Code cleanup
Description:
Replaced the comm/info freeing with H5FD_mpio_comm_info_free to make
it cleaner and consistent.
Replaced couple FUNC_LEAVE_NOAPI(SUCCEED) with FUNC_LEAVE_NOAPI(ret_value).
These were bugs since it was possible errors may occur at FUNC_ENTER...()
time and this hardcoding would not report those errors.
Documented some modification history.
Platforms tested:
Did not h5committested because these are all MPI code and well protected
by HAVE_PARALLEL.
Tested in eirene and modi4, both parallel mode.
Misc. update:
Silly bug fix.
Description:
Forgot to put the comm/info duplication code in.
Solution:
Put them back in.
Platforms tested:
Did not run h5committest since the changes are all in the MPIO driver code.
Tested on eirene and Modi4, both parallel modes.
Misc. update:
Bug fixes/API changes
Description:
Previously, the Communicator and Info object arguments supplied
to H5Pset_fapl_mpio() are stored in the property with its handle
values. This meant changes to the communicator or the Info object
after calling H5Pset_fapl_mpio would affect the how the property
list function. This was also the case when H5Fopen/create operated.
They just stored the handle value. This is not according to the
MPI-2 defined behavior of how Info objects should be handled. (MPI-2
defines Info objects must be parsed when called.)
The old design was trying to avoid numerous duplicates of the same
information (e.g., every property object holds one version, every
file opened holds another version, when all of them are referring to
the same original version.) Nevertheless it is safer to implement
it according to MPI-2 definition.
Futhermore, the library often needs to do message passing using the
supplied communicator. Using the same communicator as the application
version may result in some messages mix up.
Solution:
H5Pset_fapl_mpio now stores a duplicate of each of the communicator
and Info object.
H5Pget_fapl_mpio returns a duplicate of its stored communicator and
Info object. It is now the responsibility of the applications to free
those objects when done.
H5Fopen/create also stores a duplicate of the communicator and Info
object supplied by the File Access Property list.
H5Fclose frees those duplicates.
There are a few more internal VFL call back functions that they
follow this "make duplicates" requirement.
Platforms tested:
"h5committested".
What other platforms/configurations were tested?
Eirene (mpicc), sol(mpicc), copper(parallel)
Misc. update:
New Feature
Description:
Added support for the lock and unlock function calls for file
drivers. Only FPHDF5 uses this feature. All of these drivers don't
define lock or unlock methods.
Platforms tested:
Linux, Modi4, Sol
Misc. update:
Potential bug fix.
Description:
The code to rotate the metadata writing process may not be compiled
correctly in all cases.
Solution:
Change from using "++mpi_round" to "mpi_round+1"
Platforms tested:
Eyeballed...
Bug Fix
Description:
Metadata cache in parallel I/O can cause hangs in applications which
perform independent I/O on chunked datasets, because the metadata cache
can attempt to flush out dirty metadata from only a single process, instead
of collectively from all processes.
Solution:
Pass a dataset transfer property list down from every API function which
could possibly trigger metadata I/O.
Then, split the metadata cache into two sets of entries to allow dirty
metadata to be set aside when a hash table collision occurs during
independent I/O.
Platforms tested:
Tested h5committest {arabica (fortran), eirene (fortran, C++)
modi4 (parallel, fortran)}
FreeBSD 4.7 (sleipnir) serial & parallel
Misc. update:
Updated release_docs/RELEASE
Updated
Description:
Updated Copyright notice.
Replaced HGOTO_ERROR calls involving MPI calls with HMPI_GOTO_ERROR.
Platforms tested:
Tested h5committest {arabica (fortran), eirene (fortran, C++)
modi4 (parallel, fortran)}? Tested on Eirene (PP) only since
the code is in the MPIO module only.
Code cleanup/new feature.
Description:
Split FUNC_LEAVE into API and non-API specific versions. This allows a
solution to compiling this branch with C++, as well as reducing the size
of the binaries produced.
Platforms tested:
FreeBSD 4.7 (sleipnir) w/serial, parallel (including MPE) & thread-safe
Lots of performance improvements & a couple new internal API interfaces.
Description:
Performance Improvements:
- Cached file offset & length sizes in shared file struct, to avoid
constantly looking them up in the FCPL.
- Generic property improvements:
- Added "revision" number to generic property classes to speed
up comparisons.
- Changed method of storing properties from using a hash-table
to the TBBT routines in the library.
- Share the propery names between classes and the lists derived
from them.
- Removed redundant 'def_value' buffer from each property.
- Switching code to use a "copy on write" strategy for
properties in each list, where the properties in each list
are shared with the properties in the class, until a
property's value is changed in a list.
- Fixed error in layout code which was allocating too many buffers.
- Redefined public macros of the form (H5open()/H5check, <variable>)
internally to only be (<variable>), avoiding innumerable useless
calls to H5open() and H5check_version().
- Reuse already zeroed buffers in H5F_contig_fill instead of
constantly re-zeroing them.
- Don't write fill values if writing entire dataset.
- Use gettimeofday() system call instead of time() system when
checking the modification time of a dataset.
- Added reference counted string API and use it for tracking the
names of objects opening in a file (for the ID->name code).
- Removed redundant H5P_get() calls in B-tree routines.
- Redefine H5T datatype macros internally to the library, to avoid
calling H5check redundantly.
- Keep dataspace information for dataset locally instead of reading
from disk each time. Added new module to track open objects
in a file, to allow this (which will be useful eventually for
some FPH5 metadata caching issues).
- Remove H5AC_find macro which was inlining metadata cache lookups,
and call function instead.
- Remove redundant memset() calls from H5G_namei() routine.
- Remove redundant checking of object type when locating objects
in metadata cache and rely on the address only.
- Create default dataset object to use when default dataset creation
property list is used to create datasets, bypassing querying
for all the property list values.
- Use default I/O vector size when performing raw data with the
default dataset transfer property list, instead of querying for
I/O vector size.
- Remove H5P_DEFAULT internally to the library, replacing it with
more specific default property list based on the type of
property list needed.
- Remove redundant memset() calls in object header message (H5O*)
routines.
- Remove redunant memset() calls in data I/O routines.
- Split free-list allocation routines into malloc() and calloc()-
like routines, instead of one combined routine.
- Remove lots of indirection in H5O*() routines.
- Simplify metadata cache entry comparison routine (used when
flushing entire cache out).
- Only enable metadata cache statistics when H5AC_DEBUG is turned
on, instead of always tracking them.
- Simplify address comparison macro (H5F_addr_eq).
- Remove redundant metadata cache entry protections during dataset
creation by protecting the object header once and making all
the modifications necessary for the dataset creation before
unprotecting it.
- Reduce # of "number of element in extent" computations performed
by computing and storing the value during dataspace creation.
- Simplify checking for group location's file information, when file
has not been involving in file-mounting operations.
- Use binary encoding for modification time, instead of ASCII.
- Hoist H5HL_peek calls (to get information in a local heap)
out of loops in many group routine.
- Use static variable for iterators of selections, instead of
dynamically allocation them each time.
- Lookup & insert new entries in one step, avoiding traversing
group's B-tree twice.
- Fixed memory leak in H5Gget_objname_idx() routine (tangential to
performance improvements, but fixed along the way).
- Use free-list for reference counted strings.
- Don't bother copying object names into cached group entries,
since they are re-created when an object is opened.
The benchmark I used to measure these results created several thousand
small (2K) datasets in a file and wrote out the data for them. This is
Elena's "regular.c" benchmark.
These changes resulted in approximately ~4.3x speedup of the
development branch when compared to the previous code in the
development branch and ~1.4x speedup compared to the release
branch.
Additionally, these changes reduce the total memory used (code and
data) by the development branch by ~800KB, bringing the development
branch back into the same ballpark as the release branch.
I'll send out a more detailed description of the benchmark results
as a followup note.
New internal API routines:
Added "reference counted strings" API for tracking strings that get
used by multiple owners without duplicating the strings.
Added "ternary search tree" API for text->object mappings.
Platforms tested:
Tested h5committest {arabica (fortran), eirene (fortran, C++)
modi4 (parallel, fortran)}
Other platforms/configurations tested?
FreeBSD 4.7 (sleipnir) serial & parallel
Solaris 2.6 (baldric) serial
Purpose:
New API functions
Description:
Added API functions to return pointer to low-level file handle
(H5Fget_vfd_handle and H5FDget_vfd_handle) and related property list
setting functions(H5Pset_family_offset and H5Pset_multi_type).
Platforms tested:
Linux 2.2(eirene), Solaris 2.7(arabica), IRIX64 6.5(modi4)
Code cleanup
Description:
Combined H5P_isa_class and H5I_object functionality into a new internal
H5P API function: H5P_object_verify, which checks that a property list is
the appropriate class and then returns the property list object associated
with the property list ID.
This reduces the source code by about 200 LOC and trims the library binary
some more.
Platforms tested:
FreeBSD 4.6 (sleipnir)
Code cleanup
Description:
Changed the last HRETURN* statements in the FUNC_ENTER macros into HGOTO*
macros, which reduces the size of the library binary in certain
configurations by another 10%
Platforms tested:
FreeBSD 4.6 (sleipnir) serial & parallel, IRIX64 6.5 (modi4) serial &
parallel
Code cleanup
Description:
Change most (all?) HRETURN_ERROR macros to HGOTO_ERROR macros, along with
HRETURN macros to HGOTO_DONE macros. This unifies the error return path
from functions and reduces the size of the library by up to 10% on some
platforms.
Additionally, I improved a lot of the error cleanup code in many routines.
Platforms tested:
FreeBSD 4.6 (sleipnir) serial & parallel and IRIX64 6.5 (modi4) serial &
parallel.
Bug Fix
Description:
It was possible to create corrupted metadata information (either in memory
or in the file or both) with a parallel I/O program because of the way
metadata writes were being handled for writes out of the metadata cache.
Solution:
Added a dataset transfer property called "block before metadata write"
which is used by the MPI-I/O and MPI-posix drivers to sync up all the
processes before attempting a metadata write. This property is currently
only for metadata writes from the metadata cache.
Platforms tested:
IRIX64 6.5 (modi4) w/parallel
New feature.
Description:
Added MPI-posix VFL driver. This driver uses MPI to coordinate actions, but
performs I/O directly with posix sec(2) I/O functions. This driver should
_NOT_ be used if the file accessed is not on a parallel filesystem.
Platforms tested:
FreeBSD 4.6 (sleipnir) w/parallel & IRIX64 6.5 (modi4) w/parallel
Bug Fix
Description:
When parallel I/O is used, the MPI-I/O VFL driver uses a "lazy" model to
call MPI_File_set_view() in order to reduce the number of calls to this
function. However, this is unsafe, because if a collective I/O which uses
MPI derived types (and thus uses MPI_File_set_view()) is immediately
followed by an independent I/O, the code will attempt to call
MPI_File_set_view() in order to switch back to the default view of the
file. MPI_File_set_view() is a collective call however, and this causes
the application to hang.
Solution:
Removed "lazy" MPI_File_set_view() code, instead set the file view when it
is needed (with MPI derived types) and immediately set the file view back to
the default view before leaving the I/O routine.
Platforms tested:
IRIX64 6.5 (modi4) w/parallel. Also, tested with the latest development
and release code for the SAF library, which now works correctly with this
change. (Although the release branch of the SAF library seems to have a
bug, this 1.4.4 release candidate code gets as far as the version the SAF
library is released on top of (1.4.2-patch1, I believe)).
Code improvement
Description:
Some small code cleanups and took out the code the was turning off the
metadata cache for parallel I/O (!)
Platforms tested:
IRIX64 6.5 (modi4) w/parallel
Code optimization
Description:
Avoid creating MPI types (and thus requiring a MPI_File_set_view() call)
when contiguous selections are used for dataset I/O. This should be a
performance improvement for those sorts of selections.
Platforms tested:
Linux 2.2.x (eirene) w/parallel && IRIX64 6.5 (modi4) w/parallel & FORTRAN
Code cleanup
Description:
Use dataset transfer property list to hold information about the MPI types
for the current transfer, instead of setting pseudo-global variables in
the file's struct.
Platforms tested:
Linux 2.2.x (eirene) w/parallel & IRIX64 6.5 (modi4) w/parallel & FORTRAN
Code cleanup
Description:
Change MPI-I/O code to use the address of the dataset for the displacement,
instead of having a separate displacement value. Removed displacement
parameter from H5FD_mpio_setup parameters.
Platforms tested:
Linux 2.2.x (eirene) w/parallel & IRIX64 6.5 (modi4) w/parallel.
Code Improvement
Description:
Split the metadata accumulator code into two parts: one for allowing writes
of the accumulator buffer during reads (when the buffer is dirty and needs
to be flushed to disk in order to hold the new metadata being read in) and
another for only allowing writes of the buffer during writes.
This allows the MPI-I/O VFL driver to use the metadata accumulator (but
only during writes) and benefit from the reduced number of metadata I/O
operations that it brings.
Platforms tested:
IRIX64 6.5 (modi4) w/parallel
New feature.
Description:
Added a "small data" block allocation mechanism to the library, similar to
the mechanism used for allocating metadata currently.
See the RFC for more details:
http://hdf.ncsa.uiuc.edu/RFC/SmallData/SmallData.html
This reduces the number of I/O operations which hit the disk for my test
program from 19 to 15 (i.e. from 393 to 15, overall).
Platforms tested:
Solaris 2.7 (arabica) w/FORTRAN and FreeBSD 4.5 (sleipnir) w/C++
Code Cleanup
Description:
Removed some compiler warnings.
Solution:
In a few cases, NULL was being returned when a FAIL was supposed to
be returned instead. There were some header files which needed to be
included in a few of the sources. A couple of if-then statements had
assignments in the conditional part. The compiler warned that they
should have extra "()"s around them. Made the code check the values
instead.
Platforms tested:
Linux (parallel) Modi4 (parallel)
Code cleanup
Description:
Broke the FUNC_ENTER macro into several macros, with more specialized
uses (which followup mail will describe). This was designed to move
most/all of the checks which could be done at compile time to that point,
instead of needlessly performing them (over & over :-) at run-time.
This reduces the library's size (and thus staticly linked binaries) and
has a minor speedup effect also.
Platforms tested:
IRIX64 6.5 (modi4) with parallel & FORTRAN enabled, and additional testing
on FreeBSD and Solaris immediately after the checkin.
Code cleanup.
Description:
Took Robb's recent ideas for improving the FUNC_ENTER/FUNC_LEAVE macros
equivalents in the SAF library and adapted them to our library. I added
an additional macro which is equivalent to FUNC_ENTER:
FUNC_ENTER_NOINIT - Has the API tracing code, etc. from FUNC_ENTER but
none of the library or interface initialization code. This is to
be used _only_ for static functions and those which explicitly
cannot have the library or interface initialization code enabled
(like the API termination routines, etc.).
This allowed many more of the functions in the library [but not all yet :-(]
to be wrapped with FUNC_ENTER[_NOINIT]/FUNC_LEAVE pairs.
It also reduced the size of the library and executables (by cutting out a
bunch of code which was never executed), I'll e-mail the exact results when
I've finished editing it.
Platforms tested:
IRIX64 6.5 (modi4)
Code cleanup
Description:
Guard against getting into metadata broadcast in write routine and clean
up some error handling.
Platforms tested:
IRIX64 6.5 (modi4)
Bug fix/Code improvement.
Description:
Currently, the chunk data allocation routine invoked to allocate space for
the entire dataset is inefficient. It writes out each chunk in the dataset,
whether it is already allocated or not. Additionally, this happens not
only when it is created, but also anytime it is opened for writing, or the
dataset is extended. Worse, there's too much parallel I/O syncronization,
which slows things down even more.
Solution:
Only attempt to write out chunks that don't already exist. Additionally,
share the I/O writing between all the nodes, instead of writing everything
with process 0. Then, only block with MPI_Barrier if chunks were actually
created.
Platforms tested:
IRIX64 6.5 (modi4)
Code cleanup/Performance enhancement
Description:
The code to extend the file size in H5FD_mpio_flush is getting run even when
the file size doesn't change. Also, it's sort of sidestepping MPI-I/O when
extending the file, instead of using MPI-I/O features to set the file's size.
Solution:
Only extend the file's size when the allocated size has increased. Also
use MPI_File_set_size() to change the file's size.
Platforms tested:
IRIX64 6.5 (modi4)
Bug fix
Description:
Calling MPI_Get_count needs to be done with the same MPI type as was used
for the transfer and we are always using MPI_BYTE, even when a different
MPI type was used for the transfer.
Solution:
Only query MPI_Get_count with MPI_BYTE when we really used MPI_BYTE for the
transfer. Wait for later to query MPI_Get_count with other MPI types.
Platforms tested:
IRIX64 6.5 (modi4)
Performance enhancement
Description:
Doing an MPI_File_sync() just before a file is closed causing a large
performance loss.
Solution:
Add flag to MPI file driver to avoid performance the MPI_File_sync() when
the flag is set before a call to H5F_flush().
Platforms tested:
IRIX64 6.5 (modi4)
New Feature
Description:
The VFL flush function is called immediately before a file is closed.
This can cause duplicate syncronization actions to occur, if the VFL
close function also performs them.
Solution:
Added 'closing' parameter to VFL 'flush' operation. This allows the VFL
flush function to bypass operations that will be duplicated within the VFL
close function.
Additionally, use the 'closing' parameter to bypass calls to MPI_File_sync()
when set. Since MPI_File_close() also syncronizes the file, this avoids
the terrible performance hit taken when calling MPI_File_sync() as the file
is closing.
Platforms tested:
IRIX64 6.5 (modi4)
Code cleanup
Description:
The parallel I/O file driver is optimized to only write metadata with one
process (and broadcast the results to the other processes). This is
currently enabled by a separate call to H5FD_mpio_tas_allsame() before
each metadata write to the file. This can easily lead to problems where
the prelude function call is omitted before the actual write code or, in
a threaded environment, lead to race condititions where the value set is
reset before being used.
Solution:
Since we only want to write metadata from one process, key off of the 'type'
parameter (which has information about whether the data being written it
metadata or raw data) to H5FD_mpio_write() as the method for determining
whether to only write from one process or not.
Platforms tested:
IRIX64 6.5 (modi4)
New Feature
Description:
Currently, only process 0 writes metadata to disk, leading to a potential
performance bottleneck as the other processors wait for it to catch up.
Solution:
Rotate the metadata responsibilities among all processes, speading out the
work.
Platforms tested:
IRIX64 6.5 (modi4)
Small code cleanup/improvement
Description:
Query the MPI rank only once and store it in the file structure for each
process.
Platforms tested:
IRIX64 6.5 (modi4)
Features.
Description:
Error stack used to report only hdf5 predefined error messages
because it takes only static strings. Runtime defined messages
were not pushed to the stack.
Added the means and macros to push MPI error strings onto the
hdf5 error stack. Added a new minor error class as H5E_MPIERR
for this class of messages.
H5Epulbic.h, H5E.c:
Added H5E_MPIERR and its minor class description.
H5Eprivate.h:
Added HMPI_XXX macros to push MPI error strings to the stack.
H5FDmpio.c:
Changed couple places to use the new macros to test the new
macros. A more through changes to make use of these new
macros will be done later.
Platforms tested:
eirene (serial, parallel)
modi4(parallel)
Purpose:
Bug Fix
Description:
Committing the changes in the MPI/IO stuff so that parallel HDF5
will work on HP-UX. It seems that on HP-UX, the MPI_Status variable
needs to be initialized to 0 for it to work (i.e., in some other MPI
calls, if there's garbage in the MPI_Status variable, then it will
barf).
Solution:
Initialized to 0.
Platforms tested:
HP-UX parallel, Linux.
Purpose:
Modify H5Fclose behavior
Description:
The HDF5 actual file close behaves in several ways in terms of if there
are still objects(dataset, group, datatype) opened in file.
Solution:
Added a new file access property, file close degree. It has four values,
H5F_CLOSE_DEFAULT
H5F_CLOSE_WEAK
H5F_CLOSE_SEMI
H5F_CLOSE_STRONG
The way a file is closed is decided by these values.
Platforms tested:
IRIX64 6.5, SunOS 5.6, FreeBSD 4.4
Bug fix
Description:
A couple more compile problems from the recent elimination of IDs from
internal function calls.
Platforms tested:
IRIX64 6.5 (modi4)
Code cleanup
Description:
Get rid of IDs from internal function calls and some small cleanups from
the old-stype => generic property list conversion.
Platforms tested:
FreeBSD 4.4 (hawkwind)
