engine.c

/*******************************************************************************
 * This file is part of SWIFT.
 * Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
 *                    Matthieu Schaller (matthieu.schaller@durham.ac.uk)
 *               2015 Peter W. Draper (p.w.draper@durham.ac.uk)
 *                    Angus Lepper (angus.lepper@ed.ac.uk)
 *               2016 John A. Regan (john.a.regan@durham.ac.uk)
 *                    Tom Theuns (tom.theuns@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
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program 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 Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 ******************************************************************************/

/* Config parameters. */
#include "../config.h"

/* Some standard headers. */
#include <float.h>
#include <limits.h>
#include <sched.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

/* MPI headers. */
#ifdef WITH_MPI
#include <mpi.h>
#endif

#ifdef HAVE_LIBNUMA
#include <numa.h>
#endif

/* This object's header. */
#include "engine.h"

/* Local headers. */
#include "active.h"
#include "atomic.h"
#include "cell.h"
#include "clocks.h"
#include "cycle.h"
#include "debug.h"
#include "error.h"
#include "gravity.h"
#include "hydro.h"
#include "minmax.h"
#include "parallel_io.h"
#include "part.h"
#include "partition.h"
#include "profiler.h"
#include "proxy.h"
#include "runner.h"
#include "serial_io.h"
#include "single_io.h"
#include "sort_part.h"
#include "statistics.h"
#include "timers.h"
#include "tools.h"
#include "units.h"
#include "version.h"

/* Particle cache size. */
#define CACHE_SIZE 512

const char *engine_policy_names[] = {"none",
                                     "rand",
                                     "steal",
                                     "keep",
                                     "block",
                                     "cpu_tight",
                                     "mpi",
                                     "numa_affinity",
                                     "hydro",
                                     "self_gravity",
                                     "external_gravity",
                                     "cosmology_integration",
                                     "drift_all",
                                     "reconstruct_mpoles",
                                     "cooling",
                                     "sourceterms",
                                     "stars"};

/** The rank of the engine as a global variable (for messages). */
int engine_rank;

/**
 * @brief Link a density/force task to a cell.
 *
 * @param e The #engine.
 * @param l A pointer to the #link, will be modified atomically.
 * @param t The #task.
 *
 * @return The new #link pointer.
 */
void engine_addlink(struct engine *e, struct link **l, struct task *t) {

  /* Get the next free link. */
  const int ind = atomic_inc(&e->nr_links);
  if (ind >= e->size_links) {
    error("Link table overflow.");
  }
  struct link *res = &e->links[ind];

  /* Set it atomically. */
  res->t = t;
  res->next = atomic_swap(l, res);
}

/**
 * @brief Recursively add non-implicit ghost tasks to a cell hierarchy.
 */
void engine_add_ghosts(struct engine *e, struct cell *c, struct task *ghost_in,
                       struct task *ghost_out) {
  if (!c->split || c->count < engine_max_parts_per_ghost) {
    struct scheduler *s = &e->sched;
    c->ghost =
        scheduler_addtask(s, task_type_ghost, task_subtype_none, 0, 0, c, NULL);
    scheduler_addunlock(s, ghost_in, c->ghost);
    scheduler_addunlock(s, c->ghost, ghost_out);
  } else {
    for (int k = 0; k < 8; k++)
      if (c->progeny[k] != NULL)
        engine_add_ghosts(e, c->progeny[k], ghost_in, ghost_out);
  }
}

/**
 * @brief Generate the hydro hierarchical tasks for a hierarchy of cells -
 * i.e. all the O(Npart) tasks.
 *
 * Tasks are only created here. The dependencies will be added later on.
 *
 * Note that there is no need to recurse below the super-cell.
 *
 * @param e The #engine.
 * @param c The #cell.
 */
void engine_make_hierarchical_tasks(struct engine *e, struct cell *c) {

  struct scheduler *s = &e->sched;
  const int periodic = e->s->periodic;
  const int is_hydro = (e->policy & engine_policy_hydro);
  const int is_self_gravity = (e->policy & engine_policy_self_gravity);
  const int is_with_cooling = (e->policy & engine_policy_cooling);
  const int is_with_sourceterms = (e->policy & engine_policy_sourceterms);

  /* Are we in a super-cell ? */
  if (c->super == c) {

    /* Add the sort task. */
    c->sorts =
        scheduler_addtask(s, task_type_sort, task_subtype_none, 0, 0, c, NULL);

    /* Local tasks only... */
    if (c->nodeID == e->nodeID) {

      /* Add the drift task. */
      c->drift_part = scheduler_addtask(s, task_type_drift_part,
                                        task_subtype_none, 0, 0, c, NULL);

      /* Add the two half kicks */
      c->kick1 = scheduler_addtask(s, task_type_kick1, task_subtype_none, 0, 0,
                                   c, NULL);

      c->kick2 = scheduler_addtask(s, task_type_kick2, task_subtype_none, 0, 0,
                                   c, NULL);

      /* Add the time-step calculation task and its dependency */
      c->timestep = scheduler_addtask(s, task_type_timestep, task_subtype_none,
                                      0, 0, c, NULL);

      scheduler_addunlock(s, c->kick2, c->timestep);
      scheduler_addunlock(s, c->timestep, c->kick1);

      if (is_self_gravity) {

        /* Initialisation of the multipoles */
        c->init_grav = scheduler_addtask(s, task_type_init_grav,
                                         task_subtype_none, 0, 0, c, NULL);

        /* Gravity non-neighbouring pm calculations */
        c->grav_long_range = scheduler_addtask(
            s, task_type_grav_long_range, task_subtype_none, 0, 0, c, NULL);

        /* Gravity recursive down-pass */
        c->grav_down = scheduler_addtask(s, task_type_grav_down,
                                         task_subtype_none, 0, 0, c, NULL);

        if (periodic) scheduler_addunlock(s, c->init_grav, c->grav_ghost[0]);
        scheduler_addunlock(s, c->init_grav, c->grav_long_range);
        scheduler_addunlock(s, c->grav_long_range, c->grav_down);
        scheduler_addunlock(s, c->grav_down, c->kick2);
      }

      /* Generate the ghost tasks. */
      if (is_hydro) {
        c->ghost_in =
            scheduler_addtask(s, task_type_ghost, task_subtype_none, 0,
                              /* implicit = */ 1, c, NULL);
        c->ghost_out =
            scheduler_addtask(s, task_type_ghost, task_subtype_none, 0,
                              /* implicit = */ 1, c, NULL);
        engine_add_ghosts(e, c, c->ghost_in, c->ghost_out);
      }

#ifdef EXTRA_HYDRO_LOOP
      /* Generate the extra ghost task. */
      if (is_hydro)
        c->extra_ghost = scheduler_addtask(s, task_type_extra_ghost,
                                           task_subtype_none, 0, 0, c, NULL);
#endif

      /* Cooling task */
      if (is_with_cooling) {
        c->cooling = scheduler_addtask(s, task_type_cooling, task_subtype_none,
                                       0, 0, c, NULL);

        scheduler_addunlock(s, c->cooling, c->kick2);
      }

      /* add source terms */
      if (is_with_sourceterms) {
        c->sourceterms = scheduler_addtask(s, task_type_sourceterms,
                                           task_subtype_none, 0, 0, c, NULL);
      }
    }

  } else { /* We are above the super-cell so need to go deeper */

#ifdef SWIFT_DEBUG_CHECKS
    if (c->super != NULL) error("Incorrectly set super pointer");
#endif

    /* Recurse. */
    if (c->split)
      for (int k = 0; k < 8; k++)
        if (c->progeny[k] != NULL)
          engine_make_hierarchical_tasks(e, c->progeny[k]);
  }
}

#ifdef WITH_MPI
/**
 * Do the exchange of one type of particles with all the other nodes.
 *
 * @param counts 2D array with the counts of particles to exchange with
 *               each other node.
 * @param parts the particle data to exchange
 * @param new_nr_parts the number of particles this node will have after all
 *                     exchanges have completed.
 * @param sizeofparts sizeof the particle struct.
 * @param alignsize the memory alignment required for this particle type.
 * @param mpi_type the MPI_Datatype for these particles.
 * @param nr_nodes the number of nodes to exchange with.
 * @param nodeID the id of this node.
 *
 * @result new particle data constructed from all the exchanges with the
 *         given alignment.
 */
static void *engine_do_redistribute(int *counts, char *parts,
                                    size_t new_nr_parts, size_t sizeofparts,
                                    size_t alignsize, MPI_Datatype mpi_type,
                                    int nr_nodes, int nodeID) {

  /* Allocate a new particle array with some extra margin */
  char *parts_new = NULL;
  if (posix_memalign(
          (void **)&parts_new, alignsize,
          sizeofparts * new_nr_parts * engine_redistribute_alloc_margin) != 0)
    error("Failed to allocate new particle data.");

  /* Prepare MPI requests for the asynchronous communications */
  MPI_Request *reqs;
  if ((reqs = (MPI_Request *)malloc(sizeof(MPI_Request) * 2 * nr_nodes)) ==
      NULL)
    error("Failed to allocate MPI request list.");

  /* Only send and receive only "chunk" particles per request. So we need to
   * loop as many times as necessary here. Make 2Gb/sizeofparts so we only
   * send 2Gb packets. */
  const int chunk = INT_MAX / sizeofparts;
  int sent = 0;
  int recvd = 0;

  int activenodes = 1;
  while (activenodes) {

    for (int k = 0; k < 2 * nr_nodes; k++) reqs[k] = MPI_REQUEST_NULL;

    /* Emit the sends and recvs for the data. */
    size_t offset_send = sent;
    size_t offset_recv = recvd;
    activenodes = 0;

    for (int k = 0; k < nr_nodes; k++) {

      /* Indices in the count arrays of the node of interest */
      const int ind_send = nodeID * nr_nodes + k;
      const int ind_recv = k * nr_nodes + nodeID;

      /* Are we sending any data this loop? */
      int sending = counts[ind_send] - sent;
      if (sending > 0) {
        activenodes++;
        if (sending > chunk) sending = chunk;

        /* If the send and receive is local then just copy. */
        if (k == nodeID) {
          int receiving = counts[ind_recv] - recvd;
          if (receiving > chunk) receiving = chunk;
          memcpy(&parts_new[offset_recv * sizeofparts],
                 &parts[offset_send * sizeofparts], sizeofparts * receiving);
        } else {
          /* Otherwise send it. */
          int res =
              MPI_Isend(&parts[offset_send * sizeofparts], sending, mpi_type, k,
                        ind_send, MPI_COMM_WORLD, &reqs[2 * k + 0]);
          if (res != MPI_SUCCESS)
            mpi_error(res, "Failed to isend parts to node %i.", k);
        }
      }

      /* If we're sending to this node, then move past it to next. */
      if (counts[ind_send] > 0) offset_send += counts[ind_send];

      /* Are we receiving any data from this node? Note already done if coming
       * from this node. */
      if (k != nodeID) {
        int receiving = counts[ind_recv] - recvd;
        if (receiving > 0) {
          activenodes++;
          if (receiving > chunk) receiving = chunk;
          int res = MPI_Irecv(&parts_new[offset_recv * sizeofparts], receiving,
                              mpi_type, k, ind_recv, MPI_COMM_WORLD,
                              &reqs[2 * k + 1]);
          if (res != MPI_SUCCESS)
            mpi_error(res, "Failed to emit irecv of parts from node %i.", k);
        }
      }

      /* If we're receiving from this node, then move past it to next. */
      if (counts[ind_recv] > 0) offset_recv += counts[ind_recv];
    }

    /* Wait for all the sends and recvs to tumble in. */
    MPI_Status stats[2 * nr_nodes];
    int res;
    if ((res = MPI_Waitall(2 * nr_nodes, reqs, stats)) != MPI_SUCCESS) {
      for (int k = 0; k < 2 * nr_nodes; k++) {
        char buff[MPI_MAX_ERROR_STRING];
        MPI_Error_string(stats[k].MPI_ERROR, buff, &res);
        message("request from source %i, tag %i has error '%s'.",
                stats[k].MPI_SOURCE, stats[k].MPI_TAG, buff);
      }
      error("Failed during waitall for part data.");
    }

    /* Move to next chunks. */
    sent += chunk;
    recvd += chunk;
  }

  /* Free temps. */
  free(reqs);

  /* And return new memory. */
  return parts_new;
}
#endif

/**
 * @brief Redistribute the particles amongst the nodes according
 *      to their cell's node IDs.
 *
 * The strategy here is as follows:
 * 1) Each node counts the number of particles it has to send to each other
 * node.
 * 2) The number of particles of each type is then exchanged.
 * 3) The particles to send are placed in a temporary buffer in which the
 * part-gpart links are preserved.
 * 4) Each node allocates enough space for the new particles.
 * 5) (Asynchronous) communications are issued to transfer the data.
 *
 *
 * @param e The #engine.
 */
void engine_redistribute(struct engine *e) {

#ifdef WITH_MPI

  const int nr_nodes = e->nr_nodes;
  const int nodeID = e->nodeID;
  struct space *s = e->s;
  struct cell *cells = s->cells_top;
  const int nr_cells = s->nr_cells;
  const int *cdim = s->cdim;
  const double iwidth[3] = {s->iwidth[0], s->iwidth[1], s->iwidth[2]};
  const double dim[3] = {s->dim[0], s->dim[1], s->dim[2]};
  struct part *parts = s->parts;
  struct gpart *gparts = s->gparts;
  struct spart *sparts = s->sparts;
  ticks tic = getticks();

  /* Allocate temporary arrays to store the counts of particles to be sent
   * and the destination of each particle */
  int *counts;
  if ((counts = (int *)malloc(sizeof(int) * nr_nodes * nr_nodes)) == NULL)
    error("Failed to allocate counts temporary buffer.");
  bzero(counts, sizeof(int) * nr_nodes * nr_nodes);

  int *dest;
  if ((dest = (int *)malloc(sizeof(int) * s->nr_parts)) == NULL)
    error("Failed to allocate dest temporary buffer.");

  /* Get destination of each particle */
  for (size_t k = 0; k < s->nr_parts; k++) {

    /* Periodic boundary conditions */
    for (int j = 0; j < 3; j++) {
      if (parts[k].x[j] < 0.0)
        parts[k].x[j] += dim[j];
      else if (parts[k].x[j] >= dim[j])
        parts[k].x[j] -= dim[j];
    }
    const int cid =
        cell_getid(cdim, parts[k].x[0] * iwidth[0], parts[k].x[1] * iwidth[1],
                   parts[k].x[2] * iwidth[2]);
#ifdef SWIFT_DEBUG_CHECKS
    if (cid < 0 || cid >= s->nr_cells)
      error("Bad cell id %i for part %zu at [%.3e,%.3e,%.3e].", cid, k,
            parts[k].x[0], parts[k].x[1], parts[k].x[2]);
#endif

    dest[k] = cells[cid].nodeID;

    /* The counts array is indexed as count[from * nr_nodes + to]. */
    counts[nodeID * nr_nodes + dest[k]] += 1;
  }

  /* Sort the particles according to their cell index. */
  if (s->nr_parts > 0)
    space_parts_sort(s, dest, s->nr_parts, 0, nr_nodes - 1, e->verbose);

#ifdef SWIFT_DEBUG_CHECKS
  /* Verify that the part have been sorted correctly. */
  for (size_t k = 0; k < s->nr_parts; k++) {
    const struct part *p = &s->parts[k];

    /* New cell index */
    const int new_cid =
        cell_getid(s->cdim, p->x[0] * s->iwidth[0], p->x[1] * s->iwidth[1],
                   p->x[2] * s->iwidth[2]);

    /* New cell of this part */
    const struct cell *c = &s->cells_top[new_cid];
    const int new_node = c->nodeID;

    if (dest[k] != new_node)
      error("part's new node index not matching sorted index.");

    if (p->x[0] < c->loc[0] || p->x[0] > c->loc[0] + c->width[0] ||
        p->x[1] < c->loc[1] || p->x[1] > c->loc[1] + c->width[1] ||
        p->x[2] < c->loc[2] || p->x[2] > c->loc[2] + c->width[2])
      error("part not sorted into the right top-level cell!");
  }
#endif

  /* We need to re-link the gpart partners of parts. */
  if (s->nr_parts > 0) {
    int current_dest = dest[0];
    size_t count_this_dest = 0;
    for (size_t k = 0; k < s->nr_parts; ++k) {
      if (s->parts[k].gpart != NULL) {

        /* As the addresses will be invalidated by the communications, we will
         * instead store the absolute index from the start of the sub-array of
         * particles to be sent to a given node.
         * Recall that gparts without partners have a positive id.
         * We will restore the pointers on the receiving node later on. */
        if (dest[k] != current_dest) {
          current_dest = dest[k];
          count_this_dest = 0;
        }

#ifdef SWIFT_DEBUG_CHECKS
        if (s->parts[k].gpart->id_or_neg_offset > 0)
          error("Trying to link a partnerless gpart !");
#endif

        s->parts[k].gpart->id_or_neg_offset = -count_this_dest;
        count_this_dest++;
      }
    }
  }
  free(dest);

  /* Get destination of each s-particle */
  int *s_counts;
  if ((s_counts = (int *)malloc(sizeof(int) * nr_nodes * nr_nodes)) == NULL)
    error("Failed to allocate s_counts temporary buffer.");
  bzero(s_counts, sizeof(int) * nr_nodes * nr_nodes);

  int *s_dest;
  if ((s_dest = (int *)malloc(sizeof(int) * s->nr_sparts)) == NULL)
    error("Failed to allocate s_dest temporary buffer.");

  for (size_t k = 0; k < s->nr_sparts; k++) {

    /* Periodic boundary conditions */
    for (int j = 0; j < 3; j++) {
      if (sparts[k].x[j] < 0.0)
        sparts[k].x[j] += dim[j];
      else if (sparts[k].x[j] >= dim[j])
        sparts[k].x[j] -= dim[j];
    }
    const int cid =
        cell_getid(cdim, sparts[k].x[0] * iwidth[0], sparts[k].x[1] * iwidth[1],
                   sparts[k].x[2] * iwidth[2]);
#ifdef SWIFT_DEBUG_CHECKS
    if (cid < 0 || cid >= s->nr_cells)
      error("Bad cell id %i for spart %zu at [%.3e,%.3e,%.3e].", cid, k,
            sparts[k].x[0], sparts[k].x[1], sparts[k].x[2]);
#endif

    s_dest[k] = cells[cid].nodeID;

    /* The counts array is indexed as count[from * nr_nodes + to]. */
    s_counts[nodeID * nr_nodes + s_dest[k]] += 1;
  }

  /* Sort the particles according to their cell index. */
  if (s->nr_sparts > 0)
    space_sparts_sort(s, s_dest, s->nr_sparts, 0, nr_nodes - 1, e->verbose);

#ifdef SWIFT_DEBUG_CHECKS
  /* Verify that the spart have been sorted correctly. */
  for (size_t k = 0; k < s->nr_sparts; k++) {
    const struct spart *sp = &s->sparts[k];

    /* New cell index */
    const int new_cid =
        cell_getid(s->cdim, sp->x[0] * s->iwidth[0], sp->x[1] * s->iwidth[1],
                   sp->x[2] * s->iwidth[2]);

    /* New cell of this spart */
    const struct cell *c = &s->cells_top[new_cid];
    const int new_node = c->nodeID;

    if (s_dest[k] != new_node)
      error("spart's new node index not matching sorted index.");

    if (sp->x[0] < c->loc[0] || sp->x[0] > c->loc[0] + c->width[0] ||
        sp->x[1] < c->loc[1] || sp->x[1] > c->loc[1] + c->width[1] ||
        sp->x[2] < c->loc[2] || sp->x[2] > c->loc[2] + c->width[2])
      error("spart not sorted into the right top-level cell!");
  }
#endif

  /* We need to re-link the gpart partners of sparts. */
  if (s->nr_sparts > 0) {
    int current_dest = s_dest[0];
    size_t count_this_dest = 0;
    for (size_t k = 0; k < s->nr_sparts; ++k) {
      if (s->sparts[k].gpart != NULL) {

        /* As the addresses will be invalidated by the communications, we will
         * instead store the absolute index from the start of the sub-array of
         * particles to be sent to a given node.
         * Recall that gparts without partners have a positive id.
         * We will restore the pointers on the receiving node later on. */
        if (s_dest[k] != current_dest) {
          current_dest = s_dest[k];
          count_this_dest = 0;
        }

#ifdef SWIFT_DEBUG_CHECKS
        if (s->sparts[k].gpart->id_or_neg_offset > 0)
          error("Trying to link a partnerless gpart !");
#endif

        s->sparts[k].gpart->id_or_neg_offset = -count_this_dest;
        count_this_dest++;
      }
    }
  }

  free(s_dest);

  /* Get destination of each g-particle */
  int *g_counts;
  if ((g_counts = (int *)malloc(sizeof(int) * nr_nodes * nr_nodes)) == NULL)
    error("Failed to allocate g_gcount temporary buffer.");
  bzero(g_counts, sizeof(int) * nr_nodes * nr_nodes);

  int *g_dest;
  if ((g_dest = (int *)malloc(sizeof(int) * s->nr_gparts)) == NULL)
    error("Failed to allocate g_dest temporary buffer.");

  for (size_t k = 0; k < s->nr_gparts; k++) {

    /* Periodic boundary conditions */
    for (int j = 0; j < 3; j++) {
      if (gparts[k].x[j] < 0.0)
        gparts[k].x[j] += dim[j];
      else if (gparts[k].x[j] >= dim[j])
        gparts[k].x[j] -= dim[j];
    }
    const int cid =
        cell_getid(cdim, gparts[k].x[0] * iwidth[0], gparts[k].x[1] * iwidth[1],
                   gparts[k].x[2] * iwidth[2]);
#ifdef SWIFT_DEBUG_CHECKS
    if (cid < 0 || cid >= s->nr_cells)
      error("Bad cell id %i for gpart %zu at [%.3e,%.3e,%.3e].", cid, k,
            gparts[k].x[0], gparts[k].x[1], gparts[k].x[2]);
#endif

    g_dest[k] = cells[cid].nodeID;

    /* The counts array is indexed as count[from * nr_nodes + to]. */
    g_counts[nodeID * nr_nodes + g_dest[k]] += 1;
  }

  /* Sort the gparticles according to their cell index. */
  if (s->nr_gparts > 0)
    space_gparts_sort(s, g_dest, s->nr_gparts, 0, nr_nodes - 1, e->verbose);

#ifdef SWIFT_DEBUG_CHECKS
  /* Verify that the gpart have been sorted correctly. */
  for (size_t k = 0; k < s->nr_gparts; k++) {
    const struct gpart *gp = &s->gparts[k];

    /* New cell index */
    const int new_cid =
        cell_getid(s->cdim, gp->x[0] * s->iwidth[0], gp->x[1] * s->iwidth[1],
                   gp->x[2] * s->iwidth[2]);

    /* New cell of this gpart */
    const struct cell *c = &s->cells_top[new_cid];
    const int new_node = c->nodeID;

    if (g_dest[k] != new_node)
      error("gpart's new node index not matching sorted index (%d != %d).",
            g_dest[k], new_node);

    if (gp->x[0] < c->loc[0] || gp->x[0] > c->loc[0] + c->width[0] ||
        gp->x[1] < c->loc[1] || gp->x[1] > c->loc[1] + c->width[1] ||
        gp->x[2] < c->loc[2] || gp->x[2] > c->loc[2] + c->width[2])
      error("gpart not sorted into the right top-level cell!");
  }
#endif

  free(g_dest);

  /* Get all the counts from all the nodes. */
  if (MPI_Allreduce(MPI_IN_PLACE, counts, nr_nodes * nr_nodes, MPI_INT, MPI_SUM,
                    MPI_COMM_WORLD) != MPI_SUCCESS)
    error("Failed to allreduce particle transfer counts.");

  /* Get all the s_counts from all the nodes. */
  if (MPI_Allreduce(MPI_IN_PLACE, g_counts, nr_nodes * nr_nodes, MPI_INT,
                    MPI_SUM, MPI_COMM_WORLD) != MPI_SUCCESS)
    error("Failed to allreduce gparticle transfer counts.");

  /* Get all the g_counts from all the nodes. */
  if (MPI_Allreduce(MPI_IN_PLACE, s_counts, nr_nodes * nr_nodes, MPI_INT,
                    MPI_SUM, MPI_COMM_WORLD) != MPI_SUCCESS)
    error("Failed to allreduce sparticle transfer counts.");

  /* Report how many particles will be moved. */
  if (e->verbose) {
    if (e->nodeID == 0) {
      size_t total = 0, g_total = 0, s_total = 0;
      size_t unmoved = 0, g_unmoved = 0, s_unmoved = 0;
      for (int p = 0, r = 0; p < nr_nodes; p++) {
        for (int s = 0; s < nr_nodes; s++) {
          total += counts[r];
          g_total += g_counts[r];
          s_total += s_counts[r];
          if (p == s) {
            unmoved += counts[r];
            g_unmoved += g_counts[r];
            s_unmoved += s_counts[r];
          }
          r++;
        }
      }
      if (total > 0)
        message("%ld of %ld (%.2f%%) of particles moved", total - unmoved,
                total, 100.0 * (double)(total - unmoved) / (double)total);
      if (g_total > 0)
        message("%ld of %ld (%.2f%%) of g-particles moved", g_total - g_unmoved,
                g_total,
                100.0 * (double)(g_total - g_unmoved) / (double)g_total);
      if (s_total > 0)
        message("%ld of %ld (%.2f%%) of s-particles moved", s_total - s_unmoved,
                s_total,
                100.0 * (double)(s_total - s_unmoved) / (double)s_total);
    }
  }

  /* Now each node knows how many parts, sparts and gparts will be transferred
   * to every other node.
   * Get the new numbers of particles for this node. */
  size_t nr_parts = 0, nr_gparts = 0, nr_sparts = 0;
  for (int k = 0; k < nr_nodes; k++) nr_parts += counts[k * nr_nodes + nodeID];
  for (int k = 0; k < nr_nodes; k++)
    nr_gparts += g_counts[k * nr_nodes + nodeID];
  for (int k = 0; k < nr_nodes; k++)
    nr_sparts += s_counts[k * nr_nodes + nodeID];

  /* Now exchange the particles, type by type to keep the memory required
   * under control. */

  /* SPH particles. */
  void *new_parts = engine_do_redistribute(counts, (char *)s->parts, nr_parts,
                                           sizeof(struct part), part_align,
                                           part_mpi_type, nr_nodes, nodeID);
  free(s->parts);
  s->parts = (struct part *)new_parts;
  s->nr_parts = nr_parts;
  s->size_parts = engine_redistribute_alloc_margin * nr_parts;

  /* Extra SPH particle properties. */
  new_parts = engine_do_redistribute(counts, (char *)s->xparts, nr_parts,
                                     sizeof(struct xpart), xpart_align,
                                     xpart_mpi_type, nr_nodes, nodeID);
  free(s->xparts);
  s->xparts = (struct xpart *)new_parts;

  /* Gravity particles. */
  new_parts = engine_do_redistribute(g_counts, (char *)s->gparts, nr_gparts,
                                     sizeof(struct gpart), gpart_align,
                                     gpart_mpi_type, nr_nodes, nodeID);
  free(s->gparts);
  s->gparts = (struct gpart *)new_parts;
  s->nr_gparts = nr_gparts;
  s->size_gparts = engine_redistribute_alloc_margin * nr_gparts;

  /* Star particles. */
  new_parts = engine_do_redistribute(s_counts, (char *)s->sparts, nr_sparts,
                                     sizeof(struct spart), spart_align,
                                     spart_mpi_type, nr_nodes, nodeID);
  free(s->sparts);
  s->sparts = (struct spart *)new_parts;
  s->nr_sparts = nr_sparts;
  s->size_sparts = engine_redistribute_alloc_margin * nr_sparts;

  /* All particles have now arrived. Time for some final operations on the
     stuff we just received */

  /* Restore the part<->gpart and spart<->gpart links */
  size_t offset_parts = 0, offset_sparts = 0, offset_gparts = 0;
  for (int node = 0; node < nr_nodes; ++node) {

    const int ind_recv = node * nr_nodes + nodeID;
    const size_t count_parts = counts[ind_recv];
    const size_t count_gparts = g_counts[ind_recv];
    const size_t count_sparts = s_counts[ind_recv];

    /* Loop over the gparts received from that node */
    for (size_t k = offset_gparts; k < offset_gparts + count_gparts; ++k) {

      /* Does this gpart have a gas partner ? */
      if (s->gparts[k].type == swift_type_gas) {

        const ptrdiff_t partner_index =
            offset_parts - s->gparts[k].id_or_neg_offset;

        /* Re-link */
        s->gparts[k].id_or_neg_offset = -partner_index;
        s->parts[partner_index].gpart = &s->gparts[k];
      }

      /* Does this gpart have a star partner ? */
      if (s->gparts[k].type == swift_type_star) {

        const ptrdiff_t partner_index =
            offset_sparts - s->gparts[k].id_or_neg_offset;

        /* Re-link */
        s->gparts[k].id_or_neg_offset = -partner_index;
        s->sparts[partner_index].gpart = &s->gparts[k];
      }
    }

    offset_parts += count_parts;
    offset_gparts += count_gparts;
    offset_sparts += count_sparts;
  }

  /* Clean up the counts now we done. */
  free(counts);
  free(g_counts);
  free(s_counts);

#ifdef SWIFT_DEBUG_CHECKS
  /* Verify that all parts are in the right place. */
  for (size_t k = 0; k < nr_parts; k++) {
    const int cid =
        cell_getid(cdim, s->parts[k].x[0] * iwidth[0],
                   s->parts[k].x[1] * iwidth[1], s->parts[k].x[2] * iwidth[2]);
    if (cells[cid].nodeID != nodeID)
      error("Received particle (%zu) that does not belong here (nodeID=%i).", k,
            cells[cid].nodeID);
  }
  for (size_t k = 0; k < nr_gparts; k++) {
    const int cid = cell_getid(cdim, s->gparts[k].x[0] * iwidth[0],
                               s->gparts[k].x[1] * iwidth[1],
                               s->gparts[k].x[2] * iwidth[2]);
    if (cells[cid].nodeID != nodeID)
      error("Received g-particle (%zu) that does not belong here (nodeID=%i).",
            k, cells[cid].nodeID);
  }
  for (size_t k = 0; k < nr_sparts; k++) {
    const int cid = cell_getid(cdim, s->sparts[k].x[0] * iwidth[0],
                               s->sparts[k].x[1] * iwidth[1],
                               s->sparts[k].x[2] * iwidth[2]);
    if (cells[cid].nodeID != nodeID)
      error("Received s-particle (%zu) that does not belong here (nodeID=%i).",
            k, cells[cid].nodeID);
  }

  /* Verify that the links are correct */
  part_verify_links(s->parts, s->gparts, s->sparts, nr_parts, nr_gparts,
                    nr_sparts, e->verbose);
#endif

  /* Be verbose about what just happened. */
  if (e->verbose) {
    int my_cells = 0;
    for (int k = 0; k < nr_cells; k++)
      if (cells[k].nodeID == nodeID) my_cells += 1;
    message("node %i now has %zu parts, %zu sparts and %zu gparts in %i cells.",
            nodeID, nr_parts, nr_sparts, nr_gparts, my_cells);
  }

  if (e->verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
#else
  error("SWIFT was not compiled with MPI support.");
#endif
}

/**
 * @brief Repartition the cells amongst the nodes.
 *
 * @param e The #engine.
 */
void engine_repartition(struct engine *e) {

#if defined(WITH_MPI) && defined(HAVE_METIS)

  ticks tic = getticks();

#ifdef SWIFT_DEBUG_CHECKS
  /* Be verbose about this. */
  if (e->nodeID == 0 || e->verbose) message("repartitioning space");
  fflush(stdout);

  /* Check that all cells have been drifted to the current time */
  space_check_drift_point(e->s, e->ti_old,
                          e->policy & engine_policy_self_gravity);
#endif

  /* Clear the repartition flag. */
  e->forcerepart = 0;

  /* Nothing to do if only using a single node. Also avoids METIS
   * bug that doesn't handle this case well. */
  if (e->nr_nodes == 1) return;

  /* Do the repartitioning. */
  partition_repartition(e->reparttype, e->nodeID, e->nr_nodes, e->s,
                        e->sched.tasks, e->sched.nr_tasks);

  /* Now comes the tricky part: Exchange particles between all nodes.
     This is done in two steps, first allreducing a matrix of
     how many particles go from where to where, then re-allocating
     the parts array, and emitting the sends and receives.
     Finally, the space, tasks, and proxies need to be rebuilt. */

  /* Redistribute the particles between the nodes. */
  engine_redistribute(e);

  /* Make the proxies. */
  engine_makeproxies(e);

  /* Tell the engine it should re-build whenever possible */
  e->forcerebuild = 1;

  if (e->verbose)
    message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
#else
  if (e->reparttype->type != REPART_NONE)
    error("SWIFT was not compiled with MPI and METIS support.");
#endif
}

/**
 * @brief Decide whether trigger a repartition the cells amongst the nodes.
 *
 * @param e The #engine.
 */
void engine_repartition_trigger(struct engine *e) {

#ifdef WITH_MPI

  /* Do nothing if there have not been enough steps since the last
   * repartition, don't want to repeat this too often or immediately after
   * a repartition step. */
  if (e->step - e->last_repartition >= 2) {

    /* Old style if trigger is >1 or this is the second step (want an early
     * repartition following the initial repartition). */
    if (e->reparttype->trigger > 1 || e->step == 2) {
      if (e->reparttype->trigger > 1) {