void publish(
rclcpp::Node::SharedPtr node,
const std::string & message_type,
std::vector<typename T::SharedPtr> messages,
size_t number_of_cycles = 100)
{
auto start = std::chrono::steady_clock::now();
rmw_qos_profile_t custom_qos_profile = rmw_qos_profile_default;
custom_qos_profile.depth = messages.size();
auto publisher = node->create_publisher<T>(
std::string("test/message/") + message_type, custom_qos_profile);
rclcpp::WallRate cycle_rate(10);
rclcpp::WallRate message_rate(100);
size_t cycle_index = 0;
// publish all messages up to number_of_cycles times, longer sleep between each cycle
while (rclcpp::ok() && cycle_index < number_of_cycles) {
size_t message_index = 0;
// publish all messages one by one, shorter sleep between each message
while (rclcpp::ok() && message_index < messages.size()) {
printf("publishing message #%zu\n", message_index + 1);
publisher->publish(messages[message_index]);
++message_index;
message_rate.sleep();
}
++cycle_index;
cycle_rate.sleep();
}
auto end = std::chrono::steady_clock::now();
std::chrono::duration<float> diff = (end - start);
printf("published for %f seconds\n", diff.count());
}
void
benchmark (struct pe_vars v, long * msg_buffer)
{
static double pwrk[_SHMEM_REDUCE_SYNC_SIZE];
static long psync[_SHMEM_BCAST_SYNC_SIZE];
static double mr, mr_sum;
unsigned long size, i;
memset(psync, _SHMEM_SYNC_VALUE, sizeof(long[_SHMEM_BCAST_SYNC_SIZE]));
/*
* Warmup
*/
if (v.me < v.pairs) {
//for (i = 0; i < (ITERS_LARGE * MAX_MSG_SZ); i += MAX_MSG_SZ) {
for (i = 0; i < ITERS_LARGE; i += 1) {
//shmem_putmem(&msg_buffer[i], &msg_buffer[i], MAX_MSG_SZ, v.nxtpe);
shmem_long_put(&msg_buffer[i], &msg_buffer[i], MAX_MSG_SZ, v.nxtpe);
}
}
shmem_barrier_all();
/*
* Benchmark
*/
for (size = 1; size <= MAX_MSG_SZ; size <<= 1) {
i = size < LARGE_THRESHOLD ? ITERS_SMALL : ITERS_LARGE;
mr = message_rate(v, msg_buffer, size, i);
shmem_double_sum_to_all(&mr_sum, &mr, 1, 0, 0, v.npes, pwrk, psync);
print_message_rate(v.me, size, mr_sum);
}
}
void benchmark (long * msg_buffer, int me, int pairs, int nxtpe, MPI_Win win)
{
static double mr, mr_sum;
int iters;
if (msg_buffer == NULL) {
printf("Input buffer is NULL, no reason to proceed\n");
exit(-1);
}
/*
* Warmup
*/
if (me < pairs) {
for (int i = 0; i < ITERS_LARGE; i += 1) {
MPI_Put ((msg_buffer + i*MAX_MSG_SZ), MAX_MSG_SZ, MPI_LONG, nxtpe, i*MAX_MSG_SZ, MAX_MSG_SZ, MPI_LONG, win);
MPI_Win_flush_local (nxtpe, win);
}
}
MPI_Win_flush_all(win);
MPI_Barrier(MPI_COMM_WORLD);
/*
* Benchmark
*/
for (long size = 1; size <= MAX_MSG_SZ; size <<= 1) {
iters = size < LARGE_THRESHOLD ? ITERS_SMALL : ITERS_LARGE;
mr = message_rate(msg_buffer, size, iters, me, pairs, nxtpe, win);
MPI_Reduce(&mr, &mr_sum, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
print_message_rate(size, mr_sum, me);
}
}
