| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| | |
|
| | |
|
| |
|
|
|
|
| |
so only proc 0 writes it to disk. The elimination is activated
only when the environment variable HDF5_MPI_1_METAWRITE is nonzero;
otherwise all processes that opened the file will write the metadata to disk.
|
| |
|
|
| |
in condition.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
./INSTALL_MAINT
./README
./RELEASE
Partially updated for second alpha, but haven't updated
version numbers yet.
./src/H5.c
./src/H5A.c
./src/H5AC.c
./src/H5B.c
./src/H5D.c
./src/H5F.c
./src/H5Fcore.c
./src/H5Ffamily.c
./src/H5Fistore.c
./src/H5Fmpio.c
./src/H5Fsec2.c
./src/H5Fsplit.c
./src/H5Fstdio.c
./src/H5G.c
./src/H5Gnode.c
./src/H5HG.c
./src/H5HL.c
./src/H5I.c
./src/H5MM.c
./src/H5MMprivate.h
./src/H5O.c
./src/H5Oattr.c
./src/H5Ocomp.c
./src/H5Ocont.c
./src/H5Odtype.c
./src/H5Oefl.c
./src/H5Olayout.c
./src/H5Oname.c
./src/H5Osdspace.c
./src/H5Oshared.c
./src/H5Ostab.c
./src/H5P.c
./src/H5S.c
./src/H5T.c
./src/H5Tconv.c
./src/H5detect.c
./test/hyperslab.c
./test/istore.c
Changed memory allocation functions so they fail instead of
dumping core. The `x' was removed from the name to remind us
of that: H5MM_xmalloc() -> H5MM_malloc(), etc.
H5MM_calloc() takes one argument like H5MM_malloc() instead of
two like calloc() because we almost always called it with `1'
for one of the arguments anyway. The only difference between
the two functions is that H5MM_calloc() returns memory which
is initialized to zero.
./src/H5Gent.c
./src/H5Gprivate.h
Removed H5G_ent_calloc() since it wasn't used.
./src/H5Fistore.c
Fixed a bug found by Albert. Thanks, Albert! This fix
combined with the changes to memory allocation prevent the
library from failing an assertion if the application uses an
unreasonable size for chunks (like Alberts 10000x10000x4).
./src/H5MF.c
./src/H5MFprivate.h
Changed H5MF_free() to H5MF_xfree() since calling it with an
undefined address is allowed.
|
| |
|
|
| |
H5F_mpio_open.
|
| |
|
|
|
|
| |
which can be set by passing to H5Fcreate and H5Fopen
a string of characters (one character for each flag to turn on)
as the value of MPI_Info with key H5F_MPIO_DEBUG_KEY.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
./src/H5F.c
./src/HFcore.c
./src/H5Ffamily.c
./src/H5Flow.c
./src/H5Fmpio.c
./src/H5Fprivate.h
./src/H5Fsec2.c
./src/H5Fsplit.c
./src/H5Fstdio.c
./src/H5MF.c
./src/H5P.c
./src/H5Ppublic.h
Added H5Pset_alignment() so that it is now possible to align
file allocation requests on application-specified
boundaries. Any request >= the specified threshold will begin
on an address which is a multiple of the specified
alignment. Defaults are one for threshold and alignment. The
alignment is done on relative addresses, so the size of the
user block can affect the location of the data in the file.
./src/H5D.c
./src/dsets.c
Added a test for, and fixed the data space caching bug in
datasets. Extending a dataset through one handle will cause
all other handles to the same dataset to get the new dataset
size.
./src/H5S.c
./src/H5Sprivate.h
Removed an unused argument from H5S_read() which duplicated
information from the other argument.
./config/linux
Made `--enable-parallel' the default on my system. It used to
be that way before but then I accidently turned it off and
forgot about it.
./src/H5Fmpio.c
Qualified some function arguments with __unused__. Changed a
couple places where NULL was returned on error for herr_t
functions.
./src/H5P.c
Removed unused autos from H5Pset_mpi().
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
H5Fmpio.c was not part of the serial distribution. The auto-
dependence configuration runs for serail distribution only, thus
H5Fmpio.c has no dependence. It ended up not recompiled even
when header files it depends on are changed.
Solution:
Changed configuration to always include this file.
Enclosed the whole code in a big #ifdef HAVE_PARALLEL macro
so that they would be compiled for parallel distribution only.
Platform tested:
O2K
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Parallel File open has a racing condition if file already exists.
Some processes see file exists, then truncate it (some system
by remove and create). Some processes does not see the file
exist, then attempt to just create it. Inconsistency.
Solution:
Instead of opening the file with COMM_SELF (which results in
racing condition in routine that calls it), open it with the
communicator in access_parms. (This assumes this access call
must be called collectively.)
Platform tested:
O2K.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
mode to use for data transfer. Currently, INDEPENDENT transfer
is the default mode. Collective is supported for files opened
via the MPIO driver. Specific changes to individual files follow.
H5D.c:
Replaced previous kludge of passing xfer_mode via the file access_mode
with xfer_mode function parameter all the way to the low level I/O
drivers.
H5Dprivate.h:
The H5D_xfer_t.xfer_mode is now used in serial and parallel modes.
Removed the "HAVE_PARALLEL" macro condition to support it.
H5Dpublic.h:
Added H5D_XFER_DFLT as a new value for H5D_transfer_t and removed
the HAVE_PARALLEL condition macro to make the H5D_transfer_t
definition available to both serial and parallel version.
H5B.c:
H5F.c:
H5Farray.c:
H5Fcore.c:
H5Ffamily.c:
H5Fistore.c:
H5Flow.c:
H5Fmpio.c:
H5Fsec2.c:
H5Fsplit.c:
H5Fstdio.c:
H5Gnode.c:
H5HG.c:
H5HL.c:
H5O.c:
H5Sprivate.h:
H5Ssimp.c:
debug.c:
Added xfer_mode to the argument lists of the following routines to
specify data transfer mode. Then updated calls to these routines
with appropriate transfer mode. Whenever cases that collective
mode is inappropriate (e.g. object header read-write, flush),
H5D_XFER_DFLT is used. The H5D_XFER_DFLT mode is currently
acting the same as the independent mode.
H5F_block_read
H5F_block_write
H5F_arr_read
H5F_arr_write
H5F_core_read
H5F_core_write
H5F_fam_read
H5F_fam_write
H5F_mpio_read
H5F_mpio_write
H5F_low_write
H5F_low_read
H5F_sec2_read
H5F_sec2_write
H5F_split_read
H5F_split_write
H5F_stdio_read
H5F_stdio_write
H5S_simp_fgath
H5S_simp_fscat
H5S_simp_read
H5S_simp_write
H5Fprivate.h:
H5P.c:
Removed access_mode field from struct H5F_access_t type.
The access-mode was the initial design to control independent
or collective I/O mode. It is now supported by the data
transfer mode.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
| |
When data is read or written out, some layers did an immediate
return(SUCCEED) if number of elements to be accessed is zero.
This caused problems when collective access is used. In a
collective call, all processes in a communicator must call
the lowest layer of MPIO routines even if some of them happen
to access no date.
Solutions:
Eliminated all those pre-mature return if #element==0;
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
used to open an HDF5 file with MPIO access. The design has changed
to provide access-mode controll (independent or collective access)
as an argument in the H5D read/write calls.
Also removed the macros H5FACC_INDEPENDENT and H5FACC_COLLECTIVE.
The original thinking was to define how a file is independent or
collective access. They are replaced by the enum H5D_transfer_t
with values of H5D_XFER_INDEPENDENT and H5D_XFER_COLLECTIVE.
These values are used in the H5Pset_xfer to setup a data transfer
property list to be used during the H5D read/write calls.
The original access_mode field in the File access structure is
still there. It would be removed soon once the transfer control
can be passed down to the low level I/O routines.
|
| |
|
|
| |
write and read based on the value of the access_parms.
|
| |
|
|
|
| |
the access_template, instead of COMM_WORLD and INFO_NULL as
before.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
----------------------
./configure.in
./src/H5Fprivate.h
./src/H5Fsec2.c
We now detect and use lseek64() on systems that have it (e.g.,
Irix64) and are able to generate >2GB files on Irix XFS file
systems (and anything else that supports large files). This
change also removed some warning messages from the Irix `-64'
compiler.
> $ ls -l istore.h5
> -rw-r--r-- 1 matzke meshtv 8605436856 Feb 17 14:03 istore.h5
./configure.in
./src/H5Fprivate.h
./src/H5Fstdio.h
We now detect and use fseek64() on systems that have it (e.g.,
Irix64) and are able to generate >2GB files on Irix XFS file
systems (and anything else that supports large files). This
change also removed some warning messages from the Iris `-64'
compiler.
./src/H5E.c
./src/H5Epublic.h
Added the H5E_OVERFLOW error to signal file address overflow.
./src/H5Fpublic.h
./examples/h5_chunk_read.c
./examples/h5_compound.c
./examples/h5_extend_write.c
./examples/h5_group.c
./examples/h5_read.c
./examples/h5_write.c
./html/Datasets.html
./html/Files.html
./html/H5.api.html
./html/H5.intro.html
./html/H5.sample_code.html
./html/ph5example.c
./html/review1.html
./test/cmpd_dset.c
./test/dsets.c
./test/extend.c
./test/tfile.c
./test/th5p.c
./test/theap.c
./test/tohdr.c
./test/tstab.c
./testpar/phdf5.c
Renamed file access constants to follow the naming scheme.
Also changed the base names a little to be more accurate as to
what they do. The old names H5ACC_WRITE and H5ACC_OVERWRITE
will temporarily work.
H5ACC_DEFAULT --> H5F_ACC_RDONLY for H5Fopen()
H5ACC_DEFAULT --> H5F_ACC_EXCL for H5Fcreate()
H5ACC_WRITE --> H5F_ACC_RDWR
H5ACC_OVERWRITE --> H5F_ACC_TRUNC
Albert or Kim: The H5ACC_INDEPENDENT and H5ACC_COLLECTIVE
macros in H5Fpublic.h should be an enum typedef and have names
more like H5F_MPIO_INDEPENDENT and H5F_MPIO_COLLECTIVE. Also
change the access_mode argument of H5Cset_mpio().
H5Fcreate() and H5Fopen() are more strict now about which
flags are acceptable for the operation.
./src/H5Fprivate.h
./src/H5F.c
./src/H5C.c
Changed the file access template to make it more general. A
union contains a struct for each type of low-level driver and
the default template is initialized at run-time.
./src/H5Fpublic.h
./src/H5F.c
Added H5Fget_access_template() and cleaned up
H5Fget_create_template().
./src/H5C.c
The H5Cset_mpi() no longer trashes the file access template
when an error is detected. We check for errors and *then*
update the file access template.
Added H5C_close() so Albert and Kim have a place to release
the MPI communicator and info from the file access property
list.
Kim or Albert: I notice in H5Cset_mpi() you copy the MPI
communicator. Do you need to do something similar in
H5C_copy()?
./src/H5F.c
Added more error checking for the file creation and access
property lists because it used to be possible to make the
library dump core by swapping the creation and access property
list ID numbers of H5Fcreate().
./test/istore.c
./test/tfile.c
./test/th5p.c
./test/theap.c
./test/tohdr.c
./test/tstab.c
./testpar/testphdf5.c
One must pass H5C_DEFAULT as the file creation or access
property list in order to get the default property list. It
is no longer possible to pass zero or any other arbitrary bad
object ID.
./src/H5Fcore.c
./src/H5Ffamly.c
./src/H5Flow.c
./src/H5Fmpio.c
./src/H5Fsec2.c
./src/H5Fsplit.c
./src/H5Fstdio.c
./src/H5MF.c
The file access property list is passed to all H5F_low_...()
functions and to the drivers.
./src/H5Fcore.c
The block size can be set at run time on a per-file basis
instead of at compile time across all files. The "5000 items in
a group test" now takes 1.6 seconds.
./src/H5private.h
Removed inclusion of mpi.h and mpio.h since they're included
from H5public.h.
./src/H5Cpublic.h
./src/H5C.c
Added H5Cset_stdio(), H5Cset_sec2(), H5Cset_core(),
H5Cset_split(), and H5Cset_family() in addition to the
H5Cset_mpio() that Kim and Albert already wrote. We still
need the H5Cget_driver() and an H5Cget...() counterpart for
each of those functions. The split and family drivers still
need a little work but I'm checking this in anyway.
|
| |
|
|
| |
In H5Fprivate.h made MPI-IO the default low-level I/O module when HAVE_PARALLEL.
|
| |
|
|
|
| |
MPIO lower interface layer (H5Fmpio.c) has been commited by Kim already.
All PHDF5 codes are "bracket'ed" by #ifdef HAVE_PARALLEL macro.
|
| |
|
|
|
|
|
| |
call
MPI_File_set_size (a collective operation).
Also changed/added some error messages.
|
| |
|
|
|
| |
and removed MPI_Init and MPI_Finalize from H5.c
(these should be done in the user program).
|
| |
|
