1 #include "colls_private.h"
3 static int bcast_NTSL_segment_size_in_byte = 8192;
5 #define HEADER_SIZE 1024
8 /* Non-topology-specific pipelined linear-bcast function */
9 int smpi_coll_tuned_bcast_arrival_nb(void *buf, int count,
10 MPI_Datatype datatype, int root,
13 int tag = -COLL_TAG_BCAST;
16 MPI_Request *send_request_array;
17 MPI_Request *recv_request_array;
18 MPI_Status *send_status_array;
19 MPI_Status *recv_status_array;
21 MPI_Status temp_status_array[MAX_NODE];
28 int flag_array[MAX_NODE];
29 int already_sent[MAX_NODE];
30 int to_clean[MAX_NODE];
31 int header_buf[HEADER_SIZE];
32 char temp_buf[MAX_NODE];
35 extent = smpi_datatype_get_extent(datatype);
42 rank = smpi_comm_rank(comm);
43 size = smpi_comm_size(comm);
46 /* segment is segment size in number of elements (not bytes) */
47 int segment = bcast_NTSL_segment_size_in_byte / extent;
50 int pipe_length = count / segment;
52 /* use for buffer offset for sending and receiving data = segment size in byte */
53 int increment = segment * extent;
55 /* if the input size is not divisible by segment size =>
56 the small remainder will be done with native implementation */
57 int remainder = count % segment;
59 /* if root is not zero send to rank zero first
60 this can be modified to make it faster by using logical src, dst.
64 smpi_mpi_send(buf, count, datatype, 0, tag, comm);
65 } else if (rank == 0) {
66 smpi_mpi_recv(buf, count, datatype, root, tag, comm, &status);
70 /* value == 0 means root has not send data (or header) to the node yet */
71 for (i = 0; i < MAX_NODE; i++) {
77 /* when a message is smaller than a block size => no pipeline */
78 if (count <= segment) {
82 while (sent_count < (size - 1)) {
84 // for (j=0;j<1000;j++) {
85 for (i = 1; i < size; i++) {
86 if (already_sent[i] == 0)
87 smpi_mpi_iprobe(i, MPI_ANY_TAG, comm, &flag_array[i],
93 /* recv 1-byte message */
94 for (i = 1; i < size; i++) {
97 if ((flag_array[i] == 1) && (already_sent[i] == 0)) {
98 smpi_mpi_recv(temp_buf, 1, MPI_CHAR, i, tag, comm, &status);
99 header_buf[header_index] = i;
103 /* will send in the next step */
108 /* send header followed by data */
109 if (header_index != 0) {
110 header_buf[header_index] = -1;
112 smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
113 smpi_mpi_send(buf, count, datatype, to, tag, comm);
116 /* randomly MPI_Send to one */
118 /* search for the first node that never received data before */
119 for (i = 1; i < size; i++) {
120 if (already_sent[i] == 0) {
123 smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, i, tag, comm);
124 smpi_mpi_send(buf, count, datatype, i, tag, comm);
136 for(i=0; i<size; i++)
137 if(to_clean[i]!=0)smpi_mpi_recv(temp_buf, 1, MPI_CHAR, i, tag, comm, &status);
143 /* send 1-byte message to root */
144 smpi_mpi_send(temp_buf, 1, MPI_CHAR, 0, tag, comm);
146 /* wait for header and data, forward when required */
147 smpi_mpi_recv(header_buf, HEADER_SIZE, MPI_INT, MPI_ANY_SOURCE, tag, comm,
149 smpi_mpi_recv(buf, count, datatype, MPI_ANY_SOURCE, tag, comm, &status);
151 /* search for where it is */
153 while (rank != header_buf[myordering]) {
157 /* send header followed by data */
158 if (header_buf[myordering + 1] != -1) {
159 smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, header_buf[myordering + 1],
161 smpi_mpi_send(buf, count, datatype, header_buf[myordering + 1], tag, comm);
168 (MPI_Request *) xbt_malloc((size + pipe_length) * sizeof(MPI_Request));
170 (MPI_Request *) xbt_malloc((size + pipe_length) * sizeof(MPI_Request));
172 (MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
174 (MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
181 for (i = 0; i < 1000; i++)
183 while (sent_count < (size - 1)) {
185 //start = MPI_Wtime();
188 for (k = 0; k < 3; k++) {
189 for (i = 1; i < size; i++) {
190 if ((already_sent[i] == 0) && (will_send[i] == 0)) {
191 smpi_mpi_iprobe(i, MPI_ANY_TAG, comm, &flag_array[i],
192 &temp_status_array[i]);
193 if (flag_array[i] == 1) {
195 smpi_mpi_recv(&temp_buf[i], 1, MPI_CHAR, i, tag, comm,
203 //total = MPI_Wtime() - start;
205 //printf("Iprobe time = %.2f\n",total);
208 //start = MPI_Wtime();
209 /* recv 1-byte message */
210 for (i = 1; i < size; i++) {
212 if ((will_send[i] == 1) && (already_sent[i] == 0)) {
213 header_buf[header_index] = i;
217 /* will send in the next step */
221 //printf("sent_count = %d\n",sent_count);
224 //total = MPI_Wtime() - start;
226 //printf("Recv 1-byte time = %.2f\n",total);
229 if (header_index != 0) {
230 printf("header index = %d node = ",header_index);
231 for (i=0;i<header_index;i++) {
232 printf("%d ",header_buf[i]);
238 /* send header followed by data */
239 if (header_index != 0) {
240 header_buf[header_index] = -1;
243 //start = MPI_Wtime();
246 smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
248 //total = MPI_Wtime() - start;
250 //printf("\tSend header to %d time = %.2f\n",to,total);
252 //start = MPI_Wtime();
254 /* send data - non-pipeline case */
257 //if (header_index == 1) {
258 smpi_mpi_send(buf, count, datatype, to, tag, comm);
262 /* send data - pipeline */
264 for (i = 0; i < pipe_length; i++) {
265 smpi_mpi_send((char *)buf + (i * increment), segment, datatype, to, tag, comm);
267 //smpi_mpi_waitall((pipe_length), send_request_array, send_status_array);
269 //total = MPI_Wtime() - start;
271 //printf("\tSend data to %d time = %.2f\n",to,total);
277 /* randomly MPI_Send to one node */
279 /* search for the first node that never received data before */
280 for (i = 1; i < size; i++) {
281 if (already_sent[i] == 0) {
286 //start = MPI_Wtime();
287 smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, to, tag, comm);
289 /* still need to chop data so that we can use the same non-root code */
290 for (j = 0; j < pipe_length; j++) {
291 smpi_mpi_send((char *)buf + (j * increment), segment, datatype, to, tag,
295 //smpi_mpi_send(buf,count,datatype,to,tag,comm);
296 //smpi_mpi_wait(&request,MPI_STATUS_IGNORE);
298 //total = MPI_Wtime() - start;
300 //printf("SEND TO SINGLE node %d time = %.2f\n",i,total);
312 //total = MPI_Wtime() - start2;
314 //printf("Node zero iter = %d time = %.2f\n",iteration,total);
316 /* probe before exit in case there are messages to recv */
317 for (i = 1; i < size; i++) {
318 smpi_mpi_iprobe(i, MPI_ANY_TAG, comm, &flag_array[i],
319 &temp_status_array[i]);
320 if (flag_array[i] == 1)
321 smpi_mpi_recv(&temp_buf[i], 1, MPI_CHAR, i, tag, comm, &status);
329 /* if root already send a message to this node, don't send one-byte message */
330 smpi_mpi_iprobe(0, MPI_ANY_TAG, comm, &flag_array[0], &status);
332 /* send 1-byte message to root */
333 if (flag_array[0] == 0)
334 smpi_mpi_send(temp_buf, 1, MPI_CHAR, 0, tag, comm);
336 /* wait for header forward when required */
337 request = smpi_mpi_irecv(header_buf, HEADER_SIZE, MPI_INT, MPI_ANY_SOURCE, tag, comm);
338 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
340 /* search for where it is */
342 while (rank != header_buf[myordering]) {
346 /* send header when required */
347 if (header_buf[myordering + 1] != -1) {
348 smpi_mpi_send(header_buf, HEADER_SIZE, MPI_INT, header_buf[myordering + 1],
355 //if (header_buf[1] == -1) {
356 request = smpi_mpi_irecv(buf, count, datatype, 0, tag, comm);
357 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
358 //printf("\t\tnode %d ordering = %d receive data from root\n",rank,myordering);
360 for (i = 0; i < pipe_length; i++) {
361 recv_request_array[i] = smpi_mpi_irecv((char *)buf + (i * increment), segment, datatype, MPI_ANY_SOURCE,
367 if (header_buf[myordering + 1] != -1) {
368 for (i = 0; i < pipe_length; i++) {
369 smpi_mpi_wait(&recv_request_array[i], MPI_STATUS_IGNORE);
370 send_request_array[i] = smpi_mpi_isend((char *)buf + (i * increment), segment, datatype,
371 header_buf[myordering + 1], tag, comm);
373 smpi_mpi_waitall((pipe_length), send_request_array, send_status_array);
375 for (i = 0; i < pipe_length; i++)
376 smpi_mpi_wait(&recv_request_array[i], &recv_status_array[i]);
381 free(send_request_array);
382 free(recv_request_array);
383 free(send_status_array);
384 free(recv_status_array);
387 /* when count is not divisible by block size, use default BCAST for the remainder */
388 if ((remainder != 0) && (count > segment)) {
389 XBT_WARN("MPI_bcast_arrival_nb use default MPI_bcast.");
390 smpi_mpi_bcast((char *)buf + (pipe_length * increment), remainder, datatype, root, comm);