Adding the first round of patches about supporting collective chunk IO in HDF5
Description:
The current HDF5 library doesn't support collective MPIO with chunk storage. When users set collective option in their data transfer with chunk storage, the library silently converted the option to INDEPENDENT and that caused trememdous performance penalty. Some application like
WRF-parallel HDF5 IO module has to use contiguous storage for this reason. However, chunking storage has its own advantage(supporting compression filters and extensible dataset), so to make collective MPIO possible inside HDF5 with chunking storage is a very important task.
This check-in make collective chunk IO possible for some special cases. The condition is as follows(either case is fine with using collective chunk IO)
1. for each process, the hyperslab selection of the file data space of each dataset is regular and it is fit in one chunk.
2. for each process, the hyperslab selection of the file data space of each dataset is single and the number of chunks for the hyperslab selection should be equal.
Solution:
Lift up the contiguous storage requirement for collective IO.
Use H5D_isstore_get_addr to get the corresponding chunk address. Then the original library routines will take care of getting the correct address to make sure that MPI FILE TYPE is built correctly for collective IO>
Platforms tested:
arabica(sol), copper(AIX), eirene(Linux)
parallel test is checked at copper.
Misc. update:
Code cleanup
Description:
Clean up almost all warnings from Windows builds.
Platforms tested:
FreeBSD 4.10 (sleipnir) w/parallel
Too minor to require h5committest
Code cleanup & minor optimization
Description:
Re-work the way interface initialization routines are specified in the
library to avoid the overhead of checking for them in routines where there is
no interface initialization routine. This cleans up warnings with gcc 3.4,
reduces the library binary size a bit (about 2-3%) and should speedup the
library's execution slightly.
Platforms tested:
FreeBSD 4.10 (sleipnir) w/gcc34
h5committest
Code optimization
Description:
Use 'size_t' instead of 'hsize_t' to track the number of elements in
memory buffers, especially for type conversion.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.10 (sleipnir) w/parallel
Too minor to require h5committest
Code cleanup & optimization
Description:
Remove old structures that used a union to store information about the
dataspace extent and just store the information directly in the dataspace
extent itself.
Remove ifdef'd references to permutation ordering in dataspaces. We'll
definitely need more than this code if/when we implement this feature.
Change allocation of dataspace information from calloc() to malloc().
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.10 (sleipnir) w/parallel
Too minor to require h5committest
Code optimization
Description:
Revised dataspace selections to use a more "object oriented" mechanism
to set the function pointers for each selection and selection iterator. This
reduces the amount and number of times that dataspace selection info has to
be copied.
Additionally, change hyperslab selection information to be dynamically
allocated instead of an inline struct.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.10 (sleipnir) w/parallel
Too minor to require h5committest
Refactor code
Description:
Move chunk and contiguous cached raw data from file information to dataset
information. This simplifies a number of internal interfaces, aligns the
code with it's purpose better and should allow more optimizations to the
chunked data I/O performance.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.10 (sleipnir)
h5committest
Misc. update:
Code optimization & bug fix
Description:
When dimension information is being stored in the storage layout message
on disk, it is stored as 32-bit quantities, possibly truncating the dimension
information, if a dimension is greater than 32-bits in size.
Solution:
Fix the storage layout message problem by revising file format to not store
dimension information, since it is already available in the dataspace.
Also revise the storage layout data structures to be more compartmentalized
for the information for contiguous, chunked and compact storage.
Platforms tested:
FreeBSD 4.9 (sleipnir) w/parallel
Solaris 2.7 (arabica)
h5committest
Bug fix
Description:
Add prototypes for H5S_select_serial_size() and H5S_select_serialize()
that were causing the library to fail when built with C++ compiler.
Platforms tested:
FreeBSD 4.9 (sleipnir) w/CC=g++
Only affects C++ builds, so no h5committest
Code optimization
Description:
Expand the use of macros to inline trivial function pointer lookup and
calls to reduce the overall number of functions invoked during normal operation
of the library.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir) w/parallel
Too minor to require h5committest
Code optimization
Description:
Defer creating the span trees for hyperslab selections until they are
actually needed (which may be never, in certain circumstances).
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir) w/parallel
Code optimization
Description:
Further reduce the number of copies we make of a hyperslab selection for
chunked I/O, especially when we are only going to throw the old selection away
for a new one.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir) w/parallel
Code optimization
Description:
Reduce the number of times the number of elements in a selection is
computed.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir) w/parallel
too minor to require h5committest
Code optimization
Description:
Move the element size for the selection into the selection iterator instead
of always passing it as a parameter.
Also, eleminate another 64-bit multiply for "all" selections.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir)
too minor to require h5committest
Code optimization
Description:
Instead of dynamicly allocating various arrays for various pieces of
information about a selection or selection iterator, just use fixed size
array of size H5S_MAX_RANK (as the rest of the library does).
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir)
h5committest
Code optimization
Description:
Change algorithm to directly use coordinates describing a chunk's position
in a dataspace instead of creating a dataspace with the chunk's position
selected. This reduces the number of copies of dataspaces we need to keep
around.
Platforms tested:
Solaris 2.7 (arabica)
FreeBSD 4.9 (sleipnir) w/parallel
too minor to require h5committest
Code cleanup/optimization
Description:
Query property list values once, at the beginning of the I/O routines,
instead of querying the property list values multiple (lots!) of times in
lower level routines.
Solution:
Create "property list caches" for internal library queries of the property
list values.
Platforms tested:
IBM p690 (copper) w/parallel & fphdf5
h5committest
Bug fix.
Description:
Using a selection offset with hyperslab selections in chunked datasets
was getting into an infinite loop and hanging the application.
Solution:
Apply the selection offset to the hyperslab selection properly.
Platforms tested:
FreeBSD 4.9 (sleipnir) w & w/o parallel
h5committest
Bug fix & code cleanups
Description:
Change our use of MPI derived datatypes to not create datatypes with
"0-sized" lengths, which causes the LANL Q machine to hang.
Also, get rid of "prefer MPI derived datatypes" environment variable since
it has no advantage.
Platforms tested:
FreeBSD 4.9 (sleipnir) w & w/o parallel
h5committest
Code cleanup
Description:
Straighten out more goofiness in the MPI code dealing with collective I/O
transfers - mostly make certain that a view is set if-and-only-if collective
I/O is occurring on raw data (and vice versa for views and independent I/O)
Platforms tested:
FreeBSD 4.9 (sleipnir) w/parallel & FPHDF5
too minor to repquire h5committest
Bug fix
Description:
I/O on chunked datasets with point selections was not working correctly.
Solution:
Re-wrote some parts of raw data I/O routines that build the selections for
each chunk to correctly handle point selections.
Platforms tested:
FreeBSD 4.8 (sleipnir)
h5committest
Code cleanup & performance improvements
Description:
Optimize hyperslabs that are built to detect situations where "regular"
hyperslabs can be recovered from span tree descriptions.
Also, improve "same shape" routine to correctly work with all the different
combinations of selections.
Platforms tested:
FreeBSD 4.8 (sleipnir) w/C++
FreeBSD 4.8 (sleipnir) w/parallel
h5committested
Performance improment
Description:
Speed up chunked dataset I/O. This breaks down into several areas:
- Compute chunk selections in the file by using hyperslab operations
instead of iterating over each element in the selection.
- If the file and memory selections are the same shape, use the file
chunk selections to compute the memory chunk selections.
This required several additional dataspace, dataspace selection and
hyperslab routines.
Platforms tested:
h5committestted (although Fortran tests failed for some reason)
Code cleanup
Description:
Clean up warnings exposed by compiling on O2K. Also, revert some of Bill
and my changes to the H5S_mpi_opt_types_g, etc. and settle them back into their
original location.
Platforms tested:
h5committested.
Code cleanup.
Description:
Move many package or internal function prototypes and macro definitions
into tighter scope according to their current use.
Added more comments where appropriate.
Eliminate ancient, unused functions.
Added a couple "accessor" functions to get parts of data structures which
were moved out of scope.
Platforms tested:
h5committested
New feature/enhancement
Description:
Chunked datasets are handled poorly in several circumstances involving
certain selections and chunks that are too large for the chunk cache and/or
chunks with filters, causing the chunk to be read from disk multiple times.
Solution:
Rearrange raw data I/O infrastructure to handle chunked datasets in a much
more friendly way by creating a selection in memory and on disk for each chunk
in a chunked dataset and performing all of the I/O on that chunk at one time.
There are still some scalability (the current code attempts to
create a selection for all the chunks in the dataset, instead of just the
chunks that are accessed, requiring portions of the istore.c and fillval.c
tests to be commented out) and performance issues, but checking this in will
allow the changes to be tested by a much wider audience while I address the
remaining issues.
Platforms tested:
h5committested, FreeBSD 4.8 (sleipnir) serial & parallel, Linux 2.4 (eirene)
Code cleanup
Description:
Clean up miscellaneous warnings which have crept into the code.
Fix "_POSIX_C_SOURCE not defined" warning on FreeBSD.
Adjust gcc compiler flags to be more concise for production mode.
Refactor the H5O code so that there is a stronger boundary between code
in the H5O package and code in the library which just calls H5O routines.
Platforms tested:
Tested h5committest {arabica (fortran), eirene (fortran, C++)
modi4 (parallel, fortran)}
FreeBSD 4.7 (sleipnir) serial & parallel and gcc 2.95.4 & gcc 3.2.2
Misc. update:
Update MANIFEST if you add or remove any file.
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
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:
__DLL__ is a keyword in some platforms and __DLL__ is also defined as a macro for windows DLL applications.
That causes problems.
Description:
Solution:
Use H5_DLL*** to replace __DLL***__ at all header files.
Change the macro defination at H5api_adpt.h.
Platforms tested:
linux2.2.18smp, irix64, solaris 2.7 and windows 2000
Code cleanup
Description:
Two of the parameters to H5S_select_hyperslab are defined as
"const hsize_t foo[]", and then "foo" is assigned another value, which is
technically against the C standard and is now causing problems for a user
with a more strict compiler.
Solution:
Changed "const hsize_t foo[]" to "const hsize_t *foo".
Platforms tested:
FreeBSD 4.6 (sleipnir) - not a significant enough change to worry about
multi-platform testing.
Bug fix/Code cleanup/New Feature
Description:
Correct problems with writing fill-values to external storage and allocate
the data storage at the correct times.
Also, mostly straighten out the strange code which allocates and fills
raw data storage for datasets. Things are still a bit odd in that the
fill-values for chunked datasets are written when the space is allocated,
instead of in a separate routine, but there are two reasons for this:
it's inefficient (especially in parallel) to iterate through all the chunks
twice, and (more importantly) the space needed to store compressed chunks
isn't known until we've got a buffer of compressed fill-values ready to
write to the chunk.
Additionally, add in the H5D_SPACE_ALLOC_INCR and H5D_SPACE_ALLOC_DEFAULT
setting for the "space time", which incorporate the previous behavior of
the space allocation for chunked datasets.
The default settings for the different types of dataset storage are now
as follows:
Contiguous - Late
Chunked - Incremental
Compact - Early
This checkin also incorporates a change to the behavior of external data
storage in two ways - fill-values are _never_ written to external storage
(under the assumption that writing fill-values is triggered by allocating
space in an HDF5 file, and since space is not allocated in the file, the
fill-values should not be written) and external data files are now created
if they don't exist when data is written to them. The fill-value will
probably need to be revisited at some time in the future, this just seemed
like the safer course currently.
I think I cleaned up some compiler errors also, before getting bogged down
in the fixes for the space allocation and fill-values.
Platforms tested:
FreeBSD 4.6 (sleipnir) w/serial & parallel. Will be testing on IRIX64
6.5 (modi4) in serial & parallel shortly.
Code cleanup
Description:
Move operations on each type of selection into the source code file for
each selection type (all->H5Sall.c, hyperslab->H5Shyper.c, etc.)
Remove central H5S_select_<foo> operations, instead calling the operations
through function pointers in each selection (a much more object-oriented
approach).
Platforms tested:
FreeBSD 4.6 (sleipnir)
Large code cleanup/re-write
Description:
This is phase 1 of the data I/O re-architecture, with the following changes:
- Changed the selection drivers to not actually do any I/O, they
only generate the sequences of offset/length pairs needed for
the I/O (or memory access, in the case of iterating or filling
a selection in a memory buffer)
- Wrote more abstract I/O routines which get the sequence of offset/
length pairs for each selection and access perform the I/O or
memory access.
Benefits of this change include:
- Removed ~3400 lines of quite redundant code, with corresponding
reduction in the size of library binary.
- Any selection can now directly access memory when performing I/O,
if no type conversions are required, instead of just "regular"
hyperslab and 'all' selections, which speeds up I/O.
- Sped up I/O for hyperslab selections which have contiguous lower
dimensions by "flattening" them out into lesser dimensional objects
for the I/O.
No file format or API changes were necessary for this change.
The next phase will be to create a "selection driver" for each type of
selection, allowing each type of selection to directly call certain
methods that only apply to that type of selection, instead of passing
through dozens of functions which have switch statements to call the
appropriate method for each selection type. This will also reduce
the amount of code in the library and speed things up a bit more.
Phase 3 will involve generating an MPI datatype for all types of selections,
instead of only "regular" hyperslab and 'all' selections. This will
allow collective parallel I/O for all I/O operations which don't
require type conversions. It will also open up the door for allowing
collective I/O on datasets which require type conversion.
Phase 4 will involve changing the access pattern to deal with chunked
datasets in a more optimal way (in serial).
Phase 5 will deal with accessing chunked datasets more optimally for
collective parallel I/O operations.
Platforms tested:
FreeBSD 4.6 (sleipnir) w/ parallel & C++ and IRIX64 6.5 (modi4) w/parallel
New feature.
Description:
There is some discussion among the SAF team as to whether it is better
to use MPI derived types for raw data transfers (thus needing a
MPI_File_set_view() call), or whether it is better to use a sequence of
low-level MPI types (i.e. MPI_BYTE) for the raw data transfer.
Solution:
Added an in internal flag to determine whether derived types are preferred
(the default), or whether they should be avoided. An environment variable
("HDF5_MPI_PREFER_DERIVED_TYPES") can be set by users to control whether MPI
types should be used or not. Set the environment variable to "0" (i.e.:
'setenv HDF5_MPI_PREFER_DERIVED_TYPES 0' to avoid using MPI derived types.
Platforms tested:
IRIX64 6.5 (modi4) w/parallel
Bug fix.
Description:
Return H5S_all_opt_possible() prototype to original position in H5Spkg.h,
it is not related to parallel I/O.
Platforms tested:
FreeBSD 4.5 (sleipnir)
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:
Remove an extraneous data structure for tracking the location of a
hyperslab iterator when walking through regular hyperslab selections.
Platforms tested:
FreeBSD 4.5 (sleipnir)
Code cleanup
Description:
Previously, the I/O pipeline (pline), external file list (efl) and fill-
value (fill) structs were passed down the raw data function call chain,
even into and/or through functions which didn't use them. Since all three
of these pieces of information are available from the dataset creation
property list, just pass the dataset creation property list down the
function call chain and query for the information needed in a particular
function.
Platforms tested:
FreeBSD 4.5 (sleipnir)
Purpose:
New feature
Description:
Fill-value's behaviors for contiguous dataset have been redefined.
Basicly, dataset won't allocate space until it's necessary. Full details
are available at http://hdf.ncsa.uiuc.edu/RFC/Fill_Value, at this moment.
Platforms tested:
Linux 2.2.
Purpose:
H5Dset_extent API function
Description:
1 new public function: H5Dset_extent, that allows shrinking a dataset
several private functions that free the previously allocated chunks
H5F_istore_prune_by_extent, H5B_prune_by_extent : removes the chunks
H5F_istore_initialize_by_extent : initializes them with the fill value
H5S_select_fill is used to fill the selection in the memory buffer
Platforms tested:
Windows 2000
SUN( arabica)
LINUX (eirene)
IRIX64 (modi4)
New Feature
Description:
Added new H5Dfill() routine to fill the elements in a selection for a
memory buffer with a fill value. This is a user API wrapper around some
internal routines which were needed for the fill-value modifications
from Raymond as well as Pedro's code for reducing the size of a chunked
dataset.
Platforms tested:
FreeBSD 4.5 (sleipnir) [and IRIX64 6.5 (modi4) in parallel, in a few
minutes]
Bug Fix & Code Cleanup
Description:
The MPI-IO optimized transfer routines
(H5S_mpio_spaces_read/H5S_mpio_space_write) are not being invoked in all
the cases where they could be used.
Additionally, the code for determining if an optimized transfer is wrapped
into the actual I/O transfer routine in a very confusing way.
Solution:
Re-enabled MPI-IO optimized transfer routines in all the cases where they
should work.
Extracted all the pre-conditions for optimized transfers into separate
routines from the transfer routines.
Platforms tested:
FreeBSD 4.5 (sleipnir) & IRIX64 6.5 (modi4)
Code cleanup
Description:
The function pointer for doing parallel I/O was being set in the H5D_read
and H5D_write routines instead of in H5S_conv, where it belongs.
Solution:
Added an extra 'flags' parameter to H5S_find to allow the function to
determine when to set the MPIO functions instead of the regular optimized
I/O routines and get rid of code messing with the function pointer in the
H5D routines.
Platforms tested:
IRIX64 6.5 (modi4)
Bug Fix & Feature
Description:
The selection offset was being ignored for optimized hyperslab selection
I/O operations.
Additionally, I've found that the restrictions on optimized selection
I/O operations were too strict and found a way to allow more hyperslabs
to use the optimized I/O routines.
Solution:
Incorporate the selection offset into the selection location when performing
optimized I/O operations.
Allow optimized I/O on any single hyperslab selection and also allow
hyperslab operations on chunked datasets.
Platforms tested:
FreeBSD 4.5 (sleipnir)
Bug Fix
Description:
If a non-zero fill-value is used for a chunked dataset, any non-existent
chunked read with an "all" selection (or a contiguous hyperslab selection)
will return zero for those instead of the user's fill-value.
Solution:
Fixed I/O code to pass down fill-value to "optimized" I/O routines, so it
will be available to fill the user's buffer with.
Platforms tested:
FreeBSD 4.5 (sleipnir)
Description:
modified the below files to inlude a new public function H5Dset_extend, similar
to H5Dextend, but it can lower the dimension
this function requires 2 more new private functions:
H5D_set_extend
H5S_set_extend
Platforms tested:
