Newer
Older
Christian Godenschwager
committed
//======================================================================================================================
//
// This file is part of waLBerla. waLBerla is free software: you can
// redistribute it and/or modify it under the terms of the GNU General Public
// License as published by the Free Software Foundation, either version 3 of
// the License, or (at your option) any later version.
//
// waLBerla is distributed in the hope that it will be useful, but WITHOUT
// ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
// for more details.
//
// You should have received a copy of the GNU General Public License along
// with waLBerla (see COPYING.txt). If not, see <http://www.gnu.org/licenses/>.
//
//! \file Initialization.cpp
//! \ingroup blockforest
//! \author Florian Schornbaum <florian.schornbaum@fau.de>
//
//======================================================================================================================
#include "BlockNeighborhoodSection.h"
#include "Initialization.h"
#include "SetupBlockForest.h"
#include "loadbalancing/Cartesian.h"
#include "core/Abort.h"
#include "core/cell/CellInterval.h"
#include "core/math/IntegerFactorization.h"
#include "core/mpi/MPIManager.h"
#include "stencil/D3Q19.h"
Christian Godenschwager
committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
namespace walberla {
namespace blockforest {
//**********************************************************************************************************************
/*!
* Parses config block called 'DomainSetup' and creates a StructuredBlockForest
*
* For more information see function below.
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest > createUniformBlockGridFromConfig( const shared_ptr< Config > & config,
CellInterval * requestedDomainSize,
const bool keepGlobalBlockInformation )
{
if( !!config )
{
auto block = config->getGlobalBlock();
if( block ) {
auto subBlock = block.getBlock( "DomainSetup" );
if ( ! subBlock ) {
WALBERLA_ABORT_NO_DEBUG_INFO( "Unable to create uniform block grid from configuration file."
"\nDid you forget to specify a \"DomainSetup\" block in the configuration file?" );
}
return createUniformBlockGridFromConfig( subBlock, requestedDomainSize, keepGlobalBlockInformation );
}
}
WALBERLA_ABORT_NO_DEBUG_INFO( "No Configuration specified" );
return shared_ptr<StructuredBlockForest>();
}
//**********************************************************************************************************************
/*!
* \brief Parses config block and creates a StructuredBlockForest
*
* Two possibilities:
* 1) Using the cells per block and number of blocks for each direction
\verbatim
{
cellsPerBlock < 10, 20, 30 > ; // required
blocks < 1, 2, 3 > ; // required
periodic < 0, 0, 1 >; // not required, defaults to no periodicity
dx 0.01; // defaults to 1.0
}
\endverbatim
* An optional config parameter 'cartesianSetup' is available. Its default, true, causes one block to be
* assigned to each process. Setting it to false allows multiple blocks to be assigned to each process.
* If the number of blocks is not divisble by the number of processes, the loadbalancer tries to assign
* the blocks to processes as evenly as possible.
Christian Godenschwager
committed
* 2) Using the number of global cells, #blocks = #processes, if this does not fit, extend the domain
\verbatim
{
cells < 10,40,90>; // required
periodic < 0, 0, 1 >; // not required, defaults to no periodicity
dx 0.01; // defaults to 1.0
}
\endverbatim
* An optional config parameter 'oneBlockPerProcess' is available. Setting it to false forces all
* blocks to be assigned to a single process, which may be useful for debugging purposes. Otherwise,
* one block is assigned to each process.
Christian Godenschwager
committed
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
* Example: cells < 31,31,31> started using 8 processors <BR>
* calculated processor distribution <2,2,2> <BR>
* real domain is then extended to <32,32,32> and every processor gets a block of <16,16,16>
*
* When this setup is used and requestedDomainSize is not the null pointer, it is set to <BR>
* the requested domain size ( in the example <31,31,31> )
*
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest > createUniformBlockGridFromConfig( const Config::BlockHandle & configBlock,
CellInterval * requestedDomainSize,
const bool keepGlobalBlockInformation )
{
const Vector3<bool> periodic = configBlock.getParameter<Vector3<bool> >( "periodic", Vector3<bool> (false) );
const real_t dx = configBlock.getParameter<real_t >( "dx", real_t(1) );
Vector3<uint_t> cellsPerBlock;
Vector3<uint_t> blocks;
if ( configBlock.isDefined("cells") )
{
if ( configBlock.isDefined("cellsPerBlock") || configBlock.isDefined("blocks") )
WALBERLA_ABORT_NO_DEBUG_INFO("Config Error: Use either ('cellsPerBlock' and 'blocks') or 'cells', not both!");
Vector3<uint_t> cells = configBlock.getParameter<Vector3<uint_t> >( "cells" );
if ( requestedDomainSize )
*requestedDomainSize = CellInterval( 0,0,0,
cell_idx_c(cells[0])-1,
cell_idx_c(cells[1])-1,
cell_idx_c(cells[2])-1 );
uint_t nrOfProcesses = uint_c( MPIManager::instance()->numProcesses() );
calculateCellDistribution( cells, nrOfProcesses, blocks, cellsPerBlock );
}
else
{
cellsPerBlock = configBlock.getParameter<Vector3<uint_t> >( "cellsPerBlock" );
blocks = configBlock.getParameter<Vector3<uint_t> >( "blocks" );
if ( requestedDomainSize )
*requestedDomainSize = CellInterval( 0, 0, 0, cell_idx_c( cellsPerBlock[0] * blocks[0] ),
cell_idx_c( cellsPerBlock[1] * blocks[1] ),
cell_idx_c( cellsPerBlock[2] * blocks[2] ) );
}
const bool oneBlockPerProcess = configBlock.getParameter<bool> ( "oneBlockPerProcess", true );
const bool cartesian = configBlock.getParameter<bool> ( "cartesianSetup", true );
if ( !cartesian )
{
if ( configBlock.isDefined("oneBlockPerProcess") )
WALBERLA_ABORT_NO_DEBUG_INFO("Config Error: Set either 'oneBlockPerProcess' or set 'cartesianSetup' to false, not both!");
return createUniformBlockGrid(
blocks[0], blocks[1], blocks[2], // blocks in x/y/z direction
cellsPerBlock[0], cellsPerBlock[1], cellsPerBlock[2], // cells per block in x/y/z direction
dx, // cell size
uint_t(0), // maximum number of blocks per process
true, false, // include but don't force Metis
periodic[0], periodic[1], periodic[2], // periodicity
keepGlobalBlockInformation // keep global block information
);
}
Christian Godenschwager
committed
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
return createUniformBlockGrid(
blocks[0], blocks[1], blocks[2], // blocks/processes in x/y/z direction
cellsPerBlock[0], cellsPerBlock[1], cellsPerBlock[2], // cells per block in x/y/z direction
dx, // cell size
oneBlockPerProcess, // one block per process
periodic[0], periodic[1], periodic[2], // periodicity
keepGlobalBlockInformation // keep global block information
);
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks, each block has the same size.
* The distribution of blocks to processes also follows a Cartesian decomposition.
*
* \param domainAABB An axis-aligned bounding box that spans the entire simulation space/domain
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXProcesses Number of processes the blocks are distributed to in x direction
* \param numberOfYProcesses Number of processes the blocks are distributed to in y direction
* \param numberOfZProcesses Number of processes the blocks are distributed to in z direction
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< BlockForest >
createBlockForest( const AABB& domainAABB,
const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXProcesses, const uint_t numberOfYProcesses, const uint_t numberOfZProcesses,
const bool xPeriodic /* = false */, const bool yPeriodic /* = false */, const bool zPeriodic /* = false */,
const bool keepGlobalBlockInformation /* = false */ ) {
const uint_t numberOfProcesses = numberOfXProcesses * numberOfYProcesses * numberOfZProcesses;
if( numeric_cast< int >( numberOfProcesses ) != MPIManager::instance()->numProcesses() )
WALBERLA_ABORT( "The number of requested processes (" << numberOfProcesses << ") doesn't match the number "
"of active MPI processes (" << MPIManager::instance()->numProcesses() << ")!" );
// initialize SetupBlockForest = determine domain decomposition
SetupBlockForest sforest;
sforest.addWorkloadMemorySUIDAssignmentFunction( uniformWorkloadAndMemoryAssignment );
sforest.init( domainAABB, numberOfXBlocks, numberOfYBlocks, numberOfZBlocks, xPeriodic, yPeriodic, zPeriodic );
// if possible, create Cartesian MPI communicator
std::vector< uint_t >* processIdMap = nullptr;
Christian Godenschwager
committed
WALBERLA_MPI_SECTION()
{
auto mpiManager = MPIManager::instance();
//create cartesian communicator only if not yet a cartesian communicator (or other communicator was created)
if ( ! mpiManager->rankValid() )
{
if ( mpiManager->isCartesianCommValid() ) {
mpiManager->createCartesianComm( numberOfXProcesses, numberOfYProcesses, numberOfZProcesses, xPeriodic, yPeriodic, zPeriodic );
Christian Godenschwager
committed
processIdMap = new std::vector< uint_t >( numberOfProcesses );
Christian Godenschwager
committed
for( uint_t z = 0; z != numberOfZProcesses; ++z ) {
for( uint_t y = 0; y != numberOfYProcesses; ++y ) {
for( uint_t x = 0; x != numberOfXProcesses; ++x ) {
Christian Godenschwager
committed
(*processIdMap)[ z * numberOfXProcesses * numberOfYProcesses + y * numberOfXProcesses + x ] =
uint_c( MPIManager::instance()->cartesianRank(x,y,z) );
}
Christian Godenschwager
committed
}
}
}
else {
WALBERLA_LOG_WARNING_ON_ROOT( "Your version of OpenMPI contains a bug. See waLBerla issue #73 for more "
"information. As a workaround, MPI_COMM_WORLD instead of a "
"Cartesian MPI communicator is used." );
mpiManager->useWorldComm();
}
Christian Godenschwager
committed
}
}
// calculate process distribution
sforest.balanceLoad( blockforest::CartesianDistribution( numberOfXProcesses, numberOfYProcesses, numberOfZProcesses, processIdMap ),
numberOfXProcesses * numberOfYProcesses * numberOfZProcesses );
if( processIdMap != nullptr ) delete processIdMap;
Christian Godenschwager
committed
// create StructuredBlockForest (encapsulates a newly created BlockForest)
return std::make_shared< BlockForest >( uint_c( MPIManager::instance()->rank() ), sforest, keepGlobalBlockInformation );
Christian Godenschwager
committed
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* The distribution of blocks to processes also follows a Cartesian decomposition.
*
* \param domainAABB An axis-aligned bounding box that spans the entire simulation space/domain
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param numberOfXProcesses Number of processes the blocks are distributed to in x direction
* \param numberOfYProcesses Number of processes the blocks are distributed to in y direction
* \param numberOfZProcesses Number of processes the blocks are distributed to in z direction
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const AABB& domainAABB,
const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const uint_t numberOfXProcesses, const uint_t numberOfYProcesses, const uint_t numberOfZProcesses,
const bool xPeriodic /* = false */, const bool yPeriodic /* = false */, const bool zPeriodic /* = false */,
const bool keepGlobalBlockInformation /* = false */ )
{
auto bf = createBlockForest(
domainAABB,
numberOfXBlocks,
numberOfYBlocks,
numberOfZBlocks,
numberOfXProcesses,
numberOfYProcesses,
numberOfZProcesses,
xPeriodic,
yPeriodic,
zPeriodic,
keepGlobalBlockInformation);
auto sbf = std::make_shared< StructuredBlockForest >( bf, numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock );
Christian Godenschwager
committed
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
sbf->createCellBoundingBoxes();
return sbf;
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* The distribution of blocks to processes also follows a Cartesian decomposition.
*
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param dx Edge length of each cell (cells are assumed to be cubes)
* \param numberOfXProcesses Number of processes the blocks are distributed to in x direction
* \param numberOfYProcesses Number of processes the blocks are distributed to in y direction
* \param numberOfZProcesses Number of processes the blocks are distributed to in z direction
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const real_t dx,
const uint_t numberOfXProcesses, const uint_t numberOfYProcesses, const uint_t numberOfZProcesses,
const bool xPeriodic /* = false */, const bool yPeriodic /* = false */, const bool zPeriodic /* = false */,
const bool keepGlobalBlockInformation /* = false */ ) {
return createUniformBlockGrid( AABB( real_c(0), real_c(0), real_c(0), dx * real_c( numberOfXBlocks * numberOfXCellsPerBlock ),
dx * real_c( numberOfYBlocks * numberOfYCellsPerBlock ),
dx * real_c( numberOfZBlocks * numberOfZCellsPerBlock ) ),
numberOfXBlocks, numberOfYBlocks, numberOfZBlocks,
numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock,
numberOfXProcesses, numberOfYProcesses, numberOfZProcesses,
xPeriodic, yPeriodic, zPeriodic, keepGlobalBlockInformation );
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* Either all blocks are assigned to the same process (useful for non-parallel simulations) or each blocks is assigned
* to a different process (useful if only one block shall be assigned to each process).
*
* \param domainAABB An axis-aligned bounding box that spans the entire simulation space/domain
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param oneBlockPerProcess If true, each block is assigned to a different process. If false, all blocks are
* assigned to the same process (process 0).
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const AABB& domainAABB,
const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const bool oneBlockPerProcess,
const bool xPeriodic /* = false */, const bool yPeriodic /* = false */, const bool zPeriodic /* = false */,
const bool keepGlobalBlockInformation /* = false */ ) {
if( oneBlockPerProcess )
return createUniformBlockGrid( domainAABB, numberOfXBlocks, numberOfYBlocks, numberOfZBlocks,
numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock,
numberOfXBlocks, numberOfYBlocks, numberOfZBlocks, xPeriodic, yPeriodic, zPeriodic, keepGlobalBlockInformation );
// all blocks on the same process
return createUniformBlockGrid( domainAABB, numberOfXBlocks, numberOfYBlocks, numberOfZBlocks,
numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock, uint_c(1), uint_c(1), uint_c(1),
xPeriodic, yPeriodic, zPeriodic, keepGlobalBlockInformation );
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* Either all blocks are assigned to the same process (useful for non-parallel simulations) or each blocks is assigned
* to a different process (useful if only one block shall be assigned to each process).
*
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param dx Edge length of each cell (cells are assumed to be cubes)
* \param oneBlockPerProcess If true, each block is assigned to a different process. If false, all blocks are
* assigned to the same process (process 0).
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const real_t dx,
const bool oneBlockPerProcess,
const bool xPeriodic /* = false */, const bool yPeriodic /* = false */, const bool zPeriodic /* = false */,
const bool keepGlobalBlockInformation /* = false */ ) {
return createUniformBlockGrid( AABB( real_c(0), real_c(0), real_c(0), dx * real_c( numberOfXBlocks * numberOfXCellsPerBlock ),
dx * real_c( numberOfYBlocks * numberOfYCellsPerBlock ),
dx * real_c( numberOfZBlocks * numberOfZCellsPerBlock ) ),
numberOfXBlocks, numberOfYBlocks, numberOfZBlocks,
numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock,
oneBlockPerProcess, xPeriodic, yPeriodic, zPeriodic, keepGlobalBlockInformation );
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* The number of active MPI processes is used in order to determine the process distribution = in order to perform the
* initial, static load balancing. Each block is assumed to generate the same amount of work and to require the same
* amount of memory.
*
* \param domainAABB An axis-aligned bounding box that spans the entire simulation space/domain
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param maxBlocksPerProcess Maximum number of blocks that are allowed to be assigned to one process. If a
* value of '0' is provided, any number of blocks are allowed to be located on one
* process - meaning static load balancing doesn't try to obey any memory limit. ['0' by default]
* \param includeMetis If true (and if available!), METIS is also used during load balancing. [true by default]
* \param forceMetis If true, METIS is always preferred over space filling curves [false by default]
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const AABB& domainAABB,
const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const uint_t maxBlocksPerProcess /*= 0*/, const bool includeMetis /*= true*/, const bool forceMetis /*= false*/,
const bool xPeriodic /*= false*/, const bool yPeriodic /*= false*/, const bool zPeriodic /*= false*/,
const bool keepGlobalBlockInformation /*= false*/ ) {
// initialize SetupBlockForest = determine domain decomposition
SetupBlockForest sforest;
sforest.addWorkloadMemorySUIDAssignmentFunction( uniformWorkloadAndMemoryAssignment );
sforest.init( domainAABB, numberOfXBlocks, numberOfYBlocks, numberOfZBlocks, xPeriodic, yPeriodic, zPeriodic );
// calculate process distribution
const memory_t memoryLimit = ( maxBlocksPerProcess == 0 ) ? numeric_cast< memory_t >( sforest.getNumberOfBlocks() ) :
numeric_cast< memory_t >( maxBlocksPerProcess );
GlobalLoadBalancing::MetisConfiguration< SetupBlock > metisConfig( includeMetis, forceMetis,
std::bind( cellWeightedCommunicationCost, std::placeholders::_1, std::placeholders::_2,
Christian Godenschwager
committed
numberOfXCellsPerBlock,
numberOfYCellsPerBlock,
numberOfZCellsPerBlock ) );
sforest.calculateProcessDistribution_Default( uint_c( MPIManager::instance()->numProcesses() ), memoryLimit, "hilbert", 10, false, metisConfig );
if( !MPIManager::instance()->rankValid() )
MPIManager::instance()->useWorldComm();
// create StructuredBlockForest (encapsulates a newly created BlockForest)
auto bf = std::make_shared< BlockForest >( uint_c( MPIManager::instance()->rank() ), sforest, keepGlobalBlockInformation );
Christian Godenschwager
committed
auto sbf = std::make_shared< StructuredBlockForest >( bf, numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock );
Christian Godenschwager
committed
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
sbf->createCellBoundingBoxes();
return sbf;
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* The number of active MPI processes is used in order to determine the process distribution = in order to perform the
* initial, static load balancing. Each block is assumed to generate the same amount of work and to require the same
* amount of memory.
*
* \param numberOfXBlocks Number of blocks in x direction
* \param numberOfYBlocks Number of blocks in y direction
* \param numberOfZBlocks Number of blocks in z direction
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param dx Edge length of each cell (cells are assumed to be cubes)
* \param maxBlocksPerProcess Maximum number of blocks that are allowed to be assigned to one process. If a
* value of '0' is provided, any number of blocks are allowed to be located on one
* process - meaning static load balancing doesn't try to obey any memory limit. ['0' by default]
* \param includeMetis If true (and if available!), METIS is also used during load balancing. [true by default]
* \param forceMetis If true, METIS is always preferred over space filling curves [false by default]
* \param xPeriodic If true, the block structure is periodic in x direction [false by default]
* \param yPeriodic If true, the block structure is periodic in y direction [false by default]
* \param zPeriodic If true, the block structure is periodic in z direction [false by default]
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const uint_t numberOfXBlocks, const uint_t numberOfYBlocks, const uint_t numberOfZBlocks,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const real_t dx,
const uint_t maxBlocksPerProcess /*= 0*/, const bool includeMetis /*= true*/, const bool forceMetis /*= false*/,
const bool xPeriodic /*= false*/, const bool yPeriodic /*= false*/, const bool zPeriodic /*= false*/,
const bool keepGlobalBlockInformation /*= false*/ ) {
return createUniformBlockGrid( AABB( real_c(0), real_c(0), real_c(0), dx * real_c( numberOfXBlocks * numberOfXCellsPerBlock ),
dx * real_c( numberOfYBlocks * numberOfYCellsPerBlock ),
dx * real_c( numberOfZBlocks * numberOfZCellsPerBlock ) ),
numberOfXBlocks, numberOfYBlocks, numberOfZBlocks,
numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock,
maxBlocksPerProcess, includeMetis, forceMetis, xPeriodic, yPeriodic, zPeriodic, keepGlobalBlockInformation );
}
//**********************************************************************************************************************
/*!
* \brief Function for creating a structured block forest that represents a uniform block grid.
*
* Uniform block grid: Cartesian domain decomposition into blocks of cells, each block has the same size and contains
* the same number of cells.
* The entire block structure and its corresponding process distribution are loaded from file.
*
* \param filename A file that stores a block structure and its corresponding process distribution
* \param numberOfXCellsPerBlock Number of cells of each block in x direction
* \param numberOfYCellsPerBlock Number of cells of each block in y direction
* \param numberOfZCellsPerBlock Number of cells of each block in z direction
* \param keepGlobalBlockInformation If true, each process keeps information about remote blocks (blocks that reside
* on other processes). This information includes the process rank, the state, and
* the axis-aligned bounding box of any block (local or remote). [false by default]
*/
//**********************************************************************************************************************
shared_ptr< StructuredBlockForest >
createUniformBlockGrid( const std::string& filename,
const uint_t numberOfXCellsPerBlock, const uint_t numberOfYCellsPerBlock, const uint_t numberOfZCellsPerBlock,
const bool keepGlobalBlockInformation /*= false*/ )
{
if( !MPIManager::instance()->rankValid() )
MPIManager::instance()->useWorldComm();
auto bf = std::make_shared< BlockForest >( uint_c( MPIManager::instance()->rank() ), filename.c_str(), true, keepGlobalBlockInformation );
Christian Godenschwager
committed
if( !bf->storesUniformBlockGrid() )
WALBERLA_ABORT( "The block forest loaded from file \'" << filename << "\' does not contain a uniform block grid!" );
auto sbf = std::make_shared< StructuredBlockForest >( bf, numberOfXCellsPerBlock, numberOfYCellsPerBlock, numberOfZCellsPerBlock );
Christian Godenschwager
committed
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
sbf->createCellBoundingBoxes();
return sbf;
}
///////////////////////////////////////////
// HELPER FUNCTIONS //
///////////////////////////////////////////
//*******************************************************************************************************************
/*! Tries to distribute a given amount of total cells to a given amount of blocks
*
* It may happen that divisibility of the nr of cells requested prevents a distribution
* in this case the number of cells is chosen bigger than requested
*
*
* \param cells: total number of cells requested
* \param nrOfBlocks: total number of blocks to distribute the cells to
* \param[out] blocks: calculated number of blocks in x/y/z
* \param[out] cellsPerBlock: how many cells to put on each block
it may happen that divisibility of the number of cells requested prevents a distribution
* in this case the number of cells is chosen (slightly) bigger than requested
*
* Example: in: cells = (10,15,16)
* in: blocks = 8
* out: blocks = (2,2,2)
* out: cellsPerBlock = (5,8,8)
* out: newCells = (10,16,16)
*/
//*******************************************************************************************************************
void calculateCellDistribution( const Vector3<uint_t> & cells, uint_t nrOfBlocks,
Vector3<uint_t> & blocksOut, Vector3<uint_t> & cellsPerBlock)
{
std::vector< real_t > weighting;
weighting.push_back( real_c( cells[0]) );
weighting.push_back( real_c( cells[1]) );
weighting.push_back( real_c( cells[2]) );
std::vector<uint_t> blocks = math::getFactors( nrOfBlocks, 3, weighting );
for( uint_t i = 0; i < 3; ++i )
{
if ( uint_c( cells[i] ) % blocks[i] == 0 )
cellsPerBlock[i] = cells[i] / blocks[i];
else // extend the domain if processesCount does not divide the cell count in this direction
cellsPerBlock[i] = ( cells[i] + blocks[i] ) / blocks[i];
}
for( uint_t i = 0; i < 3; ++i )
blocksOut[i] = blocks[i];
}
void uniformWorkloadAndMemoryAssignment( SetupBlockForest& forest ) {
std::vector< SetupBlock* > blocks;
forest.getBlocks( blocks );
for( uint_t i = 0; i != blocks.size(); ++i ) {
blocks[i]->setWorkload( numeric_cast< workload_t >(1) );
blocks[i]->setMemory( numeric_cast< memory_t >(1) );
}
}
memory_t cellWeightedCommunicationCost( const SetupBlock* const a, const SetupBlock* const b,
uint_t xCellsPerBlock, uint_t yCellsPerBlock, uint_t zCellsPerBlock )
{
for ( auto dIter = stencil::D3Q19::beginNoCenter(); dIter != stencil::D3Q19::end(); ++dIter )
{
auto neighborHoodIdxA = getBlockNeighborhoodSectionIndex( *dIter );
for ( uint_t j = 0; j != a->getNeighborhoodSectionSize( neighborHoodIdxA ); ++j )
{
if( a->getNeighbor(neighborHoodIdxA,j) == b )
{
switch ( *dIter )
{
//faces
case stencil::W: return memory_c( yCellsPerBlock * zCellsPerBlock );
case stencil::E: return memory_c( yCellsPerBlock * zCellsPerBlock );
case stencil::N: return memory_c( xCellsPerBlock * zCellsPerBlock );
case stencil::S: return memory_c( xCellsPerBlock * zCellsPerBlock );
case stencil::T: return memory_c( xCellsPerBlock * yCellsPerBlock );
case stencil::B: return memory_c( xCellsPerBlock * yCellsPerBlock );
//edges
case stencil::NW: return memory_c( zCellsPerBlock );
case stencil::NE: return memory_c( zCellsPerBlock );
case stencil::SW: return memory_c( zCellsPerBlock );
case stencil::SE: return memory_c( zCellsPerBlock );
case stencil::TN: return memory_c( xCellsPerBlock );
case stencil::TS: return memory_c( xCellsPerBlock );
case stencil::TW: return memory_c( yCellsPerBlock );
case stencil::TE: return memory_c( yCellsPerBlock );
case stencil::BN: return memory_c( xCellsPerBlock );
case stencil::BS: return memory_c( xCellsPerBlock );
case stencil::BW: return memory_c( yCellsPerBlock );
case stencil::BE: return memory_c( yCellsPerBlock );
default:
WALBERLA_ABORT( "Unknown direction. Should not happen!" )
}
}
}
}
// Return 1 for corners
return numeric_cast< memory_t >(1);
}
memory_t uniformFacesDominantCommunication( const SetupBlock* const a, const SetupBlock* const b ) {
uint_t faces[] = { 4, 10, 12, 13, 15, 21 };
for( uint_t i = 0; i != 6; ++i ) {
for( uint_t j = 0; j != a->getNeighborhoodSectionSize(faces[i]); ++j )
if( a->getNeighbor(faces[i],j) == b )
return numeric_cast< memory_t >(1000);
}
return numeric_cast< memory_t >(1);
}
} // namespace blockforest
} // namespace walberla