Flow123d  JS_before_hm-1621-g63a12c7
sparse_graph.cc
Go to the documentation of this file.
1 /*!
2  *
3  * Copyright (C) 2015 Technical University of Liberec. All rights reserved.
4  *
5  * This program is free software; you can redistribute it and/or modify it under
6  * the terms of the GNU General Public License version 3 as published by the
7  * Free Software Foundation. (http://www.gnu.org/licenses/gpl-3.0.en.html)
8  *
9  * This program is distributed in the hope that it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
11  * FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
12  *
13  *
14  * @file sparse_graph.cc
15  * @ingroup la
16  * @brief Construction and partitioning of a sparse graph
17  */
18 
19 #include "system/global_defs.h"
20 #include "system/system.hh"
21 #include "la/distribution.hh"
22 #include "sparse_graph.hh"
23 
24 #include "boost/lambda/lambda.hpp"
25 #include <algorithm>
26 
27 extern "C" {
28 #include <metis.h>
29 }
30 
31 
32 /*****************************************************************************************
33  SPARSE GRAPH
34  *****************************************************************************************/
35 
36 
38  : vtx_distr(distr),
39  rows(NULL),
40  adj(NULL),
41  adj_weights(NULL),
42  part_to_check(NULL),
43  adj_of_proc( vtx_distr.np() )
44 {
45  // positions only of local vertexes
46  vtx_XYZ= new float[vtx_distr.lsize()+1];
47  vtx_weights= new int[vtx_distr.lsize()+1];
48  for(unsigned int i=0; i<vtx_distr.lsize(); i++) vtx_weights[i]=1;
49 }
50 
51 
52 
54  : vtx_distr(loc_size, comm),
55  rows(NULL),
56  adj(NULL),
57  adj_weights(NULL),
58  adj_of_proc( vtx_distr.np() )
59 {
60  // positions only of local vertexes
61  vtx_XYZ= new float[vtx_distr.lsize()+1];
62  vtx_weights= new int[vtx_distr.lsize()+1];
63  for(unsigned int i=0; i<vtx_distr.lsize(); i++) vtx_weights[i]=1;
64 }
65 
66 void SparseGraph::set_edge(const int a, const int b,int weight)
67 {
68  Edge e={a,b, weight};
69 
70  adj_of_proc[vtx_distr.get_proc(a)].push(e);
71 }
72 
73 
74 
75 
76 void SparseGraph::set_vtx_position(const int vtx, const float xyz[3],int weight)
77 {
78  OLD_ASSERT(vtx_distr.is_local(vtx),"Can not set vertex position for nonlocal vertex %d.\n",vtx);
79  int loc_index=vtx-vtx_distr.begin();
80  memcpy(vtx_XYZ+3*loc_index, xyz, 3*sizeof(float));
81  vtx_weights[loc_index]=weight;
82 }
83 
84 
85 /**
86  * Edge comparison. For lexical sorting of local edges.
87  */
88 
90 {
91  return (a.from < b.from ||
92  ( a.from == b.from && a.to < b.to));
93 }
94 
95 
96 // TODO: split into smaller functions
97 // check trivial case : Localized distribution
99 {
100  OLD_ASSERT( adj==NULL, "Graph is already finalized\n");
101 
102  unsigned int proc;
103  int total_size;
104  vector< stack<Edge> >::iterator s;
105  unsigned int edge_size=3; // 3 = number of integers in Edge to send
106 
107  /////////////////////////////////////
108  // communicate inserted edges
109 
110  // unroll data in adj_of_proc into send buffer
111  int * sdispls = new int [vtx_distr.np()];
112  int * scounts = new int [vtx_distr.np()];
113 
114  total_size=0;
115  for(proc=0, s=adj_of_proc.begin(); s!=adj_of_proc.end();++s, ++proc)
116  {
117  sdispls[proc] = total_size;
118  scounts[proc] = edge_size * (s)->size();
119  total_size += edge_size * ((s)->size)();
120  }
121  int * sendbuf = new int [(total_size+1)];
122 
123  Edge edge;
124  int buf_pos=0;
125  for(proc=0, s = adj_of_proc.begin(); s!=adj_of_proc.end(); ++s, ++proc)
126  {
127  OLD_ASSERT(sdispls[proc] == buf_pos,
128  "Mismatch between displacement %d and buffer position %d. \n", sdispls[proc], buf_pos );
129  while ( ! (s)->empty() ) {
130  edge=(s)->top();
131  sendbuf[buf_pos++]=edge.from;
132  sendbuf[buf_pos++]=edge.to;
133  sendbuf[buf_pos++]=edge.weight;
134  (s)->pop();
135  }
136  }
137 
138  /**
139  * Using MPICH valgrind reports condition on uninitialized value
140  * here. Nevertheless we cna not get rid of it even after explicitly set all allocated space
141  * to zero.
142  */
143  // communicate send sizes
144  int * rcounts = new int [vtx_distr.np()];
145  int * rdispls = new int [vtx_distr.np()];
146  MPI_Alltoall( scounts, 1, MPI_INT, rcounts, 1, MPI_INT, vtx_distr.get_comm());
147 
148  // prepare receiving buffers
149  total_size=0;
150  for(proc=0; proc<vtx_distr.np(); proc++)
151  {
152  rdispls[proc] = total_size;
153  total_size += rcounts[proc];
154  }
155 
156  int * recvbuf = new int [total_size+1];
157 
158  MPI_Alltoallv (
159  sendbuf, scounts, sdispls, MPI_INT,
160  recvbuf, rcounts, rdispls, MPI_INT,
161  vtx_distr.get_comm() );
162 
163  delete [] sendbuf;
164  delete [] scounts;
165  delete [] sdispls;
166  delete [] rcounts;
167  delete [] rdispls;
168 
169  /////////////////////////////////////
170  // construct local adj and rows arrays
171  Edge *edges= (Edge *) recvbuf;
172  int size=total_size/edge_size;
173  std::sort(edges, edges + size);
174 
175  allocate_sparse_graph(vtx_distr.lsize() + 1, size+1);
176  rows[0]=0;
177 
178  if (size != 0) {
179 
180  int i;
181  int row=0;
182  int i_adj=0;
183  int loc_from;
184  Edge unknown_edge={-1,-1,0};
185  Edge *last_edge=&unknown_edge;
186 
187  for(i=0;i<size;i++) {
188  if (! (*last_edge < edges[i]) ) continue; // skip equivalent edges
189  last_edge=edges+i;
190 
191  OLD_ASSERT(vtx_distr.is_local(edges[i].from),
192  "Received non-local edge: %d %d at position %d\n",edges[i].from, edges[i].to,i);
193 
194  loc_from=edges[i].from-vtx_distr.begin();
195  OLD_ASSERT( row <= loc_from, "Decrease in sorted edges at %d\n",i);
196 
197  while ( row < loc_from ) rows[++row]=i;
198  adj[i_adj]=edges[i].to;
199  adj_weights[i_adj]=edges[i].weight;
200  i_adj++;
201  }
202  rows[++row]=i; // i==size (of adj array)
203 
204  }
205 
206  delete [] recvbuf;
207 }
208 
209 
210 /**
211  * Check if the subgraphs of the partitions are connected.
212  */
214 {
215  OLD_ASSERT( vtx_distr.lsize(0)==vtx_distr.size() , "Check of graph continuity not yet implemented for paralel case.\n");
216  if (vtx_distr.myp()!=0) return(true);
217 
218  part_to_check=part;
219  checked_vtx.resize(vtx_distr.size(), 0);
220  std::vector<bool> checked_proc(vtx_distr.np(), false);
221 
222  unsigned int n_proc=0;
223  for(unsigned int vtx=0; n_proc<vtx_distr.np() && vtx<vtx_distr.size(); vtx++) {
224  if (checked_vtx[vtx] != 2) {
226  // check if the processor is still unvisited
227  if ( checked_proc[proc_to_check] )
228  WarningOut().fmt("Disconnected subgraph {} detected at vertex {}.\n", part_to_check[vtx], vtx);
229  // DFS unvisited vertex
230  checked_vtx[vtx]=1;
231  DFS(vtx);
232  checked_proc[proc_to_check]=true;
233  n_proc++;
234  }
235  }
236  checked_vtx.clear();
237 
238  DebugOut() << "Connectivity of subgraphs is OK.\n";
239  return (true);
240 }
241 
242 /**
243  * Recursive part of Deep First Search algorithm of the connectivity check.
244  */
245 void SparseGraph::DFS(int vtx)
246 {
247  OLD_ASSERT( vtx>=0 && vtx< (int) vtx_distr.size(),"Invalid entry vertex %d in DFS.\n",vtx);
248  int neighbour;
249  for(int i_neigh=rows[vtx]; i_neigh< rows[vtx+1];i_neigh++) {
250  neighbour = adj[i_neigh];
251  if (part_to_check[neighbour] == proc_to_check && checked_vtx[neighbour] == 0) {
252  checked_vtx[neighbour]=1;
253  DFS(neighbour);
254  }
255  }
256  checked_vtx[vtx]=2;
257 }
258 
259 
260 
262 {
263  OLD_ASSERT( adj,"Can not view non finalized graph.\n");
264  int row,col;
265  MessageOut() << "SparseGraph\n";
266  for(row=0; row < (int) vtx_distr.lsize(); row++) {
267  MessageOut().fmt("edges from this vertex: {}\n", rows[row+1]);
268  for(col=rows[row]; col<rows[row+1]; col++) {
269  MessageOut().fmt("edge (v1, v2): {} {}\n", row+vtx_distr.begin(), adj[col]);
270  }
271  }
272 }
273 
274 
276 {
277  OLD_ASSERT( rows && adj, "Graph is not yet finalized.");
278 
279  int loc_row, row, row_pos;
280  int col_pos,col,loc_col;
281 
282  for(loc_row=0;loc_row< (int) vtx_distr.lsize();loc_row++) {
283  row=loc_row+vtx_distr.begin();
284  for(row_pos=rows[loc_row];row_pos<rows[loc_row+1];row_pos++) {
285  col=adj[row_pos];
286  if (vtx_distr.is_local(col)) {
287  // find the local column
288  loc_col=col-vtx_distr.begin();
289  for(col_pos=rows[loc_col]; col_pos<rows[loc_col+1]; col_pos++)
290  if (adj[col_pos]==row) break;
291  if (adj[col_pos]!=row) return false;
292  }
293  }
294  }
295 
296  return true;
297 }
298 
299 
300 /**
301  * Free all non-null pointers.
302  */
304 {
305  delete[] vtx_XYZ;
306  delete[] vtx_weights;
307 }
308 
309 /*!
310  * @brief Output a sparse graph.
311  * @return Changed ostream.
312  */
313 ostream & operator <<(ostream & out, const SparseGraph &)
314 {
315  // TODO
316  ASSERT(0).error("Not implemented!");
317  return out;
318 }
319 
320 
321 /*****************************************************************************************
322  SPARSE GRAPH PETSC
323  *****************************************************************************************/
324 
325 void SparseGraphPETSC::allocate_sparse_graph(int lsize_vtxs, int lsize_adj)
326 {
327  PetscMalloc(lsize_vtxs*sizeof(int), &rows);
328  PetscMalloc(lsize_adj*sizeof(int),&adj);
329  PetscMalloc(lsize_adj*sizeof(int),&adj_weights);
330 }
331 
332 
333 void SparseGraphPETSC::partition(int *loc_part)
334 {
335  OLD_ASSERT( adj && rows,"Can not make partition of non finalized graph.\n");
336 
337  MatCreateMPIAdj(vtx_distr.get_comm(), vtx_distr.lsize(),vtx_distr.size(),
338  rows, adj,adj_weights, &petsc_adj_mat);
339  MatPartitioningCreate(vtx_distr.get_comm(),& petsc_part);
340  MatPartitioningSetAdjacency(petsc_part, petsc_adj_mat);
341  MatPartitioningSetFromOptions(petsc_part);
342  MatPartitioningApply(petsc_part,&part_IS);
343 
344  // copy the partition array
345  const PetscInt *is_array;
346  ISGetIndices(part_IS, &is_array);
347  memcpy(loc_part, is_array, vtx_distr.lsize()*sizeof(int));
348  ISRestoreIndices(part_IS, &is_array);
349 }
350 
351 
353 {
354  chkerr(ISDestroy(&part_IS));
355  chkerr(MatPartitioningDestroy(&petsc_part));
356  chkerr(MatDestroy(&petsc_adj_mat));
357 
358  if (adj) PetscFree(adj);
359  if (rows) PetscFree(rows);
360  if (adj_weights) PetscFree(adj_weights);
361 
362  // automaticaly calls base class destructor
363 }
364 
365 
366 /*****************************************************************************************
367  SPARSE GRAPH METIS
368  *****************************************************************************************/
369 
370 
371 void SparseGraphMETIS::allocate_sparse_graph(int lsize_vtxs, int lsize_adj)
372 {
373  rows = new int[lsize_vtxs];
374  adj = new int[lsize_adj];
375  adj_weights = new int [lsize_adj];
376 }
377 
378 
380 {
382  "METIS could be used only with localized distribution.\n");
383  if (vtx_distr.np()==1) {
384  for(unsigned int i=0;i<vtx_distr.size();i++) part[i]=0;
385  return;
386  } else {
387  if (vtx_distr.myp()==0) {
388  int n_vtx=vtx_distr.size();
389  int n_proc=vtx_distr.np();
390  int num_flag=0; // indexing style (0-C, 1-Fortran)
391  int edgecut;
392 
393 /***********************************************************************************
394  * SETTING OPTIONS
395  */
396 #if (METIS_VER_MAJOR >= 5)
397  if ( sizeof(idx_t) != sizeof(int) ) {
398  printf("ERROR in GRAPH_DIVIDE_C: Wrong type of integers for METIS.\n");
399  abort();
400  }
401  int ncon = 1;
402  // Unbalance tolerance vector. Multi-criteria balance is available in METIS 5.x.
403  real_t ubvec[1];
404  ubvec[0] = 1.001; // This is in fact default value used by METIS.
405  int options[METIS_NOPTIONS];
406 
407  for (unsigned int i = 0;i < METIS_NOPTIONS;i++) options[i] = -1;
408 
409  options[METIS_OPTION_OBJTYPE] = METIS_OBJTYPE_CUT;
410  options[METIS_OPTION_CTYPE] = METIS_CTYPE_RM;
411  options[METIS_OPTION_IPTYPE] = METIS_IPTYPE_GROW;
412  options[METIS_OPTION_RTYPE] = METIS_RTYPE_GREEDY;
413  options[METIS_OPTION_NCUTS] = 1;
414  options[METIS_OPTION_NSEPS] = 1;
415  options[METIS_OPTION_NUMBERING] = num_flag;
416  options[METIS_OPTION_NITER] = 10;
417  options[METIS_OPTION_SEED] = 12345;
418  options[METIS_OPTION_MINCONN] = 1;
419  options[METIS_OPTION_CONTIG] = 0;
420  options[METIS_OPTION_COMPRESS] = 0;
421  options[METIS_OPTION_CCORDER] = 0;
422  options[METIS_OPTION_UFACTOR] = 30;
423  //options[METIS_OPTION_DBGLVL] = METIS_DBG_INFO | METIS_DBG_TIME;
424  options[METIS_OPTION_DBGLVL] = 0;
425 #else
426  /* weights */
427  int wgtflag=3;
428  int options[5];
429  for (unsigned int i = 0; i < 5; i++ ) options[i] = 0;
430 
431 #endif
432 
433 
434  /* Initialize parts */
435  for (unsigned int i = 0; i < vtx_distr.lsize(); i++ ) part[i] = num_flag;
436 
437 /***********************************************************************************
438  * CALL METIS using optimal algorithm depending on the number of partitions
439  */
440 
441  if (n_proc <= 8) {
442 
443 #if (METIS_VER_MAJOR >= 5)
444  options[METIS_OPTION_PTYPE] = METIS_PTYPE_RB;
445  options[METIS_OPTION_UFACTOR] = 1;
446  METIS_PartGraphRecursive(&n_vtx, &ncon, rows, adj,
447  vtx_weights, NULL, adj_weights, &n_proc, NULL,
448  ubvec, options, &edgecut,part);
449 #else
450  METIS_PartGraphRecursive(&n_vtx, rows, adj,
451  vtx_weights, adj_weights, &wgtflag, &num_flag,
452  &n_proc, options, &edgecut, part);
453 #endif
454  } else {
455 
456 #if (METIS_VER_MAJOR >= 5)
457  options[METIS_OPTION_PTYPE] = METIS_PTYPE_KWAY;
458  options[METIS_OPTION_UFACTOR] = 30;
459  METIS_PartGraphKway(&n_vtx, &ncon, rows, adj,
460  vtx_weights, NULL, adj_weights, &n_proc, NULL,
461  ubvec, options, &edgecut, part);
462 #else
463  METIS_PartGraphKway(&n_vtx,rows,adj, //vtx distr, local vtx begins, edges of local vtxs
464  vtx_weights,adj_weights,&wgtflag,&num_flag, // vertex, edge weights, ...
465  &n_proc,options,&edgecut,part);
466 #endif
467  }
468  }
469  }
470 }
471 
473 {
474  if (adj) delete[](adj);
475  if (rows) delete[](rows);
476  if (adj_weights) delete[](adj_weights);
477 
478  // automaticaly calls base class destructor
479 }
unsigned int size() const
get global size
unsigned int get_proc(unsigned int idx) const
get processor of the given index
int MPI_Comm
Definition: mpi.h:141
bool check_subgraph_connectivity(int *part)
#define MessageOut()
Macro defining &#39;message&#39; record of log.
Definition: logger.hh:267
SparseGraph(const Distribution &distr)
Definition: sparse_graph.cc:37
void DFS(int vtx)
void chkerr(unsigned int ierr)
Replacement of new/delete operator in the spirit of xmalloc.
Definition: system.hh:142
#define ASSERT(expr)
Allow use shorter versions of macro names if these names is not used with external library...
Definition: asserts.hh:347
void set_vtx_position(const int vtx, const float xyz[3], int weight=1)
Definition: sparse_graph.cc:76
#define OLD_ASSERT(...)
Definition: global_defs.h:108
virtual ~SparseGraphPETSC()
int * part_to_check
created partitioning used through check of connectivity
Global macros to enhance readability and debugging, general constants.
int weight
Edge weights for communication (optional).
bool is_local(unsigned int idx) const
identify local index
#define MPI_Alltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm)
Definition: mpi.h:595
void finalize()
Make sparse graph structures: rows, adj.
Definition: sparse_graph.cc:98
Virtual class for construction and partitioning of a distributed sparse graph.
Definition: sparse_graph.hh:56
Distribution vtx_distr
distribution of vertexes
virtual void allocate_sparse_graph(int lsize_vtxs, int lsize_adj)=0
unsigned int begin(int proc) const
get starting local index
vector< stack< Edge > > adj_of_proc
storage for graph edges for individual processors
virtual void partition(int *loc_part)
virtual ~SparseGraph()
int proc_to_check
subgraph to check
unsigned int np() const
get num of processors
std::vector< int > checked_vtx
coloring of DFS algorithm
unsigned int myp() const
get my processor
Support classes for parallel programing.
float * vtx_XYZ
optional vertex coordinates (global array)
virtual ~SparseGraphMETIS()
ostream & operator<<(ostream &out, const SparseGraph &)
Output a sparse graph.
virtual void allocate_sparse_graph(int lsize_vtxs, int lsize_adj)
#define MPI_INT
Definition: mpi.h:160
MPI_Comm get_comm() const
Returns communicator.
int * adj_weights
Distributed sparse graphs, partitioning.
#define WarningOut()
Macro defining &#39;warning&#39; record of log.
Definition: logger.hh:270
virtual void allocate_sparse_graph(int lsize_vtxs, int lsize_adj)
int * vtx_weights
Vertex weights for computations (optional).
#define DebugOut()
Macro defining &#39;debug&#39; record of log.
Definition: logger.hh:276
int * adj
sparse adjency
friend bool operator<(const Edge &a, const Edge &b)
Definition: sparse_graph.cc:89
bool is_symmetric()
void set_edge(const int a, const int b, int weight=1)
Definition: sparse_graph.cc:66
virtual void partition(int *loc_part)
#define MPI_Alltoallv(sendbuf, sendcounts, sdispls, sendtype, recvbuf, recvcounts, rdispls, recvtype, comm)
Definition: mpi.h:602
void printf(BasicWriter< Char > &w, BasicCStringRef< Char > format, ArgList args)
Definition: printf.h:444
unsigned int lsize(int proc) const
get local size