Merge branch 'metis-tests' into 'master'

 METIS fixes

Code to re-scale the METIS weights in an attempt to avoid large ranges
that fail to produce a complete partition. If that still happens then
handle that case by continuing with the current partition, noting this
problem with a warning.

Also adds function to dump the METIS graphs, possibly separated
into only weights for off-line analysis.

See merge request !61

src/debug.c

 /*******************************************************************************
  * This file is part of SWIFT.
- * Copyright (c) 2013 Matthieu Schaller (matthieu.schaller@durham.ac.uk),
- *                    Pedro Gonnet (pedro.gonnet@durham.ac.uk).
+ * Copyright (c) 2013- 2015:
+ *                    Matthieu Schaller (matthieu.schaller@durham.ac.uk),
+ *                    Pedro Gonnet (pedro.gonnet@durham.ac.uk),
+ *                    Peter W. Draper (p.w.draper@durham.ac.uk).
  *
  * This program is free software: you can redistribute it and/or modify
  * it under the terms of the GNU Lesser General Public License as published
@@ -20,8 +22,10 @@
 
 #include <stdio.h>
 
+#include "config.h"
 #include "const.h"
 #include "part.h"
+#include "debug.h"
 
 /**
  * @brief Looks for the particle with the given id and prints its information to
@@ -98,3 +102,140 @@ void printParticle_single(struct part *p) {
       p->rho_dh, p->density.div_v, p->u, p->force.u_dt, p->force.balsara,
       p->force.POrho2, p->force.v_sig, p->dt);
 }
+
+#ifdef HAVE_METIS
+
+/**
+ * @brief Dump the METIS graph in standard format, simple format and weights
+ * only, to a file.
+ *
+ * @description The standard format output can be read into the METIS
+ * command-line tools. The simple format is just the cell connectivity (this
+ * should not change between calls).  The weights format is the standard one,
+ * minus the cell connectivity.
+ *
+ * The output filenames are generated from the prefix and the sequence number
+ * of calls. So the first is called <prefix>_std_001.dat, <prefix>_simple_001.dat,
+ * <prefix>_weights_001.dat, etc.
+ *
+ * @param prefix base output filename
+ * @param nvertices the number of vertices
+ * @param nvertexweights the number vertex weights
+ * @param cellconruns first part of cell connectivity info (CSR)
+ * @param cellcon second part of cell connectivity info (CSR)
+ * @param vertexweights weights of vertices
+ * @param vertexsizes size of vertices
+ * @param edgeweights weights of edges
+ */
+void dumpMETISGraph(const char *prefix, idx_t nvertices, idx_t nvertexweights,
+                    idx_t *cellconruns, idx_t *cellcon, idx_t *vertexweights, 
+                    idx_t *vertexsizes, idx_t *edgeweights) {
+  FILE *stdfile = NULL;
+  FILE *simplefile = NULL;
+  FILE *weightfile = NULL;
+  char fname[200];
+  idx_t i;
+  idx_t j;
+  int haveedgeweight = 0;
+  int havevertexsize = 0;
+  int havevertexweight = 0;
+  static int nseq = 0;
+  nseq++;
+
+  if (vertexweights != NULL) {
+    for (i = 0; i < nvertices * nvertexweights; i++) {
+      if (vertexweights[i] != 1) {
+        havevertexweight = 1;
+        break;
+      }
+    }
+  }
+
+  if (vertexsizes != NULL) {
+    for (i = 0; i < nvertices; i++) {
+      if (vertexsizes[i] != 1) {
+        havevertexsize = 1;
+        break;
+      }
+    }
+  }
+
+  if (edgeweights != NULL) {
+    for (i = 0; i < cellconruns[nvertices]; i++) {
+      if (edgeweights[i] != 1) {
+        haveedgeweight = 1;
+        break;
+      }
+    }
+  }
+
+  /*  Open output files. */
+  sprintf(fname, "%s_std_%03d.dat", prefix, nseq);
+  stdfile = fopen( fname, "w" );
+
+  sprintf(fname, "%s_simple_%03d.dat", prefix, nseq);
+  simplefile = fopen( fname, "w" );
+
+  if (havevertexweight || havevertexsize || haveedgeweight) {
+    sprintf(fname, "%s_weights_%03d.dat", prefix, nseq);
+    weightfile = fopen( fname, "w" );
+  }
+
+  /*  Write the header lines. */
+  fprintf(stdfile, "%" PRIDX " %" PRIDX, nvertices, cellconruns[nvertices] / 2);
+  fprintf(simplefile, "%" PRIDX " %" PRIDX, nvertices, cellconruns[nvertices] / 2);
+  if (havevertexweight || havevertexsize || haveedgeweight) {
+    fprintf(weightfile, "%" PRIDX " %" PRIDX, nvertices, cellconruns[nvertices] / 2);
+
+    fprintf(stdfile, " %d%d%d", havevertexsize, havevertexweight, haveedgeweight);
+    fprintf(weightfile, " %d%d%d", havevertexsize, havevertexweight, haveedgeweight);
+
+    if (havevertexweight) {
+      fprintf(stdfile, " %d", (int)nvertexweights);
+      fprintf(weightfile, " %d", (int)nvertexweights);
+    }
+  }
+
+  /*  Write the rest of the graph. */
+  for (i = 0; i < nvertices; i++) {
+    fprintf(stdfile, "\n");
+    fprintf(simplefile, "\n");
+    if (weightfile != NULL) {
+        fprintf(weightfile, "\n");
+    }
+
+    if (havevertexsize) {
+      fprintf(stdfile, " %" PRIDX, vertexsizes[i]);
+      fprintf(weightfile, " %" PRIDX, vertexsizes[i]);
+    }
+
+    if (havevertexweight) {
+      for (j = 0; j < nvertexweights; j++) {
+        fprintf(stdfile, " %" PRIDX, vertexweights[i * nvertexweights + j]);
+        fprintf(weightfile, " %" PRIDX, vertexweights[i * nvertexweights + j]);
+      }
+    }
+
+    for (j = cellconruns[i]; j < cellconruns[i + 1]; j++) {
+      fprintf(stdfile, " %" PRIDX, cellcon[j] + 1);
+      fprintf(simplefile, " %" PRIDX, cellcon[j] + 1);
+      if (haveedgeweight) {
+        fprintf(stdfile, " %" PRIDX, edgeweights[j]);
+        fprintf(weightfile, " %" PRIDX, edgeweights[j]);
+      }
+    }
+  }
+  fprintf(stdfile, "\n");
+  fprintf(simplefile, "\n");
+  if (weightfile != NULL) {
+      fprintf(weightfile, "\n");
+  }
+
+  fclose(stdfile);
+  fclose(simplefile);
+  if (weightfile != NULL) {
+    fclose(weightfile);
+  }
+}
+
+#endif

src/debug.h

@@ -27,4 +27,11 @@ void printParticle(struct part *parts, long long int i, int N);
 void printgParticle(struct gpart *parts, long long int i, int N);
 void printParticle_single(struct part *p);
 
+#ifdef HAVE_METIS
+#include "metis.h"
+void dumpMETISGraph(const char *prefix, idx_t nvtxs, idx_t ncon,
+                    idx_t *xadj, idx_t *adjncy, idx_t *vwgt, idx_t *vsize,
+                    idx_t *adjwgt);
+
+#endif
 #endif /* SWIFT_DEBUG_H */

src/engine.c

@@ -306,7 +306,8 @@ void engine_repartition(struct engine *e) {
   int nr_nodes = e->nr_nodes, nodeID = e->nodeID;
   float wscale = 1e-3, vscale = 1e-3, wscale_buff;
   idx_t wtot = 0;
-  const idx_t wmax = 1e9 / e->nr_nodes;
+  idx_t wmax = 1e9 / e->nr_nodes;
+  idx_t wmin;
 
   /* Clear the repartition flag. */
   e->forcerepart = 0;
@@ -489,6 +490,24 @@ void engine_repartition(struct engine *e) {
   /* As of here, only one node needs to compute the partition. */
   if (nodeID == 0) {
 
+    /* Final rescale of all weights to avoid a large range. Large ranges have
+     * been seen to cause an incomplete graph. */
+    wmin = wmax;
+    wmax = 0.0;
+    for (k = 0; k < 26 * nr_cells; k++) {
+      wmax = weights_e[k] > wmax ? weights_e[k] : wmax;
+      wmin = weights_e[k] < wmin ? weights_e[k] : wmin;
+    }
+    if ((wmax - wmin) > engine_maxmetisweight) {
+      wscale = engine_maxmetisweight / (wmax - wmin);
+      for (k = 0; k < 26 * nr_cells; k++) {
+        weights_e[k] = (weights_e[k] - wmin) * wscale + 1;
+      }
+      for (k = 0; k < nr_cells; k++) {
+        weights_v[k] = (weights_v[k] - wmin) * wscale + 1;
+      }
+    }
+
     /* Check that the edge weights are fully symmetric. */
     /* for ( cid = 0 ; cid < nr_cells ; cid++ )
         for ( k = 0 ; k < 26 ; k++ ) {
@@ -547,21 +566,47 @@ void engine_repartition(struct engine *e) {
     /* Call METIS. */
     idx_t one = 1, idx_nr_cells = nr_cells, idx_nr_nodes = nr_nodes;
     idx_t objval;
+
+    /* Dump graph in METIS format */
+    /*dumpMETISGraph("metis_graph", idx_nr_cells, one, offsets, inds,
+                   weights_v, NULL, weights_e);*/
+
     if (METIS_PartGraphRecursive(&idx_nr_cells, &one, offsets, inds, weights_v,
                                  NULL, weights_e, &idx_nr_nodes, NULL, NULL,
                                  options, &objval, nodeIDs) != METIS_OK)
-      error("Call to METIS_PartGraphKway failed.");
+      error("Call to METIS_PartGraphRecursive failed.");
 
     /* Dump the 3d array of cell IDs. */
     /* printf( "engine_repartition: nodeIDs = reshape( [" );
     for ( i = 0 ; i < cdim[0]*cdim[1]*cdim[2] ; i++ )
         printf( "%i " , (int)nodeIDs[ i ] );
     printf("] ,%i,%i,%i);\n",cdim[0],cdim[1],cdim[2]); */
-    
+
     /* Check that the nodeIDs are ok. */
     for (k = 0; k < nr_cells; k++)
       if (nodeIDs[k] < 0 || nodeIDs[k] >= nr_nodes)
         error("Got bad nodeID %"PRIDX" for cell %i.", nodeIDs[k], k);
+
+    /* Check that the partition is complete and all nodes have some work. */
+    int present[nr_nodes];
+    int failed = 0;
+    for (i = 0; i < nr_nodes; i++) present[i] = 0;
+    for (i = 0; i < nr_cells; i++) present[nodeIDs[i]]++;
+    for (i = 0; i < nr_nodes; i++) {
+      if (! present[i]) {
+        failed = 1;
+        message("Node %d is not present after repartition", i);
+      }
+    }
+
+    /* If partition failed continue with the current one, but make this
+     * clear. */
+    if (failed) {
+      message("WARNING: METIS repartition has failed, continuing with "
+              "the current partition, load balance will not be optimal");
+      for (k = 0; k < nr_cells; k++) nodeIDs[k] = cells[k].nodeID;
+    }
+    
   }
 
 /* Broadcast the result of the partition. */

src/engine.h

@@ -47,6 +47,8 @@
 #define engine_maxproxies 64
 #define engine_tasksreweight 10
 
+#define engine_maxmetisweight 10000.0f
+
 /* The rank of the engine as a global variable (for messages). */
 extern int engine_rank;