implemented various simulation modes (overlapping/non-overlapping...

implemented various simulation modes (overlapping/non-overlapping communication and computation, using sparse collective / p2p / persistent request communication)

cellular_automaton/c/configuration.c

@@ -8,8 +8,8 @@
 const char *verbosity_levels[] = {"OFF", "INFO", "DEBUG", "TRACE"};
 
 const char *communication_computation_modes[] = {
-   "Blocking (No overlap of communication and computation)",
-   "Non-blocking (Overlapping communication and computation)"
+   "No overlap of communication and computation",
+   "Overlapping communication and computation"
 };
 
 const char *transmission_modes[] = {
@@ -117,7 +117,7 @@ void conf_print(const conf_t *c, FILE *f)
    fprintf(f, " * Verbosity level:           %s (%d)\n", verbosity_levels[i], c->verbosity_level);
    fprintf(f, " * Input file:                %s%s\n", c->file_basename, FILE_EXT);
    fprintf(f, " * Transmission mode:         %s\n", transmission_modes[c->transmission_mode]);
-   fprintf(f, " * Communication mode:        %s\n", communication_computation_modes[c->communication_computation_mode]);
+   fprintf(f, " * Overlap mode:              %s communication and computation\n", c->communication_computation_mode == NO_OVERLAP ? "No overlap of" : "Overlapping");
    fprintf(f, " * Grid of processes:         %d x %d\n", c->nprocs[0], c->nprocs[1]);
    fprintf(f, " * Number of iterations:      %d\n", c->n_iterations);
    fprintf(f, " * Generations per iteration: %ld\n", c->n_generations_per_iteration);

cellular_automaton/c/configuration.h

@@ -21,7 +21,7 @@ enum transmission_mode_enum
 enum communication_computation_mode_enum
 {
    NO_OVERLAP = 0,
-   OVERLAP = 1
+   OVERLAP = 4
 };
 
 #define FILE_BASENAME_SZ 1024

cellular_automaton/c/main.c

@@ -23,6 +23,13 @@ int main(int argc, char* argv[])
 
    MPI_Init(&argc, &argv);
 
+   // We want the program to stop on I/O errors
+   // -> Change the default I/O error hander from MPI_ERRORS_RETURN to MPI_ERRORS_ARE_FATAL
+   // Notes:
+   //  * An individual I/O error handler can be associated to each file handle.
+   //  * The default I/O error handler is associated to the null file handle, i.e. MPI_FILE_NULL.
+   MPI_File_set_errhandler(MPI_FILE_NULL, MPI_ERRORS_ARE_FATAL);
+
    // Parse command line arguments
    conf_init_from_args(&config, argc, argv);
    if(info_enabled(&config)) conf_print(&config, stdout);
@@ -67,16 +74,17 @@ void read_input(const conf_t *c, world_t *world) {
       "Global size: %ld x %ld\n\n",
       header_length, global_sizes[0], global_sizes[1]
    );
+   // Initialize cells (determine local tile, allocate memory)
+   MPI_Cart_create(MPI_COMM_WORLD, 2, c->nprocs, periods, allow_reorder, &cart_comm);
+   world_init(world, cart_comm, global_sizes, c);
+   MPI_Comm_free(&cart_comm);
+
    if(debug_enabled(c)) printf(
       "%03d: Local tile: [%ld %ld) x [%ld %ld)\n", rank,
       world->local_start[0], world->local_start[0]+world->local_size[0],
       world->local_start[1], world->local_start[1]+world->local_size[1]
    );
 
-   // Initialize cells (determine local tile, allocate memory)
-   MPI_Cart_create(MPI_COMM_WORLD, 2, c->nprocs, periods, allow_reorder, &cart_comm);
-   world_init(world, cart_comm, global_sizes);
-
    // Collectively read cell data
    file_read_world(file, world, header_length);
    MPI_File_close(&file);
@@ -90,7 +98,7 @@ void iterate(const conf_t *c, world_t *world)
    size_t i, g;
    double total_time, sim_time = 0, io_time = 0;
 
-   char output_filename[FILE_NAME_SZ+8];
+   char output_filename[FILE_NAME_SZ+10];
    MPI_File file;
    size_t header_length;
 
@@ -105,13 +113,13 @@ void iterate(const conf_t *c, world_t *world)
 
       // Run n_gen generations
       sim_time -= MPI_Wtime();
-      do_simulation(world, n_gen); g += n_gen;
+      do_simulation(world, n_gen, c); g += n_gen;
       sim_time += MPI_Wtime();
 
       // Determine output filename
       snprintf(
          output_filename, sizeof(output_filename),
-         "%s+%07ld%s",
+         "%s+%09ld%s",
          c->file_basename, g, FILE_EXT
       );
 
@@ -127,7 +135,7 @@ void iterate(const conf_t *c, world_t *world)
       MPI_File_close(&file);
       io_time += MPI_Wtime();
 
-      if(info_enabled(c)) printf("Done iteration %-4d - Generation %-7ld - written '%s'.\n", i, g, output_filename);
+      if(info_enabled(c)) printf("Generation %-9ld - written '%s'.\n", g, output_filename);
    }
    total_time += MPI_Wtime();
 

cellular_automaton/c/simulation.c

+#include <stdlib.h>
+#include <string.h>
 #include <mpi.h>
 
 #include "simulation.h"
@@ -6,26 +8,98 @@
 #define ELECTRON_TAIL '~'
 #define WIRE '#'
 
+void do_simulation_blocking_sparse(world_t *world, size_t n_generations);
+void do_simulation_nonblocking_sparse(world_t *world, size_t n_generations);
+
+void do_simulation_p2p_overlap(world_t *world, size_t n_generations);
+void do_simulation_p2p_no_overlap(world_t *world, size_t n_generations);
+
+void do_simulation_persistent_request_overlap(world_t *world, size_t n_generations);
+void do_simulation_persistent_request_no_overlap(world_t *world, size_t n_generations);
+
+void wireworld_step_complete(world_t *world);
+void wireworld_step_interior(world_t *world);
+void wireworld_step_boundary(world_t *world);
+
 void wireworld_step(world_t *world, size_t i_start, size_t bx, size_t by);
 
-void do_simulation(world_t *world, size_t n_generations)
+// --------------------------------------------------------------------------
+
+void do_simulation(world_t *world, size_t n_generations, const conf_t *c)
 {
-   const size_t nx = world->local_size[0];
-   const size_t ny = world->local_size[1];
+   const int mode = c->transmission_mode | c->communication_computation_mode;
 
-   const size_t DOWN = nx+2; // (+2 ... for halo cells)
+   switch(mode) {
+      case SPARSE_COLLECTIVE | NO_OVERLAP:
+         do_simulation_blocking_sparse(world, n_generations);
+         break;
 
-   const size_t i_leftupper  = 1  +    DOWN;
-   const size_t i_rightupper = nx +    DOWN;
-   const size_t i_leftlower  = 1  + ny*DOWN;
+      case SPARSE_COLLECTIVE | OVERLAP:
+         do_simulation_nonblocking_sparse(world, n_generations);
+         break;
+
+      case POINT_TO_POINT | NO_OVERLAP:
+         do_simulation_p2p_no_overlap(world, n_generations);
+         break;
 
+      case POINT_TO_POINT | OVERLAP:
+         do_simulation_p2p_overlap(world, n_generations);
+         break;
+
+      case PERSISTENT_REQUEST | NO_OVERLAP:
+         do_simulation_persistent_request_no_overlap(world, n_generations);
+         break;
+
+      case PERSISTENT_REQUEST | OVERLAP:
+         do_simulation_persistent_request_overlap(world, n_generations);
+         break;
+
+      default:
+         fprintf(stderr, "Not yet implemented.\n");
+         MPI_Abort(MPI_COMM_WORLD, EXIT_FAILURE);
+   }
+}
+
+void do_simulation_blocking_sparse(world_t *world, size_t n_generations)
+{
    size_t g;
    char *tmp;
 
    const int counts[] = {1, 1, 1, 1, 1, 1, 1, 1};
    const MPI_Aint displs[] = {0, 0, 0, 0, 0, 0, 0, 0};
-   const transfer_t *transfer = &world->transfer;
-   
+
+   const MPI_Datatype *send_types = world->transfer.send_types,
+                      *recv_types = world->transfer.recv_types;
+
+   MPI_Comm comm = world->transfer.comm;
+
+   for(g = 0; g < n_generations; g++) {
+      tmp = world->cells_prev;
+      world->cells_prev = world->cells_next;
+      world->cells_next = tmp;
+
+      MPI_Neighbor_alltoallw(
+         world->cells_prev, counts, displs, send_types,
+         world->cells_prev, counts, displs, recv_types,
+         comm
+      );
+      wireworld_step_complete(world);
+   }
+}
+
+void do_simulation_nonblocking_sparse(world_t *world, size_t n_generations)
+{
+   size_t g;
+   char *tmp;
+
+   const int counts[] = {1, 1, 1, 1, 1, 1, 1, 1};
+   const MPI_Aint displs[] = {0, 0, 0, 0, 0, 0, 0, 0};
+
+   const MPI_Datatype *send_types = world->transfer.send_types,
+                      *recv_types = world->transfer.recv_types;
+
+   MPI_Comm comm = world->transfer.comm;
+
    MPI_Request request;
 
    for(g = 0; g < n_generations; g++) {
@@ -33,35 +107,190 @@ void do_simulation(world_t *world, size_t n_generations)
       world->cells_prev = world->cells_next;
       world->cells_next = tmp;
 
-      // Start halo exchange
       MPI_Ineighbor_alltoallw(
-         world->cells_prev, counts, displs, transfer->send_types,
-         world->cells_prev, counts, displs, transfer->recv_types,
-         transfer->graph_comm, &request
+         world->cells_prev, counts, displs, send_types,
+         world->cells_prev, counts, displs, recv_types,
+         comm, &request
       );
 
-      // Compute inner region
-      wireworld_step(world, i_leftupper+1+DOWN, nx-2, ny-2);
-
-      // Finish halo exchange
+      wireworld_step_interior(world);
       MPI_Wait(&request, MPI_STATUS_IGNORE);
+      wireworld_step_boundary(world);
+   }
+}
 
-      // Compute boundary regions
-      wireworld_step(world, i_leftupper, nx, 1); // upper
-      wireworld_step(world, i_leftlower, nx, 1); // lower
-      wireworld_step(world, i_leftupper+DOWN, 1, ny-2); // left
-      wireworld_step(world, i_rightupper+DOWN, 1, ny-2); // right
+void do_simulation_p2p_no_overlap(world_t *world, size_t n_generations)
+{
+   const int tag = 0;
+   const size_t n_neighbors = world->transfer.n_neighbors;
 
-/*    Blocking variant:
+   const int *neighbor_ranks = world->transfer.neighbor_ranks;
 
-      MPI_Neighbor_alltoallw(
-         world->cells_prev, counts, displs, transfer->send_types,
-         world->cells_prev, counts, displs, transfer->recv_types,
-         transfer->graph_comm
-      );
-      wireworld_step(world, i_leftupper, nx, ny);
-*/
+   const MPI_Datatype *send_types = world->transfer.send_types,
+                      *recv_types = world->transfer.recv_types;
+
+   MPI_Comm comm = world->transfer.comm;
+
+   size_t g, i;
+   char *tmp;
+   MPI_Request requests[16];
+
+   for(g = 0; g < n_generations; g++) {
+      tmp = world->cells_prev;
+      world->cells_prev = world->cells_next;
+      world->cells_next = tmp;
+
+      for(i = 0; i < n_neighbors; i++) {
+         MPI_Isend(
+            world->cells_prev, 1, send_types[i], neighbor_ranks[i],
+            tag, comm, &requests[2*i]
+         );
+         MPI_Irecv(
+            world->cells_prev, 1, recv_types[i], neighbor_ranks[i],
+            tag, comm, &requests[2*i+1]
+         );
+      }
+      MPI_Waitall(2*n_neighbors, requests, MPI_STATUSES_IGNORE);
+
+      wireworld_step_complete(world);
+   }
+}
+
+void do_simulation_p2p_overlap(world_t *world, size_t n_generations)
+{
+   const int tag = 0;
+   const size_t n_neighbors = world->transfer.n_neighbors;
+
+   const int *neighbor_ranks = world->transfer.neighbor_ranks;
+
+   const MPI_Datatype *send_types = world->transfer.send_types,
+                      *recv_types = world->transfer.recv_types;
+
+   MPI_Comm comm = world->transfer.comm;
+
+   size_t g, i;
+   char *tmp;
+   MPI_Request requests[16];
+
+   for(g = 0; g < n_generations; g++) {
+      tmp = world->cells_prev;
+      world->cells_prev = world->cells_next;
+      world->cells_next = tmp;
+
+      for(i = 0; i < n_neighbors; i++) {
+         MPI_Isend(
+            world->cells_prev, 1, send_types[i], neighbor_ranks[i],
+            tag, comm, &requests[2*i]
+         );
+         MPI_Irecv(
+            world->cells_prev, 1, recv_types[i], neighbor_ranks[i],
+            tag, comm, &requests[2*i+1]
+         );
+      }
+      wireworld_step_interior(world);
+
+      MPI_Waitall(2*n_neighbors, requests, MPI_STATUSES_IGNORE);
+
+      wireworld_step_boundary(world);
+   }
+}
+
+void do_simulation_persistent_request_no_overlap(world_t *world, size_t n_generations)
+{
+   const size_t n_neighbors = world->transfer.n_neighbors;
+   const size_t sz = world_get_storage_size(world);
+
+   size_t g, i;
+   char *tmp = malloc(sz);
+   MPI_Request *requests = world->transfer.persistent_requests;
+
+   for(g = 0; g < n_generations; g++) {
+
+      // Persistent send/receive requests have been initialized with pointer to world->cell_prev.
+      // Swapping pointers here will therefore not work and swapping memory is necessary.
+      //
+      // Possible Optimization:
+      // ----------------------
+      // Alternately using two sets of persistent requests (one for world->cells_prev and world->cells_next).
+      memcpy(tmp, world->cells_prev, sz);
+      memcpy(world->cells_prev, world->cells_next, sz);
+      memcpy(world->cells_next, tmp, sz);
+
+      MPI_Startall(2*n_neighbors, requests);
+      MPI_Waitall(2*n_neighbors, requests, MPI_STATUSES_IGNORE);
+
+      wireworld_step_complete(world);
    }
+   free(tmp);
+}
+
+void do_simulation_persistent_request_overlap(world_t *world, size_t n_generations)
+{
+   const size_t n_neighbors = world->transfer.n_neighbors;
+   const size_t sz = world_get_storage_size(world);
+
+   size_t g, i;
+   char *tmp = malloc(sz);
+
+   MPI_Request *requests = world->transfer.persistent_requests;
+
+   for(g = 0; g < n_generations; g++) {
+
+      // See above note in the '_no_overlap' funciton.
+      memcpy(tmp, world->cells_prev, sz);
+      memcpy(world->cells_prev, world->cells_next, sz);
+      memcpy(world->cells_next, tmp, sz);
+
+      MPI_Startall(2*n_neighbors, requests);
+
+      wireworld_step_interior(world);
+
+      MPI_Waitall(2*n_neighbors, requests, MPI_STATUSES_IGNORE);
+
+      wireworld_step_boundary(world);
+   }
+   free(tmp);
+}
+
+void wireworld_step_complete(world_t *world)
+{
+   const size_t nx = world->local_size[0],
+                ny = world->local_size[1];
+
+   const size_t DOWN = nx+2; // (+2 ... for halo cells)
+
+   const size_t i_leftupper  = 1  +    DOWN;
+
+   wireworld_step(world, i_leftupper, nx, ny);
+}
+
+void wireworld_step_interior(world_t *world)
+{
+   const size_t nx = world->local_size[0],
+                ny = world->local_size[1];
+
+   const size_t DOWN = nx+2; // (+2 ... for halo cells)
+
+   const size_t i_leftupper  = 1  +    DOWN;
+
+   wireworld_step(world, i_leftupper+1+DOWN, nx-2, ny-2);
+}
+
+void wireworld_step_boundary(world_t *world)
+{
+   const size_t nx = world->local_size[0],
+                ny = world->local_size[1];
+
+   const size_t DOWN = nx+2; // (+2 ... for halo cells)
+
+   const size_t i_leftupper  = 1  +    DOWN,
+                i_rightupper = nx +    DOWN,
+                i_leftlower  = 1  + ny*DOWN;
+
+   wireworld_step(world, i_leftupper, nx, 1); // upper
+   wireworld_step(world, i_leftlower, nx, 1); // lower
+   wireworld_step(world, i_leftupper+DOWN, 1, ny-2); // left
+   wireworld_step(world, i_rightupper+DOWN, 1, ny-2); // right
 }
 
 void wireworld_step(world_t *world, size_t i_start, size_t bx, size_t by)

cellular_automaton/c/simulation.h

@@ -2,7 +2,8 @@
 #define _AUTOMATON_H_
 
 #include "world.h"
+#include "configuration.h"
 
-void do_simulation(world_t *world, size_t n_generations);
+void do_simulation(world_t *world, size_t n_generations, const conf_t *c);
 
 #endif

cellular_automaton/c/world.c

@@ -6,10 +6,13 @@
 void world_init_io_type(world_t *world);
 void world_free_io_type(world_t *world);
 
-void world_init_neighborhood(world_t *world, MPI_Comm cart_comm, int nprocs[], int proc_coord[]);
+void world_init_neighborhood(world_t *world, MPI_Comm cart_comm, int nprocs[], int proc_coord[], const conf_t *c);
 void world_free_neighborhood(world_t *world);
 
-void world_init(world_t *world, MPI_Comm cart_comm, size_t *global_size)
+void world_init_persistent_requests(world_t *world, const conf_t *c);
+void world_free_persistent_requests(world_t *world);
+
+void world_init(world_t *world, MPI_Comm cart_comm, size_t *global_size, const conf_t *c)
 {
    int dim, lo, hi;
    int nprocs[2], periods[2], proc_coord[2];
@@ -34,13 +37,15 @@ void world_init(world_t *world, MPI_Comm cart_comm, size_t *global_size)
    memset(world->cells_next, ' ', storage_size);
 
    world_init_io_type(world);
-   world_init_neighborhood(world, cart_comm, nprocs, proc_coord);
+   world_init_neighborhood(world, cart_comm, nprocs, proc_coord, c);
+   world_init_persistent_requests(world, c);
 }
 
 void world_free(world_t *world)
 {
    world_free_io_type(world);
    world_free_neighborhood(world);
+   world_free_persistent_requests(world);
 
    free(world->cells_prev);
    free(world->cells_next);
@@ -49,10 +54,12 @@ void world_free(world_t *world)
 }
 
 void world_init_io_type(world_t *world) {
-   const int nx = world->local_size[0], ny = world->local_size[1];
-   const int sizes[] = {nx+2, ny+2};
-   const int subsizes[] = {nx, ny};
-   const int starts[] = {1, 1};
+   const int nx = world->local_size[0],
+             ny = world->local_size[1];
+
+   const int sizes[] = {nx+2, ny+2},
+             subsizes[] = {nx, ny},
+             starts[] = {1, 1};
 
    MPI_Type_create_subarray(2,
       sizes, subsizes, starts,
@@ -66,7 +73,7 @@ void world_free_io_type(world_t *world)
    MPI_Type_free(&world->transfer.io_type);
 }
 
-void world_init_neighborhood(world_t *world, MPI_Comm cart_comm, int nprocs[], int proc_coord[])
+void world_init_neighborhood(world_t *world, MPI_Comm cart_comm, int nprocs[], int proc_coord[], const conf_t *c)
 {
    const int px = proc_coord[0],
              py = proc_coord[1];
@@ -98,7 +105,7 @@ void world_init_neighborhood(world_t *world, MPI_Comm cart_comm, int nprocs[], i
 
    size_t i, n;
    const int sizes[] = {nx+2, ny+2};
-   int neighbor_ranks[8];
+   int *neighbor_ranks = world->transfer.neighbor_ranks;
    int weights[8];