
Cleaned up reconfigure output
Description:
Added two lines to separate output from trace script from autotools
output in reconfigure.
Solution:
Running reconfigure updated src/Makefile.in as a side effect, since
I changed config/lt_vers.am in my last checkin but forgot to
run reconfigure then.
Platforms tested:
mir (cosmetic change to reconfigure script only)
Code cleanup/update
Description:
Update compiler flags for gcc 4.1.x & start cleaning up some of the C++
compatibility
Platforms tested:
FreeBSD 4.11 (sleipnir) w/gcc 4.1 and/or g++
Feature
Description:
Added character encoding and attribute creation property lists.
Solution:
Attributes' character encoding is set via the ACPL.
The default is ASCII, with UTF-8 being the other option currently.
Platforms tested:
heping, mir, sleipnir, copper
Bug fix
Description:
Retrieving an object's name could fail (in various ways) under certain
circumstances (mostly having to do with mounted files).
Solution:
Re-write & simplify "get object name" code to fix error in a better way
than adding yet another hack to the previous pile of hacks... :-)
Platforms tested:
FreeBSD 4.11 (sleipnir)
h5committest
Patch for red-storm.
Description:
A temporary patch for redstorm, must open a file in the same parallel
data filesystem before H5_init_library. Otherwise, it hangs with error:
(client.c:568:ptlrpc_check_status()) @@@ type == PTL_RPC_MSG_ERR, err == -114 req@0x57a9660 x1/t0 o38->mds_l2@MDS_PEER_UUID:12 lens 240/272 ref 1 fl Rpc:R/100000/0 rc 0/-114
Platforms tested:
Red-storm.
Code cleanup
Description:
Fix a bunch of warnings flagged by Windows compilers.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Too minor to require h5committest
Unstable format version marker.
Description:
This marks the Check in baseline for compact group revisions,
which radically revises the source code for managing groups and object headers.
Earlier versions (1.7.52 or earlier) have the original format. This version
and later have the unstable format until further notice.
Platforms tested:
h5committested.
New feature
Description:
Add in a combination of Peter's & my code to support copying
variable-length data from one file to another, although currently only
supported with contiguous data storage.
Platforms tested:
FreeBSD 4.11 (sleipnir)
h5committest
Description: H5Eset_auto_stack() and H5Eset_auto() used FUNC_ENTER_API. It would clear the error
stack from the previous API call. It's better to use FUNC_ENTER_API_NOCLEAR to avoid clearing
the stack.
Platforms tested: fuss - simple change.
Bug fix
Description:
Fortran type generation was broken in two ways. Fixed both.
Solution:
Firstly, there were a couple of path problems. Fixed a typo and
specified the full path of a file.
Secondly, the dependencies weren't right when building with HDF5-specific
commands (make lib, make check-s, etc.). Tweaked dependencies
to fix the problem.
Platforms tested:
mir, modi4, sleipnir
slight code change for efficiency of iteration of chunk address
Description:
Quincey picked this up because of my inexperience of C. A pair of redundant malloc and free calls will be removed.
Solution:
Platforms tested:
heping(Linux 2.4)
too little to be tested.
Misc. update:
Configure feature
Description:
Added 'make trace' target.
Solution:
Added tracing to 1.7. This was done automatically in 1.6, but left out
of 1.7 until now (oops!).
Tracing in 1.7 only happens manually, when the user types 'make trace.'
Tracing automatically requires more framework than it's worth.
I also fixed a couple of tracing bugs and ran trace.
Platforms tested:
mir, sleipnir, modi4
Misc. update:
Bug fix
Description:
Before this checkin, 'gmake check-s' would fail if there was a file in
the current directory named 'check-s'.
This is fixed under gmake (not sure how to fix for other makes).
Solution:
check, progs, install, etc. are what gmake calls "phony" targets,
which means that no file should be created. These targets can be
specified by a line of the form
.PHONY: check progs install ...
Automake adds this line for targets it knows about, but HDF5 has a
lot of custom rules. This checkin adds a .PHONY line for those rules.
I believe that only gmake recognizes the .PHONY line (at least, pmake
doesn't seem to), but a partial solution is better than none.
This error should occur very rarely anyway (the user has to manually
create files with names like 'build-check-s' or '_test').
Platforms tested:
mir, sleipnir, modi4
Cleanup.
Description:
Removed the home-made version of snprintf and vsnprintf since they
do not actually enforce the size parameter correctly and could corrupt
memory or as a security hole.
This block of code has been turned off since Aug 11, 2005 for all platforms
except Tflops. No one has reported any problem since--sort of empirical
evidence that all platforms except Tflops support both functions.
Tflops has retired in September. So, no need to keep this block of code.
Also, fixed the HDxxx macro to call the corresponding xxx function.
Platforms tested:
Tested in heping, both serial and parallel.
Code cleanup
Description:
Added some comments about the thread-specific memory for the error and
function stacks.
Changed H5close() to not use the function stack (since it causes it to
be re-enabled after shutting it down)
Changed thread-safe semaphore code to not use dynamicly allocated memory.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Linux 2.4
Too minor to require h5committest
More tests
Description:
Add more tests for proper behavior of groups with different group
creation property settings.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Linux 2.4
Solaris 2.8
To support collective chunk IO, obtain the chunk addresses globally from one process.
Description:
Implement a call-back function to iterate B-tree to do the job.
Check in this code for major check-ins later.
Solution:
Platforms tested:
heping(Linux)
Should not be used by any library functions now, so won't affect library and tool tests.
No need to test on 3 platforms.
Misc. update:
Bug fix
Description:
Fixed bug in hl/src Makefile that was deleting y.tab.c.
Running bin/reconfigure changed a few other files as well (somebody
probably updated Makefiles.am but forgot to bin/reconfigure).
Solution:
Automake in its wisdom adds the line "rm *.tab.c" to every distclean
in every directory. I assume that it assumes that such files are
generated during the build. Supplied an empty distclean-compile
target to remove this line from the Makefiles.in.
Platforms tested:
mir, sleipnir, shanti
New feature
Description:
Check in baseline for compact group revisions, which radically revises the
source code for managing groups and object headers.
WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!!
WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!!
This initiates the "unstable" phase of the 1.7.x branch, leading up
to the 1.8.0 release. Please test this code, but do _NOT_ keep files created
with it - the format will change again before the release and you will not
be able to read your old files!!!
WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!!
WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!! WARNING!!!!
Solution:
There's too many changes to really describe them all, but some of them
include:
- Stop abusing the H5G_entry_t structure and split it into two separate
structures for non-symbol table node use within the library: H5O_loc_t
for object locations in a file and H5G_name_t to store the path to
an opened object. H5G_entry_t is now only used for storing symbol
table entries on disk.
- Retire H5G_namei() in favor of a more general mechanism for traversing
group paths and issuing callbacks on objects located. This gets us out
of the business of hacking H5G_namei() for new features, generally.
- Revised H5O* routines to take a H5O_loc_t instead of H5G_entry_t
- Lots more...
Platforms tested:
h5committested and maybe another dozen configurations.... :-)
New feature
Description:
Add in baseline "object copy" code from Peter [in the form of a new API
routine: H5Gcopy()]. There's still some work to do (like handling variable-
length datatypes and possibly support for references) and it hasn't been tested
on mounted files yet, but the core functionality is there and working
correctly.
I've also got a set of patches to update the 1.6 branch with tweaks to
keep the branches mostly in sync, but Elena will kill me if I import them
before the 1.6.5 release is out... :-)
Platforms tested:
FreeBSD 4.11 (sleipnir)
h5committested
Code cleanup
Description:
Clean up H5Dcreate a little, to prepare for coding standards discussion
today.
Platforms tested:
Just eyeballed, too trivial to require testing.
Bug fix/feature
Description:
Added support for -shlib in h5fc and h5c++.
Made check-install use -shlib when only shared libraries have been installed.
Solution:
h5fc and h5c++ didn't recognize -shlib. Stole code from h5cc to link against
shared libraries.
When static libraries are disabled, the examples Makefiles will automatically
use the -shlib option to link against shared libraries. Thus,
--disable-static and make check-install should work together.
Platforms tested:
heping(disable-static, enable-static, fortran, c++), modi4 (disable-static, fortran, c++, parallel, enable-static)
Bug fix
Description:
Failed parallel tests now cause make to exit with an error.
Solution:
Edited config/conclude.am to throw an error if parallel test programs fail.
Platforms tested:
heping, modi4
Code cleanup
Description:
Take out odd definition of HDalarm() and replace with standard macro
like the rest of the library.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Too minor to require h5committest
Code cleanup
Description:
Clean up & standardize a bit in preparation for coding standards
discussion.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Too minor to require h5committest
Bug fix
Description:
Changed configure.in to use an environment variable TR to set the path
to the tr utility.
Solution:
There are two kind of tr on Solaris with slightly different syntax.
HDF5's configure relies on the "standard" tr. Traditionally, HDF5ers
have needed to make sure that the "right" tr was found before the
wrong one in their path; now they can use an environment variable.
Platforms tested:
mir, shanti, sol
Misc. update:
Forgot to update release notes. Off to do that now.
Makefile bug fix
Description:
Previously, automake didn't output rules to build perform/mpi-perf or
the test/gen_* programs.
Now these can be built by typing 'make mpi-perf' (or 'make foo') or by
configuring with --enable-build-all.
Solution:
Automake doesn't like having rules for programs it doesn't build. Tricked
it by having these programs built "sometimes"--whenever the user enables
--build-all. This should be used mostly for testing and to ensure that
these helper programs compile.
***IMPORTANT***
These programs do *not* currently compile. When --enable-build-all is used
(not the default), gen_new_fill fails because it uses an old API. This is
an existing "bug" that has simply been exposed by this checkin.
Platforms tested:
sleipnir, modi4, sol
Misc. update:
Bug fix
Description:
Check in Kent's fixes to hyperslab rebuilding code that should correctly
detect situation when different sub-hyperslabs exist in a higher dimension.
Solution:
Remember previous sub-hyperslab information and compare that with newly
generated sub-hyperslab to make certain they are the same.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Too minor to require h5committest
Bug fix
Description:
Mark continuation message & chunk as dirty for case where chunk is
extended.
Also refactored code to reduce duplication.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Linux 2.4
Too minor to require h5committest
Bug fix
Description:
Some platforms complain that tcp.h is present but cannot be compiled.
Solution:
This is because tcp.h needs in.h to be included.
Checked for netinet/in.h and included it when netinet/tcp.h is being
checked.
Platforms tested:
mir, cu11
Fixed copyright notice.
Description:
A line is missing in the copyright notice. Fixed.
Platforms tested:
No test since it is only a fix in a comment block.
Bug fix
Description:
Fix "double root open" bug and apply same fix to datasets & named datatypes.
Also fix bug in symbol table entry lookup code that would "transfer" an
object from one file handle to another file handle when the two file handles
were opened on the same actual file.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Too minor to require h5committest
Code cleanup
Description:
Clean up & optimize (a little) the recent "hyperslab rebuilding" code
w/Kent.
Platforms tested:
FreeBSD 4.11 (sleipnir)
Too minor to require h5committest
bug fix.
Description:
When a parallel test script test fails, make would continue because the
way it was setup inside a for loop. Fixed it by issuing an exit 1 inside
the loop.
There was also a typo error in the newer command comparision that it
must be $${chkname} in order to be valid. Also, the test script itself
was not checked in the newer lists. All fixed.
Platforms tested:
h5committested and also hand tested in heping pp mode.
