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 <sys/stat.h>
#include <sys/types.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 "black_holes_properties.h"
#include "cell.h"
#include "chemistry.h"
#include "clocks.h"
#include "cooling.h"
#include "cosmology.h"
#include "cycle.h"
#include "debug.h"
#include "entropy_floor.h"
#include "equation_of_state.h"
#include "error.h"
#include "feedback.h"
#include "fof.h"
#include "gravity.h"
#include "gravity_cache.h"
#include "hydro.h"
#include "logger.h"
#include "logger_io.h"
#include "map.h"
#include "memuse.h"
#include "minmax.h"
#include "mpiuse.h"
#include "outputlist.h"
#include "parallel_io.h"
#include "part.h"
#include "partition.h"
#include "profiler.h"
#include "proxy.h"
#include "restart.h"
#include "runner.h"
#include "serial_io.h"
#include "single_io.h"
#include "sort_part.h"
#include "star_formation.h"
#include "star_formation_logger.h"
#include "stars_io.h"
#include "statistics.h"
#include "timers.h"
#include "tools.h"
#include "units.h"
#include "velociraptor_interface.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",
                                     "cosmological integration",
                                     "drift everything",
                                     "reconstruct multi-poles",
                                     "temperature",
                                     "cooling",
                                     "stars",
                                     "structure finding",
                                     "star formation",
                                     "feedback",
                                     "black holes",
                                     "fof search",
                                     "time-step limiter",
                                     "time-step sync",
                                     "logger"
};

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

/** The current step of the engine as a global variable (for messages). */
int engine_current_step;

extern int engine_max_parts_per_ghost;
extern int engine_max_sparts_per_ghost;

/**
 * @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) {

#ifdef SWIFT_DEBUG_CHECKS
  if (t == NULL) {
    error("Trying to link NULL task.");
  }
#endif

  /* Get the next free link. */
  const size_t ind = atomic_inc(&e->nr_links);
  if (ind >= e->size_links) {
    error(
        "Link table overflow. Increase the value of "
        "`Scheduler:links_per_tasks`.");
  }
  struct link *res = &e->links[ind];

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

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

#if defined(WITH_MPI) && (defined(HAVE_PARMETIS) || 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_current, /*check_multipoles=*/0);
#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;

  /* Generate the fixed costs include file. */
  if (e->step > 3 && e->reparttype->trigger <= 1.f) {
    task_dump_stats("partition_fixed_costs.h", e, /* header = */ 1,
                    /* allranks = */ 1);
  }

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

  /* Partitioning requires copies of the particles, so we need to reduce the
   * memory in use to the minimum, we can free the sorting indices and the
   * tasks as these will be regenerated at the next rebuild. Also the foreign
   * particle arrays can go as these will be regenerated in proxy exchange. */

  /* Sorting indices. */
  if (e->s->cells_top != NULL) space_free_cells(e->s);

  /* Task arrays. */
  scheduler_free_tasks(&e->sched);

  /* Foreign parts. (no need to nullify the cell pointers as the cells
   * will be regenerated) */
  space_free_foreign_parts(e->s, /*clear_cell_pointers=*/0);

  /* 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;

  /* Flag that a repartition has taken place */
  e->step_props |= engine_step_prop_repartition;

  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 or ParMETIS support.");

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

  const ticks tic = getticks();

  /* Do nothing if there have not been enough steps since the last repartition
   * as we don't want to repeat this too often or immediately after a
   * repartition step. Also nothing to do when requested. */
  if (e->step - e->last_repartition >= 2 &&
      e->reparttype->type != REPART_NONE) {

    /* If we have fixed costs available and this is step 2 or we are forcing
     * repartitioning then we do a fixed costs one now. */
    if (e->reparttype->trigger > 1 ||
        (e->step == 2 && e->reparttype->use_fixed_costs)) {

      if (e->reparttype->trigger > 1) {
        if ((e->step % (int)e->reparttype->trigger) == 0) e->forcerepart = 1;
      } else {
        e->forcerepart = 1;
      }
      e->reparttype->use_ticks = 0;

    } else {

      /* It is only worth checking the CPU loads when we have processed a
       * significant number of all particles as we require all tasks to have
       * timings. */
      if ((e->updates > 1 &&
           e->updates >= e->total_nr_parts * e->reparttype->minfrac) ||
          (e->g_updates > 1 &&
           e->g_updates >= e->total_nr_gparts * e->reparttype->minfrac)) {

        /* Should we are use the task timings or fixed costs. */
        if (e->reparttype->use_fixed_costs > 1) {
          e->reparttype->use_ticks = 0;
        } else {
          e->reparttype->use_ticks = 1;
        }

        /* Get CPU time used since the last call to this function. */
        double elapsed_cputime =
            clocks_get_cputime_used() - e->cputime_last_step;

        /* Gather the elapsed CPU times from all ranks for the last step. */
        double elapsed_cputimes[e->nr_nodes];
        MPI_Gather(&elapsed_cputime, 1, MPI_DOUBLE, elapsed_cputimes, 1,
                   MPI_DOUBLE, 0, MPI_COMM_WORLD);
        if (e->nodeID == 0) {

          /* Get the range and mean of cputimes. */
          double mintime = elapsed_cputimes[0];
          double maxtime = elapsed_cputimes[0];
          double sum = elapsed_cputimes[0];
          for (int k = 1; k < e->nr_nodes; k++) {
            if (elapsed_cputimes[k] > maxtime) maxtime = elapsed_cputimes[k];
            if (elapsed_cputimes[k] < mintime) mintime = elapsed_cputimes[k];
            sum += elapsed_cputimes[k];
          }
          double mean = sum / (double)e->nr_nodes;

          /* Are we out of balance? */
          double abs_trigger = fabs(e->reparttype->trigger);
          if (((maxtime - mintime) / mean) > abs_trigger) {
            if (e->verbose)
              message("trigger fraction %.3f > %.3f will repartition",
                      (maxtime - mintime) / mean, abs_trigger);
            e->forcerepart = 1;
          } else {
            if (e->verbose)
              message("trigger fraction %.3f =< %.3f will not repartition",
                      (maxtime - mintime) / mean, abs_trigger);
          }
        }
      }

      /* All nodes do this together. */
      MPI_Bcast(&e->forcerepart, 1, MPI_INT, 0, MPI_COMM_WORLD);
    }

    /* Remember we did this. */
    if (e->forcerepart) e->last_repartition = e->step;
  }

  /* We always reset CPU time for next check, unless it will not be used. */
  if (e->reparttype->type != REPART_NONE)
    e->cputime_last_step = clocks_get_cputime_used();

  if (e->verbose)
    message("took %.3f %s", clocks_from_ticks(getticks() - tic),
            clocks_getunit());
#endif
}

/**
 * @brief Exchange cell structures with other nodes.
 *
 * @param e The #engine.
 */
void engine_exchange_cells(struct engine *e) {

#ifdef WITH_MPI

  const int with_gravity = e->policy & engine_policy_self_gravity;
  const ticks tic = getticks();

  /* Exchange the cell structure with neighbouring ranks. */
  proxy_cells_exchange(e->proxies, e->nr_proxies, e->s, with_gravity);

  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 Exchange straying particles with other nodes.
 *
 * @param e The #engine.
 * @param offset_parts The index in the parts array as of which the foreign
 *        parts reside (i.e. the current number of local #part).
 * @param ind_part The foreign #cell ID of each part.
 * @param Npart The number of stray parts, contains the number of parts received
 *        on return.
 * @param offset_gparts The index in the gparts array as of which the foreign
 *        parts reside (i.e. the current number of local #gpart).
 * @param ind_gpart The foreign #cell ID of each gpart.
 * @param Ngpart The number of stray gparts, contains the number of gparts
 *        received on return.
 * @param offset_sparts The index in the sparts array as of which the foreign
 *        parts reside (i.e. the current number of local #spart).
 * @param ind_spart The foreign #cell ID of each spart.
 * @param Nspart The number of stray sparts, contains the number of sparts
 *        received on return.
 * @param offset_bparts The index in the bparts array as of which the foreign
 *        parts reside (i.e. the current number of local #bpart).
 * @param ind_bpart The foreign #cell ID of each bpart.
 * @param Nbpart The number of stray bparts, contains the number of bparts
 *        received on return.
 *
 * Note that this function does not mess-up the linkage between parts and
 * gparts, i.e. the received particles have correct linkeage.
 */
void engine_exchange_strays(struct engine *e, const size_t offset_parts,
                            const int *ind_part, size_t *Npart,
                            const size_t offset_gparts, const int *ind_gpart,
                            size_t *Ngpart, const size_t offset_sparts,
                            const int *ind_spart, size_t *Nspart,
                            const size_t offset_bparts, const int *ind_bpart,
                            size_t *Nbpart) {

#ifdef WITH_MPI
  struct space *s = e->s;
  ticks tic = getticks();

  /* Re-set the proxies. */
  for (int k = 0; k < e->nr_proxies; k++) {
    e->proxies[k].nr_parts_out = 0;
    e->proxies[k].nr_gparts_out = 0;
    e->proxies[k].nr_sparts_out = 0;
    e->proxies[k].nr_bparts_out = 0;
  }

  /* Put the parts into the corresponding proxies. */
  for (size_t k = 0; k < *Npart; k++) {

    /* Ignore the particles we want to get rid of (inhibited, ...). */
    if (ind_part[k] == -1) continue;

    /* Get the target node and proxy ID. */
    const int node_id = e->s->cells_top[ind_part[k]].nodeID;
    if (node_id < 0 || node_id >= e->nr_nodes)
      error("Bad node ID %i.", node_id);
    const int pid = e->proxy_ind[node_id];
    if (pid < 0) {
      error(
          "Do not have a proxy for the requested nodeID %i for part with "
          "id=%lld, x=[%e,%e,%e].",
          node_id, s->parts[offset_parts + k].id,
          s->parts[offset_parts + k].x[0], s->parts[offset_parts + k].x[1],
          s->parts[offset_parts + k].x[2]);
    }

    /* Re-link the associated gpart with the buffer offset of the part. */
    if (s->parts[offset_parts + k].gpart != NULL) {
      s->parts[offset_parts + k].gpart->id_or_neg_offset =
          -e->proxies[pid].nr_parts_out;
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (s->parts[offset_parts + k].time_bin == time_bin_inhibited)
      error("Attempting to exchange an inhibited particle");
#endif

    /* Load the part and xpart into the proxy. */
    proxy_parts_load(&e->proxies[pid], &s->parts[offset_parts + k],
                     &s->xparts[offset_parts + k], 1);

#ifdef WITH_LOGGER
    if (e->policy & engine_policy_logger) {
      const uint32_t logger_flag = logger_pack_flags_and_data(
        logger_flag_mpi_exit, node_id);

      /* Log the particle when leaving a rank. */
      logger_log_part(e->logger, &s->parts[offset_parts + k],
                      &s->xparts[offset_parts + k],
                      logger_masks_all_part |
                      logger_mask_data[logger_special_flags].mask,
                      logger_flag);
    }
#endif
  }

  /* Put the sparts into the corresponding proxies. */
  for (size_t k = 0; k < *Nspart; k++) {

    /* Ignore the particles we want to get rid of (inhibited, ...). */
    if (ind_spart[k] == -1) continue;

    /* Get the target node and proxy ID. */
    const int node_id = e->s->cells_top[ind_spart[k]].nodeID;
    if (node_id < 0 || node_id >= e->nr_nodes)
      error("Bad node ID %i.", node_id);
    const int pid = e->proxy_ind[node_id];
    if (pid < 0) {
      error(
          "Do not have a proxy for the requested nodeID %i for part with "
          "id=%lld, x=[%e,%e,%e].",
          node_id, s->sparts[offset_sparts + k].id,
          s->sparts[offset_sparts + k].x[0], s->sparts[offset_sparts + k].x[1],
          s->sparts[offset_sparts + k].x[2]);
    }

    /* Re-link the associated gpart with the buffer offset of the spart. */
    if (s->sparts[offset_sparts + k].gpart != NULL) {
      s->sparts[offset_sparts + k].gpart->id_or_neg_offset =
          -e->proxies[pid].nr_sparts_out;
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (s->sparts[offset_sparts + k].time_bin == time_bin_inhibited)
      error("Attempting to exchange an inhibited particle");
#endif

    /* Load the spart into the proxy */
    proxy_sparts_load(&e->proxies[pid], &s->sparts[offset_sparts + k], 1);

#ifdef WITH_LOGGER
    if (e->policy & engine_policy_logger) {
      const uint32_t logger_flag = logger_pack_flags_and_data(
        logger_flag_mpi_exit, node_id);

      /* Log the particle when leaving a rank. */
      logger_log_spart(e->logger, &s->sparts[offset_sparts + k],
                       logger_masks_all_spart |
                       logger_mask_data[logger_special_flags].mask,
                       logger_flag);
    }
#endif
  }

  /* Put the bparts into the corresponding proxies. */
  for (size_t k = 0; k < *Nbpart; k++) {

    /* Ignore the particles we want to get rid of (inhibited, ...). */
    if (ind_bpart[k] == -1) continue;

    /* Get the target node and proxy ID. */
    const int node_id = e->s->cells_top[ind_bpart[k]].nodeID;
    if (node_id < 0 || node_id >= e->nr_nodes)
      error("Bad node ID %i.", node_id);
    const int pid = e->proxy_ind[node_id];
    if (pid < 0) {
      error(
          "Do not have a proxy for the requested nodeID %i for part with "
          "id=%lld, x=[%e,%e,%e].",
          node_id, s->bparts[offset_bparts + k].id,
          s->bparts[offset_bparts + k].x[0], s->bparts[offset_bparts + k].x[1],
          s->bparts[offset_bparts + k].x[2]);
    }

    /* Re-link the associated gpart with the buffer offset of the bpart. */
    if (s->bparts[offset_bparts + k].gpart != NULL) {
      s->bparts[offset_bparts + k].gpart->id_or_neg_offset =
          -e->proxies[pid].nr_bparts_out;
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (s->bparts[offset_bparts + k].time_bin == time_bin_inhibited)
      error("Attempting to exchange an inhibited particle");
#endif

    /* Load the bpart into the proxy */
    proxy_bparts_load(&e->proxies[pid], &s->bparts[offset_bparts + k], 1);

#ifdef WITH_LOGGER
    if (e->policy & engine_policy_logger) {
      error("Not yet implemented.");
    }
#endif
  }

  /* Put the gparts into the corresponding proxies. */
  for (size_t k = 0; k < *Ngpart; k++) {

    /* Ignore the particles we want to get rid of (inhibited, ...). */
    if (ind_gpart[k] == -1) continue;

    /* Get the target node and proxy ID. */
    const int node_id = e->s->cells_top[ind_gpart[k]].nodeID;
    if (node_id < 0 || node_id >= e->nr_nodes)
      error("Bad node ID %i.", node_id);
    const int pid = e->proxy_ind[node_id];
    if (pid < 0) {
      error(
          "Do not have a proxy for the requested nodeID %i for part with "
          "id=%lli, x=[%e,%e,%e].",
          node_id, s->gparts[offset_gparts + k].id_or_neg_offset,
          s->gparts[offset_gparts + k].x[0], s->gparts[offset_gparts + k].x[1],
          s->gparts[offset_gparts + k].x[2]);
    }

#ifdef SWIFT_DEBUG_CHECKS
    if (s->gparts[offset_gparts + k].time_bin == time_bin_inhibited)
      error("Attempting to exchange an inhibited particle");
#endif

    /* Load the gpart into the proxy */
    proxy_gparts_load(&e->proxies[pid], &s->gparts[offset_gparts + k], 1);

#ifdef WITH_LOGGER
    /* Write only the dark matter particles */
    if ((e->policy & engine_policy_logger) &&
        s->gparts[offset_gparts + k].type == swift_type_dark_matter) {

      const uint32_t logger_flag = logger_pack_flags_and_data(
         logger_flag_mpi_exit, node_id);

      /* Log the particle when leaving a rank. */
      logger_log_gpart(e->logger, &s->gparts[offset_gparts + k],
                       logger_masks_all_gpart |
                       logger_mask_data[logger_special_flags].mask,
                       logger_flag);
    }
#endif
  }

  /* Launch the proxies. */
  MPI_Request reqs_in[5 * engine_maxproxies];
  MPI_Request reqs_out[5 * engine_maxproxies];
  for (int k = 0; k < e->nr_proxies; k++) {
    proxy_parts_exchange_first(&e->proxies[k]);
    reqs_in[k] = e->proxies[k].req_parts_count_in;
    reqs_out[k] = e->proxies[k].req_parts_count_out;
  }

  /* Wait for each count to come in and start the recv. */
  for (int k = 0; k < e->nr_proxies; k++) {
    int pid = MPI_UNDEFINED;
    if (MPI_Waitany(e->nr_proxies, reqs_in, &pid, MPI_STATUS_IGNORE) !=
            MPI_SUCCESS ||
        pid == MPI_UNDEFINED)
      error("MPI_Waitany failed.");
    // message( "request from proxy %i has arrived." , pid );
    proxy_parts_exchange_second(&e->proxies[pid]);
  }

  /* Wait for all the sends to have finished too. */
  if (MPI_Waitall(e->nr_proxies, reqs_out, MPI_STATUSES_IGNORE) != MPI_SUCCESS)
    error("MPI_Waitall on sends failed.");

  /* Count the total number of incoming particles and make sure we have
     enough space to accommodate them. */
  int count_parts_in = 0;
  int count_gparts_in = 0;
  int count_sparts_in = 0;
  int count_bparts_in = 0;
  for (int k = 0; k < e->nr_proxies; k++) {
    count_parts_in += e->proxies[k].nr_parts_in;
    count_gparts_in += e->proxies[k].nr_gparts_in;
    count_sparts_in += e->proxies[k].nr_sparts_in;
    count_bparts_in += e->proxies[k].nr_bparts_in;
  }
  if (e->verbose) {
    message(
        "sent out %zu/%zu/%zu/%zu parts/gparts/sparts/bparts, got %i/%i/%i/%i "
        "back.",
        *Npart, *Ngpart, *Nspart, *Nbpart, count_parts_in, count_gparts_in,
        count_sparts_in, count_bparts_in);
  }

  /* Reallocate the particle arrays if necessary */
  if (offset_parts + count_parts_in > s->size_parts) {
    s->size_parts = (offset_parts + count_parts_in) * engine_parts_size_grow;
    struct part *parts_new = NULL;
    struct xpart *xparts_new = NULL;
    if (swift_memalign("parts", (void **)&parts_new, part_align,
                       sizeof(struct part) * s->size_parts) != 0 ||
        swift_memalign("xparts", (void **)&xparts_new, xpart_align,
                       sizeof(struct xpart) * s->size_parts) != 0)
      error("Failed to allocate new part data.");
    memcpy(parts_new, s->parts, sizeof(struct part) * offset_parts);
    memcpy(xparts_new, s->xparts, sizeof(struct xpart) * offset_parts);
    swift_free("parts", s->parts);
    swift_free("xparts", s->xparts);
    s->parts = parts_new;
    s->xparts = xparts_new;

    /* Reset the links */
    for (size_t k = 0; k < offset_parts; k++) {
      if (s->parts[k].gpart != NULL) {
        s->parts[k].gpart->id_or_neg_offset = -k;
      }
    }
  }

  if (offset_sparts + count_sparts_in > s->size_sparts) {
    s->size_sparts = (offset_sparts + count_sparts_in) * engine_parts_size_grow;
    struct spart *sparts_new = NULL;
    if (swift_memalign("sparts", (void **)&sparts_new, spart_align,
                       sizeof(struct spart) * s->size_sparts) != 0)
      error("Failed to allocate new spart data.");
    memcpy(sparts_new, s->sparts, sizeof(struct spart) * offset_sparts);
    swift_free("sparts", s->sparts);
    s->sparts = sparts_new;

    /* Reset the links */
    for (size_t k = 0; k < offset_sparts; k++) {
      if (s->sparts[k].gpart != NULL) {
        s->sparts[k].gpart->id_or_neg_offset = -k;
      }
    }
  }

  if (offset_bparts + count_bparts_in > s->size_bparts) {
    s->size_bparts = (offset_bparts + count_bparts_in) * engine_parts_size_grow;
    struct bpart *bparts_new = NULL;
    if (swift_memalign("bparts", (void **)&bparts_new, bpart_align,
                       sizeof(struct bpart) * s->size_bparts) != 0)
      error("Failed to allocate new bpart data.");
    memcpy(bparts_new, s->bparts, sizeof(struct bpart) * offset_bparts);
    swift_free("bparts", s->bparts);
    s->bparts = bparts_new;

    /* Reset the links */
    for (size_t k = 0; k < offset_bparts; k++) {
      if (s->bparts[k].gpart != NULL) {
        s->bparts[k].gpart->id_or_neg_offset = -k;
      }
    }
  }

  if (offset_gparts + count_gparts_in > s->size_gparts) {
    s->size_gparts = (offset_gparts + count_gparts_in) * engine_parts_size_grow;
    struct gpart *gparts_new = NULL;
    if (swift_memalign("gparts", (void **)&gparts_new, gpart_align,
                       sizeof(struct gpart) * s->size_gparts) != 0)
      error("Failed to allocate new gpart data.");
    memcpy(gparts_new, s->gparts, sizeof(struct gpart) * offset_gparts);
    swift_free("gparts", s->gparts);
    s->gparts = gparts_new;

    /* Reset the links */
    for (size_t k = 0; k < offset_gparts; k++) {
      if (s->gparts[k].type == swift_type_gas) {
        s->parts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
      } else if (s->gparts[k].type == swift_type_stars) {
        s->sparts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
      } else if (s->gparts[k].type == swift_type_black_hole) {
        s->bparts[-s->gparts[k].id_or_neg_offset].gpart = &s->gparts[k];
      }
    }
  }

  /* Collect the requests for the particle data from the proxies. */
  int nr_in = 0, nr_out = 0;
  for (int k = 0; k < e->nr_proxies; k++) {
    if (e->proxies[k].nr_parts_in > 0) {
      reqs_in[5 * k] = e->proxies[k].req_parts_in;
      reqs_in[5 * k + 1] = e->proxies[k].req_xparts_in;
      nr_in += 2;
    } else {
      reqs_in[5 * k] = reqs_in[5 * k + 1] = MPI_REQUEST_NULL;
    }
    if (e->proxies[k].nr_gparts_in > 0) {
      reqs_in[5 * k + 2] = e->proxies[k].req_gparts_in;
      nr_in += 1;
    } else {
      reqs_in[5 * k + 2] = MPI_REQUEST_NULL;
    }
    if (e->proxies[k].nr_sparts_in > 0) {
      reqs_in[5 * k + 3] = e->proxies[k].req_sparts_in;
      nr_in += 1;
    } else {
      reqs_in[5 * k + 3] = MPI_REQUEST_NULL;
    }
    if (e->proxies[k].nr_bparts_in > 0) {
      reqs_in[5 * k + 4] = e->proxies[k].req_bparts_in;
      nr_in += 1;
    } else {
      reqs_in[5 * k + 4] = MPI_REQUEST_NULL;
    }

    if (e->proxies[k].nr_parts_out > 0) {
      reqs_out[5 * k] = e->proxies[k].req_parts_out;
      reqs_out[5 * k + 1] = e->proxies[k].req_xparts_out;
      nr_out += 2;
    } else {
      reqs_out[5 * k] = reqs_out[5 * k + 1] = MPI_REQUEST_NULL;
    }
    if (e->proxies[k].nr_gparts_out > 0) {
      reqs_out[5 * k + 2] = e->proxies[k].req_gparts_out;
      nr_out += 1;
    } else {
      reqs_out[5 * k + 2] = MPI_REQUEST_NULL;
    }
    if (e->proxies[k].nr_sparts_out > 0) {
      reqs_out[5 * k + 3] = e->proxies[k].req_sparts_out;
      nr_out += 1;
    } else {
      reqs_out[5 * k + 3] = MPI_REQUEST_NULL;
    }
    if (e->proxies[k].nr_bparts_out > 0) {
      reqs_out[5 * k + 4] = e->proxies[k].req_bparts_out;
      nr_out += 1;
    } else {
      reqs_out[5 * k + 4] = MPI_REQUEST_NULL;
    }
  }

  /* Wait for each part array to come in and collect the new
     parts from the proxies. */
  int count_parts = 0, count_gparts = 0, count_sparts = 0, count_bparts = 0;
  for (int k = 0; k < nr_in; k++) {
    int err, pid;
    if ((err = MPI_Waitany(5 * e->nr_proxies, reqs_in, &pid,
                           MPI_STATUS_IGNORE)) != MPI_SUCCESS) {
      char buff[MPI_MAX_ERROR_STRING];
      int res;
      MPI_Error_string(err, buff, &res);
      error("MPI_Waitany failed (%s).", buff);
    }
    if (pid == MPI_UNDEFINED) break;
    // message( "request from proxy %i has arrived." , pid / 5 );
    pid = 5 * (pid / 5);

    /* If all the requests for a given proxy have arrived... */
    if (reqs_in[pid + 0] == MPI_REQUEST_NULL &&
        reqs_in[pid + 1] == MPI_REQUEST_NULL &&
        reqs_in[pid + 2] == MPI_REQUEST_NULL &&
        reqs_in[pid + 3] == MPI_REQUEST_NULL &&
        reqs_in[pid + 4] == MPI_REQUEST_NULL) {
      /* Copy the particle data to the part/xpart/gpart arrays. */
      struct proxy *prox = &e->proxies[pid / 5];
      memcpy(&s->parts[offset_parts + count_parts], prox->parts_in,
             sizeof(struct part) * prox->nr_parts_in);
      memcpy(&s->xparts[offset_parts + count_parts], prox->xparts_in,
             sizeof(struct xpart) * prox->nr_parts_in);
      memcpy(&s->gparts[offset_gparts + count_gparts], prox->gparts_in,
             sizeof(struct gpart) * prox->nr_gparts_in);
      memcpy(&s->sparts[offset_sparts + count_sparts], prox->sparts_in,
             sizeof(struct spart) * prox->nr_sparts_in);
      memcpy(&s->bparts[offset_bparts + count_bparts], prox->bparts_in,
             sizeof(struct bpart) * prox->nr_bparts_in);

#ifdef WITH_LOGGER
      if (e->policy & engine_policy_logger) {
        const uint32_t flag = logger_pack_flags_and_data(logger_flag_mpi_enter,
                                              prox->nodeID);

        struct part *parts = &s->parts[offset_parts + count_parts];
        struct xpart *xparts = &s->xparts[offset_parts + count_parts];
        struct spart *sparts = &s->sparts[offset_sparts + count_sparts];
        struct gpart *gparts = &s->gparts[offset_gparts + count_gparts];

        /* Log the gas particles */
        logger_log_parts(e->logger, parts, xparts,
                         prox->nr_parts_in, logger_masks_all_part |
                         logger_mask_data[logger_special_flags].mask,
                         flag);

        /* Log the stellar particles */
        logger_log_sparts(e->logger, sparts, prox->nr_sparts_in,
                          logger_masks_all_spart |
                          logger_mask_data[logger_special_flags].mask,
                          flag);

        /* Log the gparts */
        logger_log_gparts(e->logger, gparts, prox->nr_gparts_in,
                          logger_masks_all_gpart |
                          logger_mask_data[logger_special_flags].mask,
                          flag);

        /* Log the bparts */
        if (prox->nr_bparts_in > 0) {
          error("TODO");
        }
      }
#endif