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 /* All processes concurrently append NUM_ELEMS elements to the list */
102 for (i = 0; i < NUM_ELEMS; i++) {
103 llist_ptr_t new_elem_ptr;
106 /* Create a new list element and register it with the window */
107 new_elem_ptr.rank = procid;
108 new_elem_ptr.disp = alloc_elem(procid, llist_win);
110 /* Append the new node to the list. This might take multiple attempts if
111 * others have already appended and our tail pointer is stale. */
113 llist_ptr_t next_tail_ptr = nil;
115 MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
117 MPI_Compare_and_swap((void *) &new_elem_ptr.rank, (void *) &nil.rank,
118 (void *) &next_tail_ptr.rank, MPI_INT, tail_ptr.rank,
119 (MPI_Aint) & (((llist_elem_t *) tail_ptr.disp)->next.rank),
122 MPI_Win_unlock(tail_ptr.rank, llist_win);
123 success = (next_tail_ptr.rank == nil.rank);
129 MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
131 MPI_Fetch_and_op(&new_elem_ptr.disp, &result, MPI_AINT, tail_ptr.rank,
132 (MPI_Aint) & (((llist_elem_t *) tail_ptr.disp)->next.disp),
133 MPI_REPLACE, llist_win);
135 /* Note: accumulate is faster, since we don't need the result. Replacing with
136 * Fetch_and_op to create a more complete test case. */
138 * MPI_Accumulate(&new_elem_ptr.disp, 1, MPI_AINT, tail_ptr.rank,
139 * (MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp), 1,
140 * MPI_AINT, MPI_REPLACE, llist_win);
143 MPI_Win_unlock(tail_ptr.rank, llist_win);
144 tail_ptr = new_elem_ptr;
146 /* For implementations that use pt-to-pt messaging, force progress for other threads'
148 for (i = 0; i < NPROBE; i++)
149 MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag,
154 /* Tail pointer is stale, fetch the displacement. May take multiple tries
155 * if it is being updated. */
157 MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
159 MPI_Fetch_and_op(NULL, &next_tail_ptr.disp, MPI_AINT, tail_ptr.rank,
160 (MPI_Aint) & (((llist_elem_t *) tail_ptr.disp)->next.disp),
161 MPI_NO_OP, llist_win);
163 MPI_Win_unlock(tail_ptr.rank, llist_win);
164 } while (next_tail_ptr.disp == nil.disp);
165 tail_ptr = next_tail_ptr;
170 MPI_Barrier(MPI_COMM_WORLD);
172 /* Traverse the list and verify that all processes inserted exactly the correct
173 * number of elements. */
177 int *counts, count = 0;
179 counts = (int *) malloc(sizeof(int) * nproc);
180 assert(counts != NULL);
182 for (i = 0; i < nproc; i++)
187 /* Walk the list and tally up the number of elements inserted by each rank */
188 while (tail_ptr.disp != nil.disp) {
191 MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
193 MPI_Get(&elem, sizeof(llist_elem_t), MPI_BYTE,
194 tail_ptr.rank, tail_ptr.disp, sizeof(llist_elem_t), MPI_BYTE, llist_win);
196 MPI_Win_unlock(tail_ptr.rank, llist_win);
198 tail_ptr = elem.next;
200 /* This is not the root */
202 assert(elem.value >= 0 && elem.value < nproc);
203 counts[elem.value]++;
207 int last_elem = tail_ptr.disp == nil.disp;
208 printf("%2d%s", elem.value, last_elem ? "" : " -> ");
209 if (count % ELEM_PER_ROW == 0 && !last_elem)
214 /* This is the root */
216 assert(elem.value == -1);
224 /* Verify the counts we collected */
225 for (i = 0; i < nproc; i++) {
226 int expected = NUM_ELEMS;
228 if (counts[i] != expected) {
229 printf("Error: Rank %d inserted %d elements, expected %d\n", i, counts[i],
235 printf("%s\n", errors == 0 ? " No Errors" : "FAIL");
239 MPI_Win_free(&llist_win);
241 /* Free all the elements in the list */
242 for (; my_elems_count > 0; my_elems_count--)
243 MPI_Free_mem(my_elems[my_elems_count - 1]);