Merge branch 'partition-memory-edges' into 'master'

Initial partition with memory weighted edges

See merge request !1009

doc/RTD/source/ParameterFiles/parameter_description.rst

@@ -986,26 +986,34 @@ should contain. The type of partitioning attempted is controlled by the::
   DomainDecomposition:
     initial_type:
 
-parameter. Which can have the values *memory*, *region*, *grid* or
+parameter. Which can have the values *memory*, *edgememory*, *region*, *grid* or
 *vectorized*:
 
+    * *edgememory*
+
+    This is the default if METIS or ParMETIS is available. It performs a
+    partition based on the memory use of all the particles in each cell.
+    The total memory per cell is used to weight the cell vertex and all the
+    associated edges. This attempts to equalize the memory used by all the
+    ranks but with some consideration given to the need to not cut dense
+    regions (by also minimizing the edge cut). How successful this
+    attempt is depends on the granularity of cells and particles and the
+    number of ranks, clearly if most of the particles are in one cell, or a
+    small region of the volume, balance is impossible or difficult. Having
+    more top-level cells makes it easier to calculate a good distribution
+    (but this comes at the cost of greater overheads).
 
     * *memory*
 
-    This is the default if METIS or ParMETIS is available. It performs a
-    partition based on the memory use of all the particles in each cell,
-    attempting to equalize the memory used by all the ranks.
-    How successful this attempt is depends on the granularity of cells and particles
-    and the number of ranks, clearly if most of the particles are in one cell,
-    or a small region of the volume, balance is impossible or
-    difficult. Having more top-level cells makes it easier to calculate a
-    good distribution (but this comes at the cost of greater overheads).
+    This is like *edgememory*, but doesn't include any edge weights, it should
+    balance the particle memory use per rank more exactly (but note effects
+    like the numbers of cells per rank will also have an effect, as that
+    changes the need for foreign cells).
 
     * *region*
 
     The one other METIS/ParMETIS option is "region". This attempts to assign equal
-    numbers of cells to each rank, with the surface area of the regions minimised
-    (so we get blobs, rather than rectangular volumes of cells).
+    numbers of cells to each rank, with the surface area of the regions minimised.
 
 If ParMETIS and METIS are not available two other options are possible, but
 will give a poorer partition:

src/debug.c

@@ -375,31 +375,11 @@ static void dumpCells_map(struct cell *c, void *data) {
   size_t *ldata = (size_t *)data;
   FILE *file = (FILE *)ldata[0];
   struct engine *e = (struct engine *)ldata[1];
-  float ntasks = c->nr_tasks;
   int super = (int)ldata[2];
   int active = (int)ldata[3];
   int mpiactive = (int)ldata[4];
   int pactive = (int)ldata[5];
 
-#if SWIFT_DEBUG_CHECKS
-  /* The c->nr_tasks field does not include all the tasks. So let's check this
-   * the hard way. Note pairs share the task 50/50 with the other cell. */
-  ntasks = 0.0f;
-  struct task *tasks = e->sched.tasks;
-  int nr_tasks = e->sched.nr_tasks;
-  for (int k = 0; k < nr_tasks; k++) {
-    if (tasks[k].cj == NULL) {
-      if (c == tasks[k].ci) {
-        ntasks = ntasks + 1.0f;
-      }
-    } else {
-      if (c == tasks[k].ci || c == tasks[k].cj) {
-        ntasks = ntasks + 0.5f;
-      }
-    }
-  }
-#endif
-
   /* Only cells with particles are dumped. */
   if (c->hydro.count > 0 || c->grav.count > 0 || c->stars.count > 0) {
 
@@ -429,6 +409,36 @@ static void dumpCells_map(struct cell *c, void *data) {
         (!super || ((super && c->super == c) || (c->parent == NULL))) &&
         active && mpiactive) {
 
+      /* The c->nr_tasks field does not include all the tasks. So let's check
+       * this the hard way. Note pairs share the task 50/50 with the other
+       * cell. Also accumulate all the time used by tasks of this cell and
+       * form some idea of the effective task depth. */
+      float ntasks = 0.0f;
+      struct task *tasks = e->sched.tasks;
+      int nr_tasks = e->sched.nr_tasks;
+      double ticsum = 0.0; /* Sum of work for this cell. */
+      double dsum = 0.0;
+      for (int k = 0; k < nr_tasks; k++) {
+        if (tasks[k].cj == NULL) {
+          if (tasks[k].ci != NULL) {
+            if (c == tasks[k].ci || c == tasks[k].ci->super) {
+              ntasks = ntasks + 1.0f;
+              ticsum += (tasks[k].toc - tasks[k].tic);
+              dsum += tasks[k].ci->depth;
+            }
+          }
+        } else {
+          if (c == tasks[k].ci || c == tasks[k].ci->super || c == tasks[k].cj ||
+              c == tasks[k].cj->super) {
+            ntasks = ntasks + 0.5f;
+            ticsum += 0.5 * (tasks[k].toc - tasks[k].tic);
+            if (tasks[k].ci != NULL) dsum += (tasks[k].ci->depth * 0.5);
+            dsum += (tasks[k].cj->depth * 0.5);
+          }
+        }
+      }
+      dsum /= (double)ntasks;
+
       /* If requested we work out how many particles are active in this cell. */
       int pactcount = 0;
       if (pactive) {
@@ -443,15 +453,16 @@ static void dumpCells_map(struct cell *c, void *data) {
           if (spart_is_active(&sparts[k], e)) pactcount++;
       }
 
-      fprintf(file,
-              "  %6.3f %6.3f %6.3f %6.3f %6.3f %6.3f %6d %6d %6d %6d %6d %6d "
-              "%6.1f %20lld %6d %6d %6d %6d %6d %6d %6d\n",
-              c->loc[0], c->loc[1], c->loc[2], c->width[0], c->width[1],
-              c->width[2], e->step, c->hydro.count, c->grav.count,
-              c->stars.count, pactcount, c->depth, ntasks, c->hydro.ti_end_min,
-              get_time_bin(c->hydro.ti_end_min), (c->super == c),
-              (c->parent == NULL), cell_is_active_hydro(c, e), c->nodeID,
-              c->nodeID == e->nodeID, ismpiactive);
+      fprintf(
+          file,
+          "  %6.3f %6.3f %6.3f %6.3f %6.3f %6.3f %6d %6d %6d %6d %6d %6d %6d "
+          "%6.1f %20lld %6d %6d %6d %6d %6d %6d %6d %f %f\n",
+          c->loc[0], c->loc[1], c->loc[2], c->width[0], c->width[1],
+          c->width[2], e->step, c->hydro.count, c->grav.count, c->stars.count,
+          pactcount, c->depth, c->maxdepth, ntasks, c->hydro.ti_end_min,
+          get_time_bin(c->hydro.ti_end_min), (c->super == c),
+          (c->parent == NULL), cell_is_active_hydro(c, e), c->nodeID,
+          c->nodeID == e->nodeID, ismpiactive, ticsum, dsum);
     }
   }
 }
@@ -483,11 +494,12 @@ void dumpCells(const char *prefix, int super, int active, int mpiactive,
 
   /* Header. */
   fprintf(file,
-          "# %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s "
-          "%20s %6s %6s %6s %6s %6s %6s %6s\n",
+          "# %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s %6s "
+          "%20s %6s %6s %6s %6s %6s %6s %6s %6s %6s\n",
           "x", "y", "z", "xw", "yw", "zw", "step", "count", "gcount", "scount",
-          "actcount", "depth", "tasks", "ti_end_min", "timebin", "issuper",
-          "istop", "active", "rank", "local", "mpiactive");
+          "actcount", "depth", "maxdepth", "tasks", "ti_end_min", "timebin",
+          "issuper", "istop", "active", "rank", "local", "mpiactive", "ticsum",
+          "avedepth");
 
   size_t data[6];
   data[0] = (size_t)file;

src/partition.c

@@ -66,7 +66,8 @@
 const char *initial_partition_name[] = {
     "axis aligned grids of cells", "vectorized point associated cells",
     "memory balanced, using particle weighted cells",
-    "similar sized regions, using unweighted cells"};
+    "similar sized regions, using unweighted cells",
+    "memory and edge balanced cells using particle weights"};
 
 /* Simple descriptions of repartition types for reports. */
 const char *repartition_name[] = {
@@ -399,55 +400,158 @@ static void ACCUMULATE_SIZES_MAPPER(gpart);
  */
 static void ACCUMULATE_SIZES_MAPPER(spart);
 
+/* qsort support. */
+static int ptrcmp(const void *p1, const void *p2) {
+  const double *v1 = *(const double **)p1;
+  const double *v2 = *(const double **)p2;
+  return (*v1) - (*v2);
+}
+
 /**
  * @brief Accumulate total memory size in particles per cell.
  *
  * @param s the space containing the cells.
+ * @param verbose whether to report any clipped cell counts.
  * @param counts the number of bytes in particles per cell. Should be
  *               allocated as size s->nr_cells.
  */
-static void accumulate_sizes(struct space *s, double *counts) {
+static void accumulate_sizes(struct space *s, int verbose, double *counts) {
 
   bzero(counts, sizeof(double) * s->nr_cells);
 
   struct counts_mapper_data mapper_data;
-  mapper_data.counts = counts;
   mapper_data.s = s;
+  double gsize = 0.0;
+  double *gcounts = NULL;
+  double hsize = 0.0;
+  double ssize = 0.0;
 
-  double hsize = (double)sizeof(struct part);
-  if (s->nr_parts > 0) {
-    mapper_data.size = hsize;
-    threadpool_map(&s->e->threadpool, accumulate_sizes_mapper_part, s->parts,
-                   s->nr_parts, sizeof(struct part), space_splitsize,
-                   &mapper_data);
-  }
-
-  double gsize = (double)sizeof(struct gpart);
   if (s->nr_gparts > 0) {
+    /* Self-gravity gets more efficient with density (see gitlab issue #640)
+     * so we end up with too much weight in cells with larger numbers of
+     * gparts, to suppress this we fix a upper weight limit based on a
+     * percentile clip to on the numbers of cells. Should be relatively
+     * harmless when not really needed. */
+    if ((gcounts = (double *)malloc(sizeof(double) * s->nr_cells)) == NULL)
+      error("Failed to allocate gcounts buffer.");
+    bzero(gcounts, sizeof(double) * s->nr_cells);
+    gsize = (double)sizeof(struct gpart);
+
+    mapper_data.counts = gcounts;
     mapper_data.size = gsize;
     threadpool_map(&s->e->threadpool, accumulate_sizes_mapper_gpart, s->gparts,
                    s->nr_gparts, sizeof(struct gpart), space_splitsize,
                    &mapper_data);
+
+    /* Get all the counts from all the nodes. */
+    if (MPI_Allreduce(MPI_IN_PLACE, gcounts, s->nr_cells, MPI_DOUBLE, MPI_SUM,
+                      MPI_COMM_WORLD) != MPI_SUCCESS)
+      error("Failed to allreduce particle cell gpart weights.");
+
+    /* Now we need to sort... */
+    double **ptrs = NULL;
+    if ((ptrs = (double **)malloc(sizeof(double *) * s->nr_cells)) == NULL)
+      error("Failed to allocate pointers buffer.");
+    for (int k = 0; k < s->nr_cells; k++) {
+      ptrs[k] = &gcounts[k];
+    }
+
+    /* Sort pointers, not counts... */
+    qsort(ptrs, s->nr_cells, sizeof(double *), ptrcmp);
+
+    /* Percentile cut keeps 99.8% of cells and clips above. */
+    int cut = ceil(s->nr_cells * 0.998);
+
+    /* And clip. */
+    int nadj = 0;
+    double clip = *ptrs[cut];
+    for (int k = cut + 1; k < s->nr_cells; k++) {
+      *ptrs[k] = clip;
+      nadj++;
+    }
+    if (verbose) message("clipped gravity part counts of %d cells", nadj);
+    free(ptrs);
+  }
+
+  /* Other particle types are assumed to correlate with processing time. */
+  if (s->nr_parts > 0) {
+    mapper_data.counts = counts;
+    hsize = (double)sizeof(struct part);
+    mapper_data.size = hsize;
+    threadpool_map(&s->e->threadpool, accumulate_sizes_mapper_part, s->parts,
+                   s->nr_parts, sizeof(struct part), space_splitsize,
+                   &mapper_data);
   }
 
-  double ssize = (double)sizeof(struct spart);
   if (s->nr_sparts > 0) {
+    ssize = (double)sizeof(struct spart);
     mapper_data.size = ssize;
     threadpool_map(&s->e->threadpool, accumulate_sizes_mapper_spart, s->sparts,
                    s->nr_sparts, sizeof(struct spart), space_splitsize,
                    &mapper_data);
   }
 
+  /* Merge the counts arrays across all nodes, if needed. Doesn't include any
+   * gparts. */
+  if (s->nr_parts > 0 || s->nr_sparts > 0) {
+    if (MPI_Allreduce(MPI_IN_PLACE, counts, s->nr_cells, MPI_DOUBLE, MPI_SUM,
+                      MPI_COMM_WORLD) != MPI_SUCCESS)
+      error("Failed to allreduce particle cell weights.");
+  }
+
+  /* And merge with gravity counts. */
+  double sum = 0.0;
+  if (s->nr_gparts > 0) {
+    for (int k = 0; k < s->nr_cells; k++) {
+      counts[k] += gcounts[k];
+      sum += counts[k];
+    }
+    free(gcounts);
+  } else {
+    for (int k = 0; k < s->nr_cells; k++) {
+      sum += counts[k];
+    }
+  }
+
   /* Keep the sum of particles across all ranks in the range of IDX_MAX. */
-  if ((s->e->total_nr_parts * hsize + s->e->total_nr_gparts * gsize +
-       s->e->total_nr_sparts * ssize) > (double)IDX_MAX) {
-    double vscale =
-        (double)(IDX_MAX - 1000) /
-        (double)(s->e->total_nr_parts * hsize + s->e->total_nr_gparts * gsize +
-                 s->e->total_nr_sparts * ssize);
+  if (sum > (double)(IDX_MAX - 10000)) {
+    double vscale = (double)(IDX_MAX - 10000) / sum;
     for (int k = 0; k < s->nr_cells; k++) counts[k] *= vscale;
   }
 }
+
+/**
+ * @brief Make edge weights from the accumulated particle sizes per cell.
+ *
+ * @param s the space containing the cells.
+ * @param counts the number of bytes in particles per cell.
+ * @param edges weights for the edges of these regions. Should be 26 * counts.
+ */
+static void sizes_to_edges(struct space *s, double *counts, double *edges) {
+
+  bzero(edges, sizeof(double) * s->nr_cells * 26);
+
+  for (int l = 0; l < s->nr_cells; l++) {
+    int p = 0;
+    for (int i = -1; i <= 1; i++) {
+      int isid = ((i < 0) ? 0 : ((i > 0) ? 2 : 1));
+      for (int j = -1; j <= 1; j++) {
+        int jsid = isid * 3 + ((j < 0) ? 0 : ((j > 0) ? 2 : 1));
+        for (int k = -1; k <= 1; k++) {
+          int ksid = jsid * 3 + ((k < 0) ? 0 : ((k > 0) ? 2 : 1));
+
+          /* If not self, we work out the sort indices to get the expected
+           * fractional weight and add that. Scale to keep sum less than
+           * counts and a bit of tuning... */
+          if (i || j || k) {
+            edges[l * 26 + p] = counts[l] * sid_scale[sortlistID[ksid]] / 26.0;
+            p++;
+          }
+        }
+      }
+    }
+  }
+}
 #endif
 
 #if defined(WITH_MPI) && (defined(HAVE_METIS) || defined(HAVE_PARMETIS))
@@ -464,7 +568,7 @@ static void split_metis(struct space *s, int nregions, int *celllist) {
 
   /* To check or visualise the partition dump all the cells. */
   /*if (engine_rank == 0) dumpCellRanks("metis_partition", s->cells_top,
-   * s->nr_cells);*/
+                                      s->nr_cells);*/
 }
 #endif
 
@@ -780,7 +884,7 @@ static void pick_parmetis(int nodeID, struct space *s, int nregions,
 
     /* Dump graphs to disk files for testing. */
     /*dumpMETISGraph("parmetis_graph", ncells, 1, std_xadj, full_adjncy,
-      full_weights_v, NULL, full_weights_e);*/
+                   full_weights_v, NULL, full_weights_e); */
 
     /* xadj is set for each rank, different to serial version in that each
      * rank starts with 0, so we need to re-offset. */
@@ -1213,7 +1317,7 @@ static void pick_metis(int nodeID, struct space *s, int nregions,
 
     /* Dump graph in METIS format */
     /*dumpMETISGraph("metis_graph", idx_ncells, one, xadj, adjncy, weights_v,
-      NULL, weights_e);*/
+                   NULL, weights_e);*/
 
     if (METIS_PartGraphKway(&idx_ncells, &one, xadj, adjncy, weights_v, NULL,
                             weights_e, &idx_nregions, NULL, NULL, options,
@@ -1536,22 +1640,22 @@ static void repart_edge_metis(int vweights, int eweights, int timebins,
   if (vweights && eweights) {
     if (vsum > esum) {
       if (vsum > (double)IDX_MAX) {
-        vscale = (double)(IDX_MAX - 1000) / vsum;
+        vscale = (double)(IDX_MAX - 10000) / vsum;
         escale = vscale;
       }
     } else {
       if (esum > (double)IDX_MAX) {
-        escale = (double)(IDX_MAX - 1000) / esum;
+        escale = (double)(IDX_MAX - 10000) / esum;
         vscale = escale;
       }
     }
   } else if (vweights) {
     if (vsum > (double)IDX_MAX) {
-      vscale = (double)(IDX_MAX - 1000) / vsum;
+      vscale = (double)(IDX_MAX - 10000) / vsum;
     }
   } else if (eweights) {
     if (esum > (double)IDX_MAX) {
-      escale = (double)(IDX_MAX - 1000) / esum;
+      escale = (double)(IDX_MAX - 10000) / esum;
     }
   }
 
@@ -1655,17 +1759,11 @@ static void repart_memory_metis(struct repartition *repartition, int nodeID,
   double *weights = NULL;
   if ((weights = (double *)malloc(sizeof(double) * s->nr_cells)) == NULL)
     error("Failed to allocate cell weights buffer.");
-  bzero(weights, sizeof(double) * s->nr_cells);
 
   /* Check each particle and accumulate the sizes per cell. */
-  accumulate_sizes(s, weights);
+  accumulate_sizes(s, s->e->verbose, weights);
 
-  /* Get all the counts from all the nodes. */
-  if (MPI_Allreduce(MPI_IN_PLACE, weights, s->nr_cells, MPI_DOUBLE, MPI_SUM,
-                    MPI_COMM_WORLD) != MPI_SUCCESS)
-    error("Failed to allreduce particle cell weights.");
-
-    /* Allocate cell list for the partition. If not already done. */
+  /* Allocate cell list for the partition. If not already done. */
 #ifdef HAVE_PARMETIS
   int refine = 1;
 #endif
@@ -1839,27 +1937,38 @@ void partition_initial_partition(struct partition *initial_partition,
     }
 
   } else if (initial_partition->type == INITPART_METIS_WEIGHT ||
+             initial_partition->type == INITPART_METIS_WEIGHT_EDGE ||
              initial_partition->type == INITPART_METIS_NOWEIGHT) {
 #if defined(WITH_MPI) && (defined(HAVE_METIS) || defined(HAVE_PARMETIS))
     /* Simple k-way partition selected by METIS using cell particle
      * counts as weights or not. Should be best when starting with a
      * inhomogeneous dist.
      */
-
-    /* Space for particles sizes per cell, which will be used as weights. */
-    double *weights = NULL;
+    double *weights_v = NULL;
+    double *weights_e = NULL;
     if (initial_partition->type == INITPART_METIS_WEIGHT) {
-      if ((weights = (double *)malloc(sizeof(double) * s->nr_cells)) == NULL)
-        error("Failed to allocate weights buffer.");
-      bzero(weights, sizeof(double) * s->nr_cells);
+      /* Particles sizes per cell, which will be used as weights. */
+      if ((weights_v = (double *)malloc(sizeof(double) * s->nr_cells)) == NULL)
+        error("Failed to allocate weights_v buffer.");
+
+      /* Check each particle and accumulate the sizes per cell. */
+      accumulate_sizes(s, s->e->verbose, weights_v);
+
+    } else if (initial_partition->type == INITPART_METIS_WEIGHT_EDGE) {
 
-      /* Check each particle and accumilate the sizes per cell. */
-      accumulate_sizes(s, weights);
+      /* Particle sizes also counted towards the edges. */
 
-      /* Get all the counts from all the nodes. */
-      if (MPI_Allreduce(MPI_IN_PLACE, weights, s->nr_cells, MPI_DOUBLE, MPI_SUM,
-                        MPI_COMM_WORLD) != MPI_SUCCESS)
-        error("Failed to allreduce particle cell weights.");
+      if ((weights_v = (double *)malloc(sizeof(double) * s->nr_cells)) == NULL)
+        error("Failed to allocate weights_v buffer.");
+      if ((weights_e = (double *)malloc(sizeof(double) * s->nr_cells * 26)) ==
+          NULL)
+        error("Failed to allocate weights_e buffer.");
+
+      /* Check each particle and accumulate the sizes per cell. */
+      accumulate_sizes(s, s->e->verbose, weights_v);
+
+      /* Spread these into edge weights. */
+      sizes_to_edges(s, weights_v, weights_e);
     }
 
     /* Do the calculation. */
@@ -1868,12 +1977,13 @@ void partition_initial_partition(struct partition *initial_partition,
       error("Failed to allocate celllist");
 #ifdef HAVE_PARMETIS
     if (initial_partition->usemetis) {
-      pick_metis(nodeID, s, nr_nodes, weights, NULL, celllist);
+      pick_metis(nodeID, s, nr_nodes, weights_v, weights_e, celllist);
     } else {
-      pick_parmetis(nodeID, s, nr_nodes, weights, NULL, 0, 0, 0.0f, celllist);
+      pick_parmetis(nodeID, s, nr_nodes, weights_v, weights_e, 0, 0, 0.0f,
+                    celllist);
     }
 #else
-    pick_metis(nodeID, s, nr_nodes, weights, NULL, celllist);
+    pick_metis(nodeID, s, nr_nodes, weights_v, weights_e, celllist);
 #endif
 
     /* And apply to our cells */
@@ -1888,7 +1998,8 @@ void partition_initial_partition(struct partition *initial_partition,
       partition_initial_partition(initial_partition, nodeID, nr_nodes, s);
     }
 
-    if (weights != NULL) free(weights);
+    if (weights_v != NULL) free(weights_v);
+    if (weights_e != NULL) free(weights_e);
     free(celllist);
 #else
     error("SWIFT was not compiled with METIS or ParMETIS support");
@@ -1943,7 +2054,7 @@ void partition_init(struct partition *partition,
 /* Defaults make use of METIS if available */
 #if defined(HAVE_METIS) || defined(HAVE_PARMETIS)
   const char *default_repart = "fullcosts";
-  const char *default_part = "memory";
+  const char *default_part = "edgememory";
 #else
   const char *default_repart = "none";
   const char *default_part = "grid";
@@ -1974,10 +2085,13 @@ void partition_init(struct partition *partition,
     case 'm':
       partition->type = INITPART_METIS_WEIGHT;
       break;
+    case 'e':
+      partition->type = INITPART_METIS_WEIGHT_EDGE;
+      break;
     default:
       message("Invalid choice of initial partition type '%s'.", part_type);
       error(
-          "Permitted values are: 'grid', 'region', 'memory' or "
+          "Permitted values are: 'grid', 'region', 'memory', 'edgememory' or "
           "'vectorized'");
 #else
     default:

src/partition.h

@@ -28,7 +28,8 @@ enum partition_type {
   INITPART_GRID = 0,
   INITPART_VECTORIZE,
   INITPART_METIS_WEIGHT,
-  INITPART_METIS_NOWEIGHT
+  INITPART_METIS_NOWEIGHT,
+  INITPART_METIS_WEIGHT_EDGE
 };
 
 /* Simple descriptions of types for reports. */