- use local function for the decision making, whether worldComm or a
dedicated communicator is needed (and which sibling ranks are
involved)
Co-authored-by: mattijs <mattijs>
- previously checked on destruction, but it is robuster to check for a
locally defined communicator during construction
- add InfoProxy output for fileOperation
ENH: add fileOperation::storeComm()
- transfers management of the communicator from external to internal.
Use with caution
- make fileHandler deletion mechanism more
transparent by providing a nullptr signature. A nullptr parameter
is already being used in the argList destructor for shutdown, but that
relied on an implicit conversion to autoPtr to trigger things.
- improved handling of file handler replacement.
Previously had a very basic check on old vs new handlers using their
type() values (string comparison!!), which would unfortunately
prevent proper swapping of the contents.
Check the actual pointers instead.
As part of the change, treat any empty autoPtr as no-op instead of as
deletion (which is handled explicitly as nullptr instead).
In addition to making the internal logic simpler, it means that the
current file handler always changes to a valid state without
inadvertently removing everything and falling back to creating a new
default handler (again).
This handling of no-ops also simplifies call code. For example,
<code>
autoPtr<fileHandler> oldHandler;
autoPtr<fileHandler> writeHandler;
word handlerName;
if (arg.readIfPresent("writeHandler", handlerName))
{
writeHandler = fileOperation::New(handlerName);
}
oldHandler = fileHandler(std::move(writeHandler));
... do something
writeHandler = fileHandler(std::move(oldHandler));
</code>
If the "writeHandler" is not specified, each call is a no-op.
If it is specified, the handlers are swapped out each time.
- the management of the fileHandler communicators is now encapsulated
privately (managedComm_) with the final layer being responsible for
cleaning up after itself. This makes delegation/inheritance clearer
and avoids the risk of freeing an MPI communicator twice.
STYLE: uniformFile static check relocated to fileOperation layer
- recover the target of symbolic links.
This is needed when re-creating a file tree on another rank.
ENH: handle checkGzip, followLink flags in fileHander filePath()
- previously just relied on the backend defaults, now pass through
- separate init(...) for common constructor init steps
ENH: extend rmDir to handle removal of empty directories only
- recursively remove directories that only contain other directories
but no other contents. Treats dead links as non-content.
- additional Pstream::broadcasts() method to serialize/deserialize
multiple items.
- revoke the broadcast specialisations for std::string and List(s) and
use a generic broadcasting template. In most cases, the previous
specialisations would have required two broadcasts:
(1) for the size
(2) for the contiguous content.
Now favour reduced communication over potential local (intermediate)
storage that would have only benefited a few select cases.
ENH: refine PstreamBuffers access methods
- replace 'bool hasRecvData(label)' with 'label recvDataCount(label)'
to recover the number of unconsumed receive bytes from specified
processor. Can use 'labelList recvDataCounts()' to recover the
number of unconsumed receive bytes from all processor.
- additional peekRecvData() method (for transcribing contiguous data)
ENH: globalIndex whichProcID - check for isLocal first
- reasonable to assume that local items are searched for more
frequently, so do preliminary check for isLocal before performing
a more costly binary search of globalIndex offsets
ENH: masterUncollatedFileOperation - bundled scatter of status
BUG: masterUncollatedFileOperation checking of file-size
- used Foam:fileSize check to decide on scheduled/nonBlocking but this
was being done on all ranks and subsequently broadcast.
Now avoid unnecessary filesystem access on non-master ranks.
- split off a Pstream::genericBroadcast() which uses UOPBstream during
serialization and UOPBstream during de-serialization.
This function will not normally be used directly by callers, but
provides a base layer for higher-level broadcast calls.
- low-level UPstream broadcast of string content.
Since std::string has length and contiguous content, it is possible
to handle directly by the following:
1. broadcast size
2. resize
3. broadcast content when size != 0
Although this is a similar amount of communication as the generic
streaming version (min 1, max 2 broadcasts) it is more efficient
by avoiding serialization/de-serialization overhead.
- handle broadcast of List content distinctly.
Allows an optimized path for contiguous data, similar to how
std::string is handled (broadcast size, resize container, broadcast
content when size != 0), but can revert to genericBroadcast (streamed)
for non-contiguous data.
- make various scatter variants simple aliases for broadcast, since
that is what they are doing behind the scenes anyhow:
* scatter()
* combineScatter()
* listCombineScatter()
* mapCombineScatter()
Except scatterList() which remains somewhat different.
Beyond the additional (size == nProcs) check, the only difference to
using broadcast(List<T>&) or a regular scatter(List<T>&) is that
processor-local data is skipped. So leave this variant as-is.
STYLE: rename/prefix implementation code with 'Pstream'
- better association with its purpose and provides a unique name
- native MPI min/max/sum reductions for float/double
irrespective of WM_PRECISION_OPTION
- native MPI min/max/sum reductions for (u)int32_t/(u)int64_t types,
irrespective of WM_LABEL_SIZE
- replace rarely used vector2D sum reduction with FixedList as a
indicator of its intent and also generalizes to different lengths.
OLD:
vector2D values; values.x() = ...; values.y() = ...;
reduce(values, sumOp<vector2D>());
NEW:
FixedList<scalar,2> values; values[0] = ...; values[1] = ...;
reduce(values, sumOp<scalar>());
- allow returnReduce() to use native reductions. Previous code (with
linear/tree selector) would have bypassed them inadvertently.
ENH: added support for MPI broadcast (for a memory span)
ENH: select communication schedule as a static method
- UPstream::whichCommunication(comm) to select linear/tree
communication instead of ternary or
if (Pstream::nProcs() < Pstream::nProcsSimpleSum) ...
STYLE: align nProcsSimpleSum static value with etc/controlDict override
- simplifies local toggling.
- centralize fileModification static variables into IOobject.
They were previously scattered between IOobject and regIOobject
- support selective enable/disable of the file banner.
ENH: improve code isolation for decomposedBlockData
- use readBlockEntry/writeBlockEntry to encapsulate the IO handling,
which ensures more consistency
- new decomposedBlockData::readHeader for chaining into the
block header information.
- remove unused constructors for decomposedBlockData
ENH: minor cleanup of collated fileOperations
- improves interface and data consistency.
Older signatures are still active (via the Foam_IOstream_extras
define).
- refine internals for IOstreamOption streamFormat, versionNumber
ENH: improve data alignment for IOstream and IOobject
- fit sizeof label/scalar into unsigned char
STYLE: remove dead code
- robuster matching behaviour when encountering paths that themselves
contain the word "processor" in them. For example,
"/path/processor0generation2/case1/processor10/system"
will now correctly match on processor10 instead of failing.
- use procRangeType for encapsulating the processor ranges
- provision for information of distributed vs non-distributed roots.
The information is currently available from the initial setup, but
can useful to access directly within fileOperation.
STYLE: modernize list iteration
- improves flexibility. Can tag a tmp as allowing non-const access to
the reference and skip additional const_cast in following code. For
example,
tmp<volScalarField> tfld(nullptr);
auto* ptr = getObjectPtr<volScalarField>("field");
if (ptr)
{
tfld.ref(*ptr);
}
else
{
tfld.reset(volScalarField::New(...));
}
auto& fld = tfld.ref();
ENH: renamed tmpNrc to refPtr
- the name 'refPtr' (reference|pointer) should be easier to remember
than tmpNrc (tmp, but non-ref-counted).
- provide tmpNrc typedef and header for code compatibility
NOTE
- in some places refPtr and tmp can be used instead of a
std::reference_wrapper for handling external references.
Unlike std::reference_wrapper, it can be default constructed
(holding nothing), whereas reference_wrapper may need a dummy
reference. However, the lifetime extension of references _may_ be
better with reference_wrapper.
- clearer than passing a reference to a dummy variable,
or relying on move occuring within the copy constructor
(historical, but should be deprecated)
STYLE: consistent autoPtr syntax for uncollated file operations
Improvements to existing functionality
--------------------------------------
- MPI is initialised without thread support if it is not needed e.g. uncollated
- Use native c++11 threading; avoids problem with static destruction order.
- etc/cellModels now only read if needed.
- etc/controlDict can now be read from the environment variable FOAM_CONTROLDICT
- Uniform files (e.g. '0/uniform/time') are now read only once on the master only
(with the masterUncollated or collated file handlers)
- collated format writes to 'processorsNNN' instead of 'processors'. The file
format is unchanged.
- Thread buffer and file buffer size are no longer limited to 2Gb.
The global controlDict file contains parameters for file handling. Under some
circumstances, e.g. running in parallel on a system without NFS, the user may
need to set some parameters, e.g. fileHandler, before the global controlDict
file is read from file. To support this, OpenFOAM now allows the global
controlDict to be read as a string set to the FOAM_CONTROLDICT environment
variable.
The FOAM_CONTROLDICT environment variable can be set to the content the global
controlDict file, e.g. from a sh/bash shell:
export FOAM_CONTROLDICT=$(foamDictionary $FOAM_ETC/controlDict)
FOAM_CONTROLDICT can then be passed to mpirun using the -x option, e.g.:
mpirun -np 2 -x FOAM_CONTROLDICT simpleFoam -parallel
Note that while this avoids the need for NFS to read the OpenFOAM configuration
the executable still needs to load shared libraries which must either be copied
locally or available via NFS or equivalent.
New: Multiple IO ranks
----------------------
The masterUncollated and collated fileHandlers can now use multiple ranks for
writing e.g.:
mpirun -np 6 simpleFoam -parallel -ioRanks '(0 3)'
In this example ranks 0 ('processor0') and 3 ('processor3') now handle all the
I/O. Rank 0 handles 0,1,2 and rank 3 handles 3,4,5. The set of IO ranks should always
include 0 as first element and be sorted in increasing order.
The collated fileHandler uses the directory naming processorsNNN_XXX-YYY where
NNN is the total number of processors and XXX and YYY are first and last
processor in the rank, e.g. in above example the directories would be
processors6_0-2
processors6_3-5
and each of the collated files in these contains data of the local ranks
only. The same naming also applies when e.g. running decomposePar:
decomposePar -fileHandler collated -ioRanks '(0 3)'
New: Distributed data
---------------------
The individual root directories can be placed on different hosts with different
paths if necessary. In the current framework it is necessary to specify the
root per slave process but this has been simplified with the option of specifying
the root per host with the -hostRoots command line option:
mpirun -np 6 simpleFoam -parallel -ioRanks '(0 3)' \
-hostRoots '("machineA" "/tmp/" "machineB" "/tmp")'
The hostRoots option is followed by a list of machine name + root directory, the
machine name can contain regular expressions.
New: hostCollated
-----------------
The new hostCollated fileHandler automatically sets the 'ioRanks' according to
the host name with the lowest rank e.g. to run simpleFoam on 6 processors with
ranks 0-2 on machineA and ranks 3-5 on machineB with the machines specified in
the hostfile:
mpirun -np 6 --hostfile hostfile simpleFoam -parallel -fileHandler hostCollated
This is equivalent to
mpirun -np 6 --hostfile hostfile simpleFoam -parallel -fileHandler collated -ioRanks '(0 3)'
This example will write directories:
processors6_0-2/
processors6_3-5/
A typical example would use distributed data e.g. no two nodes, machineA and
machineB, each with three processes:
decomposePar -fileHandler collated -case cavity
# Copy case (constant/*, system/*, processors6/) to master:
rsync -a cavity machineA:/tmp/
# Create root on slave:
ssh machineB mkdir -p /tmp/cavity
# Run
mpirun --hostfile hostfile icoFoam \
-case /tmp/cavity -parallel -fileHandler hostCollated \
-hostRoots '("machineA" "/tmp" "machineB" "/tmp")'
Contributed by Mattijs Janssens
Original commit message:
------------------------
Parallel IO: New collated file format
When an OpenFOAM simulation runs in parallel, the data for decomposed fields and
mesh(es) has historically been stored in multiple files within separate
directories for each processor. Processor directories are named 'processorN',
where N is the processor number.
This commit introduces an alternative "collated" file format where the data for
each decomposed field (and mesh) is collated into a single file, which is
written and read on the master processor. The files are stored in a single
directory named 'processors'.
The new format produces significantly fewer files - one per field, instead of N
per field. For large parallel cases, this avoids the restriction on the number
of open files imposed by the operating system limits.
The file writing can be threaded allowing the simulation to continue running
while the data is being written to file. NFS (Network File System) is not
needed when using the the collated format and additionally, there is an option
to run without NFS with the original uncollated approach, known as
"masterUncollated".
The controls for the file handling are in the OptimisationSwitches of
etc/controlDict:
OptimisationSwitches
{
...
//- Parallel IO file handler
// uncollated (default), collated or masterUncollated
fileHandler uncollated;
//- collated: thread buffer size for queued file writes.
// If set to 0 or not sufficient for the file size threading is not used.
// Default: 2e9
maxThreadFileBufferSize 2e9;
//- masterUncollated: non-blocking buffer size.
// If the file exceeds this buffer size scheduled transfer is used.
// Default: 2e9
maxMasterFileBufferSize 2e9;
}
When using the collated file handling, memory is allocated for the data in the
thread. maxThreadFileBufferSize sets the maximum size of memory in bytes that
is allocated. If the data exceeds this size, the write does not use threading.
When using the masterUncollated file handling, non-blocking MPI communication
requires a sufficiently large memory buffer on the master node.
maxMasterFileBufferSize sets the maximum size in bytes of the buffer. If the
data exceeds this size, the system uses scheduled communication.
The installation defaults for the fileHandler choice, maxThreadFileBufferSize
and maxMasterFileBufferSize (set in etc/controlDict) can be over-ridden within
the case controlDict file, like other parameters. Additionally the fileHandler
can be set by:
- the "-fileHandler" command line argument;
- a FOAM_FILEHANDLER environment variable.
A foamFormatConvert utility allows users to convert files between the collated
and uncollated formats, e.g.
mpirun -np 2 foamFormatConvert -parallel -fileHandler uncollated
An example case demonstrating the file handling methods is provided in:
$FOAM_TUTORIALS/IO/fileHandling
The work was undertaken by Mattijs Janssens, in collaboration with Henry Weller.
