1 /* pmm - double broadcast parallel matrix multiplication */
3 /* Copyright (c) 2006-2016. The SimGrid Team. All rights reserved. */
5 /* This program is free software; you can redistribute it and/or modify it
6 * under the terms of the license (GNU LGPL) which comes with this package. */
8 #include "simgrid/msg.h"
9 #include "xbt/matrix.h"
10 #include "xbt/xbt_os_time.h"
12 XBT_LOG_NEW_DEFAULT_CATEGORY(msg_pmm, "Messages specific for this msg example");
14 /* This example should always be executed using a deployment of GRID_SIZE * GRID_SIZE nodes. */
15 #define GRID_SIZE 3 /* Modify to adjust the grid's size */
16 #define NODE_MATRIX_SIZE 300 /* Amount of work done by each node*/
18 #define GRID_NUM_NODES GRID_SIZE * GRID_SIZE
19 #define MATRIX_SIZE NODE_MATRIX_SIZE * GRID_SIZE
20 #define MAILBOX_NAME_SIZE 10
21 #define NEIGHBOURS_COUNT GRID_SIZE - 1
23 /* The job sent to every node */
24 typedef struct s_node_job{
27 int nodes_in_row[NEIGHBOURS_COUNT];
28 int nodes_in_col[NEIGHBOURS_COUNT];
33 typedef s_node_job_t *node_job_t;
35 /* Structure for recovering results */
36 typedef struct s_result {
42 typedef s_result_t *result_t;
44 int node(int argc, char **argv);
45 static void create_jobs(xbt_matrix_t A, xbt_matrix_t B, node_job_t *jobs);
46 static void broadcast_jobs(node_job_t *jobs);
47 static node_job_t wait_job(int selfid);
48 static void broadcast_matrix(xbt_matrix_t M, int num_nodes, int *nodes);
49 static void get_sub_matrix(xbt_matrix_t *sM, int selfid);
50 static void receive_results(result_t *results);
51 static void task_cleanup(void *arg);
53 int node(int argc, char **argv)
55 char my_mbox[MAILBOX_NAME_SIZE];
57 node_job_t jobs[GRID_NUM_NODES];
58 xbt_matrix_t A = NULL;
59 xbt_matrix_t B = NULL;
60 xbt_matrix_t C = NULL;
63 xbt_assert(argc != 1, "Wrong number of arguments for this node");
65 /* Initialize the node's data-structures */
66 int myid = xbt_str_parse_int(argv[1], "Invalid ID received as first node parameter: %s");
67 snprintf(my_mbox, MAILBOX_NAME_SIZE - 1, "%d", myid);
68 xbt_matrix_t sC = xbt_matrix_double_new_zeros(NODE_MATRIX_SIZE, NODE_MATRIX_SIZE);
71 /* Create the matrices to multiply and one to store the result */
72 A = xbt_matrix_double_new_id(MATRIX_SIZE, MATRIX_SIZE);
73 B = xbt_matrix_double_new_seq(MATRIX_SIZE, MATRIX_SIZE);
74 C = xbt_matrix_double_new_zeros(MATRIX_SIZE, MATRIX_SIZE);
76 /* Create the nodes' jobs */
77 create_jobs(A, B, jobs);
79 /* Get own job first */
82 /* Broadcast the rest of the jobs to the other nodes */
83 broadcast_jobs(jobs + 1);
86 myjob = wait_job(myid);
89 /* Multiplication main-loop */
90 XBT_VERB("Start Multiplication's Main-loop");
91 for (int k=0; k < GRID_SIZE; k++){
96 XBT_VERB("Broadcast sA(%d,%d) to row %d", myjob->row, k, myjob->row);
97 broadcast_matrix(myjob->A, NEIGHBOURS_COUNT, myjob->nodes_in_row);
101 XBT_VERB("Broadcast sB(%d,%d) to col %d", k, myjob->col, myjob->col);
102 broadcast_matrix(myjob->B, NEIGHBOURS_COUNT, myjob->nodes_in_col);
105 if(myjob->row == k && myjob->col == k){
106 xbt_matrix_double_addmult(myjob->A, myjob->B, sC);
107 }else if(myjob->row == k){
108 get_sub_matrix(&sA, myid);
109 xbt_matrix_double_addmult(sA, myjob->B, sC);
111 }else if(myjob->col == k){
112 get_sub_matrix(&sB, myid);
113 xbt_matrix_double_addmult(myjob->A, sB, sC);
116 get_sub_matrix(&sA, myid);
117 get_sub_matrix(&sB, myid);
118 xbt_matrix_double_addmult(sA, sB, sC);
124 /* Node 0: gather the results and reconstruct the final matrix */
127 result_t results[GRID_NUM_NODES] = {0};
129 XBT_VERB("Multiplication done.");
131 /* Get the result from the nodes in the GRID */
132 receive_results(results);
134 /* First add our results */
135 xbt_matrix_copy_values(C, sC, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE, 0, 0, 0, 0, NULL);
137 /* Reconstruct the rest of the result matrix */
138 for (node = 1; node < GRID_NUM_NODES; node++){
139 xbt_matrix_copy_values(C, results[node]->sC, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE,
140 NODE_MATRIX_SIZE * results[node]->row, NODE_MATRIX_SIZE * results[node]->col,
142 xbt_matrix_free(results[node]->sC);
143 xbt_free(results[node]);
146 //xbt_matrix_dump(C, "C:res", 0, xbt_matrix_dump_display_double);
152 /* The rest: return the result to node 0 */
156 XBT_VERB("Multiplication done. Send the sub-result.");
158 result = xbt_new0(s_result_t, 1);
159 result->row = myjob->row;
160 result->col = myjob->col;
161 result->sC = xbt_matrix_new_sub(sC, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE, 0, 0, NULL);
162 task = MSG_task_create("result",100,100,result);
163 MSG_task_send(task, "0");
166 /* Clean up and finish*/
168 xbt_matrix_free(myjob->A);
169 xbt_matrix_free(myjob->B);
175 * Broadcast the jobs to the nodes of the grid (except to node 0)
177 static void broadcast_jobs(node_job_t *jobs)
179 char node_mbox[MAILBOX_NAME_SIZE];
180 msg_comm_t comms[GRID_NUM_NODES - 1] = {0};
182 XBT_VERB("Broadcast Jobs");
183 for (int node = 1; node < GRID_NUM_NODES; node++){
184 msg_task_t task = MSG_task_create("Job", 100, 100, jobs[node-1]);
185 snprintf(node_mbox, MAILBOX_NAME_SIZE - 1, "%d", node);
186 comms[node-1] = MSG_task_isend(task, node_mbox);
189 MSG_comm_waitall(comms, GRID_NUM_NODES-1, -1);
190 for (int node = 1; node < GRID_NUM_NODES; node++)
191 MSG_comm_destroy(comms[node - 1]);
194 static node_job_t wait_job(int selfid)
196 msg_task_t task = NULL;
197 char self_mbox[MAILBOX_NAME_SIZE];
198 snprintf(self_mbox, MAILBOX_NAME_SIZE - 1, "%d", selfid);
199 msg_error_t err = MSG_task_receive(&task, self_mbox);
200 xbt_assert(err == MSG_OK, "Error while receiving from %s (%d)", self_mbox, (int)err);
201 node_job_t job = (node_job_t)MSG_task_get_data(task);
202 MSG_task_destroy(task);
203 XBT_VERB("Got Job (%d,%d)", job->row, job->col);
208 static void broadcast_matrix(xbt_matrix_t M, int num_nodes, int *nodes)
210 char node_mbox[MAILBOX_NAME_SIZE];
212 for(int node=0; node < num_nodes; node++){
213 snprintf(node_mbox, MAILBOX_NAME_SIZE - 1, "%d", nodes[node]);
214 xbt_matrix_t sM = xbt_matrix_new_sub(M, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE, 0, 0, NULL);
215 msg_task_t task = MSG_task_create("sub-matrix", 100, 100, sM);
216 MSG_task_dsend(task, node_mbox, task_cleanup);
217 XBT_DEBUG("sub-matrix sent to %s", node_mbox);
221 static void get_sub_matrix(xbt_matrix_t *sM, int selfid)
223 msg_task_t task = NULL;
224 char node_mbox[MAILBOX_NAME_SIZE];
226 XBT_VERB("Get sub-matrix");
228 snprintf(node_mbox, MAILBOX_NAME_SIZE - 1, "%d", selfid);
229 msg_error_t err = MSG_task_receive(&task, node_mbox);
230 xbt_assert(err == MSG_OK, "Error while receiving from %s (%d)", node_mbox, (int)err);
231 *sM = (xbt_matrix_t)MSG_task_get_data(task);
232 MSG_task_destroy(task);
235 static void task_cleanup(void *arg){
236 msg_task_t task = (msg_task_t)arg;
237 xbt_matrix_t m = (xbt_matrix_t)MSG_task_get_data(task);
239 MSG_task_destroy(task);
242 int main(int argc, char *argv[])
244 MSG_init(&argc, argv);
245 const char *platform = "../../platforms/cluster.xml";
248 MSG_create_environment(platform);
250 MSG_function_register("node", node);
251 for(int i = 0 ; i< 9; i++) {
252 char *hostname = bprintf("node-%d.acme.org", i);
253 char **argvF = xbt_new(char *, 3);
254 argvF[0] = xbt_strdup("node");
255 argvF[1] = bprintf("%d", i);
257 MSG_process_create_with_arguments("node", node, NULL, MSG_host_by_name(hostname), 2, argvF);
261 msg_error_t res = MSG_main();
262 XBT_INFO("Simulated time: %g", MSG_get_clock());
264 return res != MSG_OK;
267 static void create_jobs(xbt_matrix_t A, xbt_matrix_t B, node_job_t *jobs)
272 for (int node = 0; node < GRID_NUM_NODES; node++){
273 XBT_VERB("Create job %d", node);
274 jobs[node] = xbt_new0(s_node_job_t, 1);
275 jobs[node]->row = row;
276 jobs[node]->col = col;
278 /* Compute who are the nodes in the same row and column than the node receiving this job */
280 for (int j = 0; j < GRID_SIZE; j++) {
281 if (node != (GRID_SIZE * row) + j) {
282 jobs[node]->nodes_in_row[k] = (GRID_SIZE * row) + j;
288 for (int j = 0; j < GRID_SIZE; j++) {
289 if (node != (GRID_SIZE * j) + col) {
290 jobs[node]->nodes_in_col[k] = (GRID_SIZE * j) + col;
295 /* Assign a sub matrix of A and B to the job */
297 xbt_matrix_new_sub(A, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE * row, NODE_MATRIX_SIZE * col, NULL);
299 xbt_matrix_new_sub(B, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE, NODE_MATRIX_SIZE * row, NODE_MATRIX_SIZE * col, NULL);
302 if (col >= GRID_SIZE){
309 static void receive_results(result_t *results) {
310 msg_comm_t comms[GRID_NUM_NODES-1] = {0};
311 msg_task_t tasks[GRID_NUM_NODES-1] = {0};
313 XBT_VERB("Receive Results.");
315 /* Get the result from the nodes in the GRID */
316 for (int node = 1; node < GRID_NUM_NODES; node++)
317 comms[node-1] = MSG_task_irecv(&tasks[node-1], "0");
319 MSG_comm_waitall(comms, GRID_NUM_NODES - 1, -1);
320 for (int node = 1; node < GRID_NUM_NODES; node++)
321 MSG_comm_destroy(comms[node - 1]);
323 /* Reconstruct the result matrix */
324 for (int node = 1; node < GRID_NUM_NODES; node++){
325 results[node] = (result_t)MSG_task_get_data(tasks[node-1]);
326 MSG_task_destroy(tasks[node-1]);