/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
/*
* (C) 2011 by Argonne National Laboratory.
* See COPYRIGHT in top-level directory.
*/
/* This test attempts to execute multiple simultaneous nonblocking collective
* (NBC) MPI routines at the same time, and manages their completion with a
* variety of routines (MPI_{Wait,Test}{,_all,_any,_some}). It also throws a
* few point-to-point operations into the mix.
*
* Possible improvements:
* - post operations on multiple comms from multiple threads
*/
#include "mpi.h"
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <assert.h>
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif
static int errs = 0;
/* Constants that control the high level test harness behavior. */
/* MAIN_ITERATIONS is how many NBC ops the test will attempt to issue. */
#define MAIN_ITERATIONS (100000)
/* WINDOW is the maximum number of outstanding NBC requests at any given time */
#define WINDOW (20)
/* we sleep with probability 1/CHANCE_OF_SLEEP */
#define CHANCE_OF_SLEEP (1000)
/* JITTER_DELAY is denominated in microseconds (us) */
#define JITTER_DELAY (50000) /* 0.05 seconds */
/* NUM_COMMS is the number of communicators on which ops will be posted */
#define NUM_COMMS (4)
/* Constants that control behavior of the individual testing operations.
* Altering these can help to explore the testing space, but increasing them too
* much can consume too much memory (often O(n^2) usage). */
/* FIXME is COUNT==10 too limiting? should we try a larger count too (~500)? */
#define COUNT (10)
#define PRIME (17)
#define my_assert(cond_) \
do { \
if (!(cond_)) { \
++errs; \
if (errs < 10) { \
fprintf(stderr, "assertion (%s) failed on line %d\n", #cond_, __LINE__); \
} \
} \
} while (0)
/* Intended to act like "rand_r", but we can be sure that it will exist and be
* consistent across all of comm world. Returns a number in the range
* [0,GEN_PRN_MAX] */
#define GEN_PRN_MAX (4294967291-1)
static unsigned int gen_prn(unsigned int x)
{
/* a simple "multiplicative congruential method" PRNG, with parameters:
* m=4294967291, largest 32-bit prime
* a=279470273, good primitive root of m from "TABLES OF LINEAR
* CONGRUENTIAL GENERATORS OF DIFFERENT SIZES AND GOOD
* LATTICE STRUCTURE", by Pierre L’Ecuyer */
return (279470273UL * (unsigned long) x) % 4294967291UL;
}
/* given a random unsigned int value "rndval_" from gen_prn, this evaluates to a
* value in the range [min_,max_) */
#define rand_range(rndval_,min_,max_) \
((unsigned int)((min_) + ((rndval_) * (1.0 / (GEN_PRN_MAX+1.0)) * ((max_) - (min_)))))
static void sum_fn(void *invec, void *inoutvec, int *len, MPI_Datatype * datatype)
{
int i;
int *in = invec;
int *inout = inoutvec;
for (i = 0; i < *len; ++i) {
inout[i] = in[i] + inout[i];
}
}
/* used to keep track of buffers that should be freed after the corresponding
* operation has completed */
struct laundry {
int case_num; /* which test case initiated this req/laundry */
MPI_Comm comm;
int *buf;
int *recvbuf;
int *sendcounts;
int *recvcounts;
int *sdispls;
int *rdispls;
int *sendtypes;
int *recvtypes;
};
static void cleanup_laundry(struct laundry *l)
{
l->case_num = -1;
l->comm = MPI_COMM_NULL;
if (l->buf)
free(l->buf);
if (l->recvbuf)
free(l->recvbuf);
if (l->sendcounts)
free(l->sendcounts);
if (l->recvcounts)
free(l->recvcounts);
if (l->sdispls)
free(l->sdispls);
if (l->rdispls)
free(l->rdispls);
if (l->sendtypes)
free(l->sendtypes);
if (l->recvtypes)
free(l->recvtypes);
}
/* Starts a "random" operation on "comm" corresponding to "rndnum" and returns
* in (*req) a request handle corresonding to that operation. This call should
* be considered collective over comm (with a consistent value for "rndnum"),
* even though the operation may only be a point-to-point request. */
static void start_random_nonblocking(MPI_Comm comm, unsigned int rndnum, MPI_Request * req,
struct laundry *l)
{
int i, j;
int rank, size;
int *buf = NULL;
int *recvbuf = NULL;
int *sendcounts = NULL;
int *recvcounts = NULL;
int *sdispls = NULL;
int *rdispls = NULL;
int *sendtypes = NULL;
int *recvtypes = NULL;
signed char *buf_alias = NULL;
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &size);
*req = MPI_REQUEST_NULL;
l->case_num = -1;
l->comm = comm;
l->buf = buf = malloc(COUNT * size * sizeof(int));
l->recvbuf = recvbuf = malloc(COUNT * size * sizeof(int));
l->sendcounts = sendcounts = malloc(size * sizeof(int));
l->recvcounts = recvcounts = malloc(size * sizeof(int));
l->sdispls = sdispls = malloc(size * sizeof(int));
l->rdispls = rdispls = malloc(size * sizeof(int));
l->sendtypes = sendtypes = malloc(size * sizeof(MPI_Datatype));
l->recvtypes = recvtypes = malloc(size * sizeof(MPI_Datatype));
#define NUM_CASES (21)
l->case_num = rand_range(rndnum, 0, NUM_CASES);
switch (l->case_num) {
case 0: /* MPI_Ibcast */
for (i = 0; i < COUNT; ++i) {
if (rank == 0) {
buf[i] = i;
}
else {
buf[i] = 0xdeadbeef;
}
}
MPI_Ibcast(buf, COUNT, MPI_INT, 0, comm, req);
break;
case 1: /* MPI_Ibcast (again, but designed to stress scatter/allgather impls) */
/* FIXME fiddle with PRIME and buffer allocation s.t. PRIME is much larger (1021?) */
buf_alias = (signed char *) buf;
my_assert(COUNT * size * sizeof(int) > PRIME); /* sanity */
for (i = 0; i < PRIME; ++i) {
if (rank == 0)
buf_alias[i] = i;
else
buf_alias[i] = 0xdb;
}
for (i = PRIME; i < COUNT * size * sizeof(int); ++i) {
buf_alias[i] = 0xbf;
}
MPI_Ibcast(buf_alias, PRIME, MPI_SIGNED_CHAR, 0, comm, req);
break;
case 2: /* MPI_Ibarrier */
MPI_Ibarrier(comm, req);
break;
case 3: /* MPI_Ireduce */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Ireduce(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, 0, comm, req);
break;
case 4: /* same again, use a user op and free it before the wait */
{
MPI_Op op = MPI_OP_NULL;
MPI_Op_create(sum_fn, /*commute= */ 1, &op);
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Ireduce(buf, recvbuf, COUNT, MPI_INT, op, 0, comm, req);
MPI_Op_free(&op);
}
break;
case 5: /* MPI_Iallreduce */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iallreduce(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 6: /* MPI_Ialltoallv (a weak test, neither irregular nor sparse) */
for (i = 0; i < size; ++i) {
sendcounts[i] = COUNT;
recvcounts[i] = COUNT;
sdispls[i] = COUNT * i;
rdispls[i] = COUNT * i;
for (j = 0; j < COUNT; ++j) {
buf[i * COUNT + j] = rank + (i * j);
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Ialltoallv(buf, sendcounts, sdispls, MPI_INT, recvbuf, recvcounts, rdispls, MPI_INT,
comm, req);
break;
case 7: /* MPI_Igather */
for (i = 0; i < size * COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Igather(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, 0, comm, req);
break;
case 8: /* same test again, just use a dup'ed datatype and free it before the wait */
{
MPI_Datatype type = MPI_DATATYPE_NULL;
MPI_Type_dup(MPI_INT, &type);
for (i = 0; i < size * COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Igather(buf, COUNT, MPI_INT, recvbuf, COUNT, type, 0, comm, req);
MPI_Type_free(&type); /* should cause implementations that don't refcount
* correctly to blow up or hang in the wait */
}
break;
case 9: /* MPI_Iscatter */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
if (rank == 0)
buf[i * COUNT + j] = i + j;
else
buf[i * COUNT + j] = 0xdeadbeef;
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Iscatter(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, 0, comm, req);
break;
case 10: /* MPI_Iscatterv */
for (i = 0; i < size; ++i) {
/* weak test, just test the regular case where all counts are equal */
sendcounts[i] = COUNT;
sdispls[i] = i * COUNT;
for (j = 0; j < COUNT; ++j) {
if (rank == 0)
buf[i * COUNT + j] = i + j;
else
buf[i * COUNT + j] = 0xdeadbeef;
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Iscatterv(buf, sendcounts, sdispls, MPI_INT, recvbuf, COUNT, MPI_INT, 0, comm, req);
break;
case 11: /* MPI_Ireduce_scatter */
for (i = 0; i < size; ++i) {
recvcounts[i] = COUNT;
for (j = 0; j < COUNT; ++j) {
buf[i * COUNT + j] = rank + i;
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Ireduce_scatter(buf, recvbuf, recvcounts, MPI_INT, MPI_SUM, comm, req);
break;
case 12: /* MPI_Ireduce_scatter_block */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
buf[i * COUNT + j] = rank + i;
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Ireduce_scatter_block(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 13: /* MPI_Igatherv */
for (i = 0; i < size * COUNT; ++i) {
buf[i] = 0xdeadbeef;
recvbuf[i] = 0xdeadbeef;
}
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
}
for (i = 0; i < size; ++i) {
recvcounts[i] = COUNT;
rdispls[i] = i * COUNT;
}
MPI_Igatherv(buf, COUNT, MPI_INT, recvbuf, recvcounts, rdispls, MPI_INT, 0, comm, req);
break;
case 14: /* MPI_Ialltoall */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
buf[i * COUNT + j] = rank + (i * j);
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Ialltoall(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, comm, req);
break;
case 15: /* MPI_Iallgather */
for (i = 0; i < size * COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iallgather(buf, COUNT, MPI_INT, recvbuf, COUNT, MPI_INT, comm, req);
break;
case 16: /* MPI_Iallgatherv */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
recvcounts[i] = COUNT;
rdispls[i] = i * COUNT;
}
for (i = 0; i < COUNT; ++i)
buf[i] = rank + i;
MPI_Iallgatherv(buf, COUNT, MPI_INT, recvbuf, recvcounts, rdispls, MPI_INT, comm, req);
break;
case 17: /* MPI_Iscan */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iscan(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 18: /* MPI_Iexscan */
for (i = 0; i < COUNT; ++i) {
buf[i] = rank + i;
recvbuf[i] = 0xdeadbeef;
}
MPI_Iexscan(buf, recvbuf, COUNT, MPI_INT, MPI_SUM, comm, req);
break;
case 19: /* MPI_Ialltoallw (a weak test, neither irregular nor sparse) */
for (i = 0; i < size; ++i) {
sendcounts[i] = COUNT;
recvcounts[i] = COUNT;
sdispls[i] = COUNT * i * sizeof(int);
rdispls[i] = COUNT * i * sizeof(int);
sendtypes[i] = MPI_INT;
recvtypes[i] = MPI_INT;
for (j = 0; j < COUNT; ++j) {
buf[i * COUNT + j] = rank + (i * j);
recvbuf[i * COUNT + j] = 0xdeadbeef;
}
}
MPI_Ialltoallw(buf, sendcounts, sdispls, sendtypes, recvbuf, recvcounts, rdispls, recvtypes,
comm, req);
break;
case 20: /* basic pt2pt MPI_Isend/MPI_Irecv pairing */
/* even ranks send to odd ranks, but only if we have a full pair */
if ((rank % 2 != 0) || (rank != size - 1)) {
for (j = 0; j < COUNT; ++j) {
buf[j] = j;
recvbuf[j] = 0xdeadbeef;
}
if (rank % 2 == 0)
MPI_Isend(buf, COUNT, MPI_INT, rank + 1, 5, comm, req);
else
MPI_Irecv(recvbuf, COUNT, MPI_INT, rank - 1, 5, comm, req);
}
break;
default:
fprintf(stderr, "unexpected value for l->case_num=%d)\n", (l->case_num));
MPI_Abort(comm, 1);
break;
}
}
static void check_after_completion(struct laundry *l)
{
int i, j;
int rank, size;
MPI_Comm comm = l->comm;
int *buf = l->buf;
int *recvbuf = l->recvbuf;
int *sendcounts = l->sendcounts;
int *recvcounts = l->recvcounts;
int *sdispls = l->sdispls;
int *rdispls = l->rdispls;
int *sendtypes = l->sendtypes;
int *recvtypes = l->recvtypes;
char *buf_alias = (char *) buf;
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &size);
/* these cases all correspond to cases in start_random_nonblocking */
switch (l->case_num) {
case 0: /* MPI_Ibcast */
for (i = 0; i < COUNT; ++i) {
if (buf[i] != i)
printf("buf[%d]=%d i=%d\n", i, buf[i], i);
my_assert(buf[i] == i);
}
break;
case 1: /* MPI_Ibcast (again, but designed to stress scatter/allgather impls) */
for (i = 0; i < PRIME; ++i) {
if (buf_alias[i] != i)
printf("buf_alias[%d]=%d i=%d\n", i, buf_alias[i], i);
my_assert(buf_alias[i] == i);
}
break;
case 2: /* MPI_Ibarrier */
/* nothing to check */
break;
case 3: /* MPI_Ireduce */
if (rank == 0) {
for (i = 0; i < COUNT; ++i) {
if (recvbuf[i] != ((size * (size - 1) / 2) + (i * size)))
printf("got recvbuf[%d]=%d, expected %d\n", i, recvbuf[i],
((size * (size - 1) / 2) + (i * size)));
my_assert(recvbuf[i] == ((size * (size - 1) / 2) + (i * size)));
}
}
break;
case 4: /* same again, use a user op and free it before the wait */
if (rank == 0) {
for (i = 0; i < COUNT; ++i) {
if (recvbuf[i] != ((size * (size - 1) / 2) + (i * size)))
printf("got recvbuf[%d]=%d, expected %d\n", i, recvbuf[i],
((size * (size - 1) / 2) + (i * size)));
my_assert(recvbuf[i] == ((size * (size - 1) / 2) + (i * size)));
}
}
break;
case 5: /* MPI_Iallreduce */
for (i = 0; i < COUNT; ++i) {
if (recvbuf[i] != ((size * (size - 1) / 2) + (i * size)))
printf("got recvbuf[%d]=%d, expected %d\n", i, recvbuf[i],
((size * (size - 1) / 2) + (i * size)));
my_assert(recvbuf[i] == ((size * (size - 1) / 2) + (i * size)));
}
break;
case 6: /* MPI_Ialltoallv (a weak test, neither irregular nor sparse) */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
/*printf("recvbuf[%d*COUNT+%d]=%d, expecting %d\n", i, j, recvbuf[i*COUNT+j], (i + (rank * j))); */
my_assert(recvbuf[i * COUNT + j] == (i + (rank * j)));
}
}
break;
case 7: /* MPI_Igather */
if (rank == 0) {
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[i * COUNT + j] == i + j);
}
}
}
else {
for (i = 0; i < size * COUNT; ++i) {
my_assert(recvbuf[i] == 0xdeadbeef);
}
}
break;
case 8: /* same test again, just use a dup'ed datatype and free it before the wait */
if (rank == 0) {
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[i * COUNT + j] == i + j);
}
}
}
else {
for (i = 0; i < size * COUNT; ++i) {
my_assert(recvbuf[i] == 0xdeadbeef);
}
}
break;
case 9: /* MPI_Iscatter */
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[j] == rank + j);
}
if (rank != 0) {
for (i = 0; i < size * COUNT; ++i) {
/* check we didn't corrupt the sendbuf somehow */
my_assert(buf[i] == 0xdeadbeef);
}
}
break;
case 10: /* MPI_Iscatterv */
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[j] == rank + j);
}
if (rank != 0) {
for (i = 0; i < size * COUNT; ++i) {
/* check we didn't corrupt the sendbuf somehow */
my_assert(buf[i] == 0xdeadbeef);
}
}
for (i = 1; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
/* check we didn't corrupt the rest of the recvbuf */
my_assert(recvbuf[i * COUNT + j] == 0xdeadbeef);
}
}
break;
case 11: /* MPI_Ireduce_scatter */
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[j] == (size * rank + ((size - 1) * size) / 2));
}
for (i = 1; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
/* check we didn't corrupt the rest of the recvbuf */
my_assert(recvbuf[i * COUNT + j] == 0xdeadbeef);
}
}
break;
case 12: /* MPI_Ireduce_scatter_block */
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[j] == (size * rank + ((size - 1) * size) / 2));
}
for (i = 1; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
/* check we didn't corrupt the rest of the recvbuf */
my_assert(recvbuf[i * COUNT + j] == 0xdeadbeef);
}
}
break;
case 13: /* MPI_Igatherv */
if (rank == 0) {
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[i * COUNT + j] == i + j);
}
}
}
else {
for (i = 0; i < size * COUNT; ++i) {
my_assert(recvbuf[i] == 0xdeadbeef);
}
}
break;
case 14: /* MPI_Ialltoall */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
/*printf("recvbuf[%d*COUNT+%d]=%d, expecting %d\n", i, j, recvbuf[i*COUNT+j], (i + (i * j))); */
my_assert(recvbuf[i * COUNT + j] == (i + (rank * j)));
}
}
break;
case 15: /* MPI_Iallgather */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[i * COUNT + j] == i + j);
}
}
break;
case 16: /* MPI_Iallgatherv */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
my_assert(recvbuf[i * COUNT + j] == i + j);
}
}
break;
case 17: /* MPI_Iscan */
for (i = 0; i < COUNT; ++i) {
my_assert(recvbuf[i] == ((rank * (rank + 1) / 2) + (i * (rank + 1))));
}
break;
case 18: /* MPI_Iexscan */
for (i = 0; i < COUNT; ++i) {
if (rank == 0)
my_assert(recvbuf[i] == 0xdeadbeef);
else
my_assert(recvbuf[i] == ((rank * (rank + 1) / 2) + (i * (rank + 1)) - (rank + i)));
}
break;
case 19: /* MPI_Ialltoallw (a weak test, neither irregular nor sparse) */
for (i = 0; i < size; ++i) {
for (j = 0; j < COUNT; ++j) {
/*printf("recvbuf[%d*COUNT+%d]=%d, expecting %d\n", i, j, recvbuf[i*COUNT+j], (i + (rank * j))); */
my_assert(recvbuf[i * COUNT + j] == (i + (rank * j)));
}
}
break;
case 20: /* basic pt2pt MPI_Isend/MPI_Irecv pairing */
/* even ranks send to odd ranks, but only if we have a full pair */
if ((rank % 2 != 0) || (rank != size - 1)) {
for (j = 0; j < COUNT; ++j) {
/* only odd procs did a recv */
if (rank % 2 == 0) {
my_assert(recvbuf[j] == 0xdeadbeef);
}
else {
if (recvbuf[j] != j)
printf("recvbuf[%d]=%d j=%d\n", j, recvbuf[j], j);
my_assert(recvbuf[j] == j);
}
}
}
break;
default:
printf("invalid case_num (%d) detected\n", l->case_num);
assert(0);
break;
}
}
#undef NUM_CASES
static void complete_something_somehow(unsigned int rndnum, int numreqs, MPI_Request reqs[],
int *outcount, int indices[])
{
int i, idx, flag;
#define COMPLETION_CASES (8)
switch (rand_range(rndnum, 0, COMPLETION_CASES)) {
case 0:
MPI_Waitall(numreqs, reqs, MPI_STATUSES_IGNORE);
*outcount = numreqs;
for (i = 0; i < numreqs; ++i) {
indices[i] = i;
}
break;
case 1:
MPI_Testsome(numreqs, reqs, outcount, indices, MPI_STATUS_IGNORE);
if (*outcount == MPI_UNDEFINED) {
*outcount = 0;
}
break;
case 2:
MPI_Waitsome(numreqs, reqs, outcount, indices, MPI_STATUS_IGNORE);
if (*outcount == MPI_UNDEFINED) {
*outcount = 0;
}
break;
case 3:
MPI_Waitany(numreqs, reqs, &idx, MPI_STATUS_IGNORE);
if (idx == MPI_UNDEFINED) {
*outcount = 0;
}
else {
*outcount = 1;
indices[0] = idx;
}
break;
case 4:
MPI_Testany(numreqs, reqs, &idx, &flag, MPI_STATUS_IGNORE);
if (idx == MPI_UNDEFINED) {
*outcount = 0;
}
else {
*outcount = 1;
indices[0] = idx;
}
break;
case 5:
MPI_Testall(numreqs, reqs, &flag, MPI_STATUSES_IGNORE);
if (flag) {
*outcount = numreqs;
for (i = 0; i < numreqs; ++i) {
indices[i] = i;
}
}
else {
*outcount = 0;
}
break;
case 6:
/* select a new random index and wait on it */
rndnum = gen_prn(rndnum);
idx = rand_range(rndnum, 0, numreqs);
MPI_Wait(&reqs[idx], MPI_STATUS_IGNORE);
*outcount = 1;
indices[0] = idx;
break;
case 7:
/* select a new random index and wait on it */
rndnum = gen_prn(rndnum);
idx = rand_range(rndnum, 0, numreqs);
MPI_Test(&reqs[idx], &flag, MPI_STATUS_IGNORE);
*outcount = (flag ? 1 : 0);
indices[0] = idx;
break;
default:
assert(0);
break;
}
#undef COMPLETION_CASES
}
int main(int argc, char **argv)
{
int i, num_posted, num_completed;
int wrank, wsize;
unsigned int seed = 0x10bc;
unsigned int post_seq, complete_seq;
struct laundry larr[WINDOW];
MPI_Request reqs[WINDOW];
int outcount;
int indices[WINDOW];
MPI_Comm comms[NUM_COMMS];
MPI_Comm comm;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &wrank);
MPI_Comm_size(MPI_COMM_WORLD, &wsize);
/* it is critical that all processes in the communicator start with a
* consistent value for "post_seq" */
post_seq = complete_seq = gen_prn(seed);
num_completed = 0;
num_posted = 0;
/* construct all of the communicators, just dups of comm world for now */
for (i = 0; i < NUM_COMMS; ++i) {
MPI_Comm_dup(MPI_COMM_WORLD, &comms[i]);
}
/* fill the entire window of ops */
for (i = 0; i < WINDOW; ++i) {
reqs[i] = MPI_REQUEST_NULL;
memset(&larr[i], 0, sizeof(struct laundry));
larr[i].case_num = -1;
/* randomly select a comm, using a new seed to avoid correlating
* particular kinds of NBC ops with particular communicators */
comm = comms[rand_range(gen_prn(post_seq), 0, NUM_COMMS)];
start_random_nonblocking(comm, post_seq, &reqs[i], &larr[i]);
++num_posted;
post_seq = gen_prn(post_seq);
}
/* now loop repeatedly, completing ops with "random" completion functions,
* until we've posted and completed MAIN_ITERATIONS ops */
while (num_completed < MAIN_ITERATIONS) {
complete_something_somehow(complete_seq, WINDOW, reqs, &outcount, indices);
complete_seq = gen_prn(complete_seq);
for (i = 0; i < outcount; ++i) {
int idx = indices[i];
assert(reqs[idx] == MPI_REQUEST_NULL);
if (larr[idx].case_num != -1) {
check_after_completion(&larr[idx]);
cleanup_laundry(&larr[idx]);
++num_completed;
if (num_posted < MAIN_ITERATIONS) {
comm = comms[rand_range(gen_prn(post_seq), 0, NUM_COMMS)];
start_random_nonblocking(comm, post_seq, &reqs[idx], &larr[idx]);
++num_posted;
post_seq = gen_prn(post_seq);
}
}
}
/* "randomly" and infrequently introduce some jitter into the system */
if (0 == rand_range(gen_prn(complete_seq + wrank), 0, CHANCE_OF_SLEEP)) {
usleep(JITTER_DELAY); /* take a short nap */
}
}
for (i = 0; i < NUM_COMMS; ++i) {
MPI_Comm_free(&comms[i]);
}
if (wrank == 0) {
if (errs)
printf("found %d errors\n", errs);
else
printf(" No errors\n");
}
MPI_Finalize();
return 0;
}