- gather/scatter types of operations can avoid AllToAll communication
and use simple MPI gather (or scatter) to establish the receive sizes.
New methods: finishedGathers() / finishedScatters()
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.
- both schemes and solutions data are treated as MUST_READ_IF_MODIFIED
even if the requested readOption is nominally MUST_READ or
READ_IF_PRESENT, but now delay this change.
- do not need contruct or move assign from SortableList.
Rarely (never) used and can simply treat like a normal list
by applying shrink beforehand.
- make append() methods return void instead of returning self, which
makes it easier to derive from. Having them return self was a bit of
an original design mistake.
Chaining appends do not actually occur anywhere. Even if they were
to be used, would not want to rely on them (fear of slicing on any
derived classes).
BUG: IndirectList iterator comparison loses constness
- eliminate redundant size_ accounting
- drop extra 'Container' template parameter and replace functionality
with more flexible pack/unpack methods.
There is also a pack() method that handles indirect lists of lists
that can be used, for example, to pack a patch slice of faces.
Drop the 'operator()' method in favour of unpack to expose and properly
document the conversion. Should revisit the corresponding code in
some places for optimization potential.
- align some method names with globalIndex:
totalSize(), maxSize() etc
- less communication than gatherList/scatterList
ENH: refine send granularity in Pstream::exchange
STYLE: ensure PstreamBuffers and defaultCommsType agree
- simpler loops for lduSchedule
- can restrict calculation of D32 and other spray properties to a
subset of parcels. Uses a predicate selection mechanism similar to
vtkCloud etc.
ENH: code cleanup in scalar predicates
- pass by value not reference in predicates
- additional assign() method to refactor common code
- with the special setFormat "probes", all of the sampled sets are
treated more similarly to probes, with an ensemble output to raw
probed format.
This is of course less useful when the number of sampled points
becomes very large.
- in v2112 the functionObject results were only delivering values from
the last set listed (ie, overwritten).
Now that the values are properly scoped by the name of the set itself
Eg, `average(lines,p)` for the average for 'lines' set, existing
workflows will break.
It thus makes reasonble sense to also handle results without a
qualifier as ensemble values.
average(p) // Ensemble average of all listed sets
- the very old 'writer' class was fully stateless and always templated
on an particular output type.
This is now replaced with a 'coordSetWriter' with similar concepts
as previously introduced for surface writers (#1206).
- writers change from being a generic state-less set of routines to
more properly conforming to the normal notion of a writer.
- Parallel data is done *outside* of the writers, since they are used
in a wide variety of contexts and the caller is currently still in
a better position for deciding how to combine parallel data.
ENH: update sampleSets to sample on per-field basis (#2347)
- sample/write a field in a single step.
- support for 'sampleOnExecute' to obtain values at execution
intervals without writing.
- support 'sets' input as a dictionary entry (as well as a list),
which is similar to the changes for sampled-surface and permits use
of changeDictionary to modify content.
- globalIndex for gather to reduce parallel communication, less code
- qualify the sampleSet results (properties) with the name of the set.
The sample results were previously without a qualifier, which meant
that only the last property value was actually saved (previous ones
overwritten).
For example,
```
sample1
{
scalar
{
average(line,T) 349.96521;
min(line,T) 349.9544281;
max(line,T) 350;
average(cells,T) 349.9854619;
min(cells,T) 349.6589286;
max(cells,T) 350.4967271;
average(line,epsilon) 0.04947733869;
min(line,epsilon) 0.04449639927;
max(line,epsilon) 0.06452856475;
}
label
{
size(line,T) 79;
size(cells,T) 1720;
size(line,epsilon) 79;
}
}
```
ENH: update particleTracks application
- use globalIndex to manage original parcel addressing and
for gathering. Simplify code by introducing a helper class,
storing intermediate fields in hash tables instead of
separate lists.
ADDITIONAL NOTES:
- the regionSizeDistribution largely retains separate writers since
the utility of placing sum/dev/count for all fields into a single file
is questionable.
- the streamline writing remains a "soft" upgrade, which means that
scalar and vector fields are still collected a priori and not
on-the-fly. This is due to how the streamline infrastructure is
currently handled (should be upgraded in the future).
Automatic hole closure:
- introduces 'holeToFace' topoSet source
- used when detecting a 'leak-path'
- creates additional baffles to close the leak
Multi-stage layer addition:
- Can add layers in multiple passes
See issues: #2403, #2404
- for metis-like graphs there is no guarantee that a zero-sized graph
has an offsets list with size 1 or size 0, so always use
numCells = max(0, xadj.size()-1)
this was already done in most places, but missed in the
decomposeGeneral method
STYLE: use sumOp<label>() instead of plusOp<label>()
- the internal data are contiguous so can broadcast size and internals
directly without an intermediate stream.
ENH: split out broadcast time for profilingPstream information
STYLE: minor Pstream cleanup
- UPstream::commsType_ from protected to private, since it already has
inlined noexcept getters/setters that should be used.
- don't pass unused/unneed tag into low-level MPI reduction templates.
Document where tags are not needed
- had Pstream::broadcast instead of UPstream::broadcast in internals
- used Pstream::maxCommsSize (bytes) for the lower limit when sending.
This would have send more data on each iteration than expected based
on maxCommsSize and finish with a number of useless iterations.
Was generally not a serious bug since maxCommsSize (if used) was
likely still far away from the MPI limits and exchange() is primarily
harnessed by PstreamBuffers, which is sending character data
(ie, number of elements and number of bytes is identical).
- For v2112 and earlier: pre-assembled lists of particles
to be transferred and target patch on a per processor basis.
Apart from memory overhead of assembling the lists this adds
allocations/de-allocation when building linked-lists.
- Now stream particle transfer tuples directly into PstreamBuffers.
Use a local cache of UOPstream wrappers for the formatters
(since there are potentially many particles being shifted about).
On the receiving size, read out tuple-wise.
- Communication on transfers now restricted to the immediate
neighbours instead of using an all-to-all to exchange sizes.
Applied to Cloud::move and RecycleInteraction
- now largely encapsulated using PstreamBuffers methods,
which makes it simpler to centralize and maintain
- avoid building intermediate structures when sending data,
remove unused methods/data
TUT: parallel version of depthCharge2D
STYLE: minor update in ProcessorTopology
- PstreamBuffers nProcs() and allProcs() methods to recover the rank
information consistent with the communicator used for construction
- allowClearRecv() methods for more control over buffer reuse
For example,
pBufs.allowClearRecv(false);
forAll(particles, particlei)
{
pBufs.clear();
fill...
read via IPstream(..., pBufs);
}
This preserves the receive buffers memory allocation between calls.
- finishedNeighbourSends() method as compact wrapper for
finishedSends() when send/recv ranks are identically
(eg, neighbours)
- hasSendData()/hasRecvData() methods for PstreamBuffers.
Can be useful for some situations to skip reading entirely.
For example,
pBufs.finishedNeighbourSends(neighProcs);
if (!returnReduce(pBufs.hasRecvData(), orOp<bool>()))
{
// Nothing to do
continue;
}
...
On an individual basis:
for (const int proci : pBufs.allProcs())
{
if (pBufs.hasRecvData(proci))
{
...
}
}
Also conceivable to do the following instead (nonBlocking only):
if (!returnReduce(pBufs.hasSendData(), orOp<bool>()))
{
// Nothing to do
pBufs.clear();
continue;
}
pBufs.finishedNeighbourSends(neighProcs);
...
- a somewhat specialized use case, but can be useful when there are
many ranks with sparse communication but for which the access
pattern is established during inner loops.
PstreamBuffers pBufs(Pstream::commsTypes::nonBlocking);
pBufs.allowClearRecv(false);
PtrList<OPstream> output(Pstream::nProcs());
while (condition)
{
// Rewind existing streams
forAll(output, proci)
{
auto* osptr = output.get(proci);
if (osptr)
{
(*osptr).rewind();
}
}
for (Particle& p : myCloud)
{
label toProci = ...;
// Get or create output stream
auto* osptr = output.get(toProci);
if (!osptr)
{
osptr = new OPstream(toProci, pBufs);
output.set(toProci, osptr);
}
// Append more data...
(*osptr) << p;
}
pBufs.finishedSends();
... reads
}
- 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
- reduces later surprises and simplifies effort for the caller
- more flexible globalIndex scatter with auto-sized return field.
- Avoid communication for scattering into zero-sized fields.
- the data front for isoAdvection can be particularly sparse and at
higher processor counts there is an advantage to avoiding all-to-all
communication for the PstreamBuffers exchange
Based on code changes from T.Aoyagi(RIST), A.Azami(RIST)
- use MPI_Bcast intrinsic instead of manual tree to reduce the overall
number of messages.
Old behaviour can be re-enabled with
`#define Foam_Pstream_scatter_nobroadcast`
- The idea of broadcast streams is to replace multiple master to
subProcs communications with a single MPI_Bcast.
if (Pstream::master())
{
OPBstream toAll(Pstream::masterNo());
toAll << data;
}
else
{
IPBstream fromMaster(Pstream::masterNo());
fromMaster >> data;
}
// vs.
if (Pstream::master())
{
for (const int proci : Pstream::subProcs())
{
OPstream os(Pstream::commsTypes::scheduled, proci);
os << data;
}
}
else
{
IPstream is(Pstream::commsTypes::scheduled, Pstream::masterNo());
is >> data;
}
Can simply use UPstream::broadcast() directly for contiguous data
with known lengths.
Based on ideas from T.Aoyagi(RIST), A.Azami(RIST)
- 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
- refactor as an MPI-independent base class.
Add bufferIPC{send,recv} private methods for construct/destruct.
Eliminates code duplication from two constructor forms and reduces
additional constructor definitions in dummy library.
- add PstreamBuffers access methods, refactor common finish sends
code, tweak member packing
ENH: resize_nocopy for processorLduInterface buffers
- content is immediately overwritten
STYLE: cull unneeded includes in processorFa*
- handled by processorLduInterface
- this can be used to apply a uniform field level to remove from
a sampled field. For example,
fieldLevel
{
"p.*" 1e5; // Absolute -> gauge [Pa]
T 273.15; // [K] -> [C]
U #eval{ 10/sqrt(3) }; // Uniform mag(U)=10
}
After the fieldLevel has been removed, any fieldScale is applied.
For example
fieldScale
{
"p.*" 0.01; // [Pa] -> [mbar]
}
The fieldLevel for vector and tensor fields may still need some
further refinement.
