1 /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
4 * (C) 2003 by Argonne National Laboratory.
5 * See COPYRIGHT in top-level directory.
8 /* MPI-3 distributed linked list construction example
9 * --------------------------------------------------
11 * Construct a distributed shared linked list using proposed MPI-3 dynamic
12 * windows. Initially process 0 creates the head of the list, attaches it to
13 * the window, and broadcasts the pointer to all processes. All processes then
14 * concurrently append N new elements to the list. When a process attempts to
15 * attach its element to the tail of list it may discover that its tail pointer
16 * is stale and it must chase ahead to the new tail before the element can be
32 #define ELEM_PER_ROW 16
34 /* Linked list pointer */
40 /* Linked list element */
46 static const llist_ptr_t nil = { -1, (MPI_Aint) MPI_BOTTOM };
48 static const int verbose = 0;
50 /* List of locally allocated list elements. */
51 static llist_elem_t **my_elems = NULL;
52 static int my_elems_size = 0;
53 static int my_elems_count = 0;
55 /* Allocate a new shared linked list element */
56 MPI_Aint alloc_elem(int value, MPI_Win win)
59 llist_elem_t *elem_ptr;
61 /* Allocate the new element and register it with the window */
62 MPI_Alloc_mem(sizeof(llist_elem_t), MPI_INFO_NULL, &elem_ptr);
63 elem_ptr->value = value;
65 MPI_Win_attach(win, elem_ptr, sizeof(llist_elem_t));
67 /* Add the element to the list of local elements so we can free it later. */
68 if (my_elems_size == my_elems_count) {
70 my_elems = realloc(my_elems, my_elems_size * sizeof(void *));
72 my_elems[my_elems_count] = elem_ptr;
75 MPI_Get_address(elem_ptr, &disp);
79 int main(int argc, char **argv)
83 llist_ptr_t head_ptr, tail_ptr;
85 MPI_Init(&argc, &argv);
87 MPI_Comm_rank(MPI_COMM_WORLD, &procid);
88 MPI_Comm_size(MPI_COMM_WORLD, &nproc);
90 MPI_Win_create_dynamic(MPI_INFO_NULL, MPI_COMM_WORLD, &llist_win);
92 /* Process 0 creates the head node */
94 head_ptr.disp = alloc_elem(-1, llist_win);
96 /* Broadcast the head pointer to everyone */
98 MPI_Bcast(&head_ptr.disp, 1, MPI_AINT, 0, MPI_COMM_WORLD);
101 /* Lock the window for shared access to all targets */
102 MPI_Win_lock_all(0, llist_win);
104 /* All processes concurrently append NUM_ELEMS elements to the list */
105 for (i = 0; i < NUM_ELEMS; i++) {
106 llist_ptr_t new_elem_ptr;
109 /* Create a new list element and register it with the window */
110 new_elem_ptr.rank = procid;
111 new_elem_ptr.disp = alloc_elem(procid, llist_win);
113 /* Append the new node to the list. This might take multiple attempts if
114 * others have already appended and our tail pointer is stale. */
116 llist_ptr_t next_tail_ptr = nil;
118 MPI_Compare_and_swap((void *) &new_elem_ptr.rank, (void *) &nil.rank,
119 (void *) &next_tail_ptr.rank, MPI_INT, tail_ptr.rank,
120 (MPI_Aint) & (((llist_elem_t *) tail_ptr.disp)->next.rank),
123 MPI_Win_flush(tail_ptr.rank, llist_win);
124 success = (next_tail_ptr.rank == nil.rank);
129 MPI_Accumulate(&new_elem_ptr.disp, 1, MPI_AINT, tail_ptr.rank,
130 (MPI_Aint) & (((llist_elem_t *) tail_ptr.disp)->next.disp), 1,
131 MPI_AINT, MPI_REPLACE, llist_win);
133 MPI_Win_flush(tail_ptr.rank, llist_win);
134 tail_ptr = new_elem_ptr;
136 /* For implementations that use pt-to-pt messaging, force progress for other threads'
138 for (i = 0; i < NPROBE; i++)
139 MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag,
144 /* Tail pointer is stale, fetch the displacement. May take multiple tries
145 * if it is being updated. */
147 MPI_Get_accumulate(NULL, 0, MPI_DATATYPE_NULL, &next_tail_ptr.disp,
148 1, MPI_AINT, tail_ptr.rank,
149 (MPI_Aint) & (((llist_elem_t *) tail_ptr.disp)->next.disp),
150 1, MPI_AINT, MPI_NO_OP, llist_win);
151 MPI_Win_flush(tail_ptr.rank, llist_win);
152 } while (next_tail_ptr.disp == nil.disp);
153 tail_ptr = next_tail_ptr;
158 MPI_Win_unlock_all(llist_win);
159 MPI_Barrier(MPI_COMM_WORLD);
161 /* Traverse the list and verify that all processes inserted exactly the correct
162 * number of elements. */
166 int *counts, count = 0;
168 counts = (int *) malloc(sizeof(int) * nproc);
169 assert(counts != NULL);
171 for (i = 0; i < nproc; i++)
176 MPI_Win_lock_all(0, llist_win);
178 /* Walk the list and tally up the number of elements inserted by each rank */
179 while (tail_ptr.disp != nil.disp) {
182 MPI_Get(&elem, sizeof(llist_elem_t), MPI_BYTE,
183 tail_ptr.rank, tail_ptr.disp, sizeof(llist_elem_t), MPI_BYTE, llist_win);
185 MPI_Win_flush(tail_ptr.rank, llist_win);
187 tail_ptr = elem.next;
189 /* This is not the root */
191 assert(elem.value >= 0 && elem.value < nproc);
192 counts[elem.value]++;
196 int last_elem = tail_ptr.disp == nil.disp;
197 printf("%2d%s", elem.value, last_elem ? "" : " -> ");
198 if (count % ELEM_PER_ROW == 0 && !last_elem)
203 /* This is the root */
205 assert(elem.value == -1);
210 MPI_Win_unlock_all(llist_win);
215 /* Verify the counts we collected */
216 for (i = 0; i < nproc; i++) {
217 int expected = NUM_ELEMS;
219 if (counts[i] != expected) {
220 printf("Error: Rank %d inserted %d elements, expected %d\n", i, counts[i],
226 printf("%s\n", errors == 0 ? " No Errors" : "FAIL");
230 MPI_Win_free(&llist_win);
232 /* Free all the elements in the list */
233 for (; my_elems_count > 0; my_elems_count--)
234 MPI_Free_mem(my_elems[my_elems_count - 1]);