static ucs_status_t ucp_perf_setup(ucx_perf_context_t *perf, ucx_perf_params_t *params)
{
ucp_params_t ucp_params;
ucp_config_t *config;
ucs_status_t status;
uint64_t features;
status = ucp_perf_test_check_params(params, &features);
if (status != UCS_OK) {
goto err;
}
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
goto err;
}
ucp_params.features = features;
ucp_params.request_size = 0;
ucp_params.request_init = NULL;
ucp_params.request_cleanup = NULL;
status = ucp_init(&ucp_params, config, &perf->ucp.context);
ucp_config_release(config);
if (status != UCS_OK) {
goto err;
}
status = ucp_worker_create(perf->ucp.context, params->thread_mode,
&perf->ucp.worker);
if (status != UCS_OK) {
goto err_cleanup;
}
status = ucp_perf_test_alloc_mem(perf, params);
if (status != UCS_OK) {
ucs_warn("ucp test failed to alocate memory");
goto err_destroy_worker;
}
status = ucp_perf_test_setup_endpoints(perf, features);
if (status != UCS_OK) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Failed to setup endpoints: %s", ucs_status_string(status));
}
goto err_free_mem;
}
return UCS_OK;
err_free_mem:
ucp_perf_test_free_mem(perf);
err_destroy_worker:
ucp_worker_destroy(perf->ucp.worker);
err_cleanup:
ucp_cleanup(perf->ucp.context);
err:
return status;
}
static int mca_spml_ucx_component_close(void)
{
if (mca_spml_ucx.ucp_context) {
ucp_cleanup(mca_spml_ucx.ucp_context);
mca_spml_ucx.ucp_context = NULL;
}
return OSHMEM_SUCCESS;
}
static void ucp_perf_cleanup(ucx_perf_context_t *perf)
{
ucp_perf_test_cleanup_endpoints(perf);
rte_call(perf, barrier);
ucp_perf_test_free_mem(perf);
ucp_worker_destroy(perf->ucp.worker);
ucp_cleanup(perf->ucp.context);
}
int mca_pml_ucx_close(void)
{
PML_UCX_VERBOSE(1, "mca_pml_ucx_close");
if (ompi_pml_ucx.ucp_context != NULL) {
ucp_cleanup(ompi_pml_ucx.ucp_context);
ompi_pml_ucx.ucp_context = NULL;
}
return OMPI_SUCCESS;
}
static int component_finalize(void) {
int i;
for (i = 0; i < ompi_proc_world_size(); i++) {
ucp_ep_h ep = OSC_UCX_GET_EP(&(ompi_mpi_comm_world.comm), i);
if (ep != NULL) {
ucp_ep_destroy(ep);
}
}
if (mca_osc_ucx_component.ucp_worker != NULL) {
ucp_worker_destroy(mca_osc_ucx_component.ucp_worker);
}
assert(mca_osc_ucx_component.num_incomplete_req_ops == 0);
OBJ_DESTRUCT(&mca_osc_ucx_component.requests);
opal_progress_unregister(progress_callback);
ucp_cleanup(mca_osc_ucx_component.ucp_context);
return OMPI_SUCCESS;
}
static ucs_status_t ucp_perf_setup(ucx_perf_context_t *perf,
ucx_perf_params_t *params)
{
ucp_params_t ucp_params;
ucp_worker_params_t worker_params;
ucp_config_t *config;
ucs_status_t status;
ucp_params.field_mask = UCP_PARAM_FIELD_FEATURES;
ucp_params.features = 0;
status = ucp_perf_test_fill_params(params, &ucp_params);
if (status != UCS_OK) {
goto err;
}
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
goto err;
}
status = ucp_init(&ucp_params, config, &perf->ucp.context);
ucp_config_release(config);
if (status != UCS_OK) {
goto err;
}
worker_params.field_mask = UCP_WORKER_PARAM_FIELD_THREAD_MODE;
worker_params.thread_mode = params->thread_mode;
status = ucp_worker_create(perf->ucp.context, &worker_params,
&perf->ucp.worker);
if (status != UCS_OK) {
goto err_cleanup;
}
status = ucp_perf_test_alloc_mem(perf, params);
if (status != UCS_OK) {
ucs_warn("ucp test failed to alocate memory");
goto err_destroy_worker;
}
status = ucp_perf_test_setup_endpoints(perf, ucp_params.features);
if (status != UCS_OK) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Failed to setup endpoints: %s", ucs_status_string(status));
}
goto err_free_mem;
}
return UCS_OK;
err_free_mem:
ucp_perf_test_free_mem(perf);
err_destroy_worker:
ucp_worker_destroy(perf->ucp.worker);
err_cleanup:
ucp_cleanup(perf->ucp.context);
err:
return status;
}
void print_ucp_info(int print_opts, ucs_config_print_flags_t print_flags,
uint64_t features)
{
ucp_config_t *config;
ucs_status_t status;
ucp_context_h context;
ucp_worker_h worker;
ucp_params_t params;
ucp_worker_params_t worker_params;
ucp_ep_params_t ep_params;
ucp_address_t *address;
size_t address_length;
ucp_ep_h ep;
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
return;
}
memset(¶ms, 0, sizeof(params));
params.features = features;
status = ucp_init(¶ms, config, &context);
if (status != UCS_OK) {
printf("<Failed to create UCP context>\n");
goto out_release_config;
}
if (print_opts & PRINT_UCP_CONTEXT) {
ucp_context_print_info(context, stdout);
}
if (!(print_opts & (PRINT_UCP_WORKER|PRINT_UCP_EP))) {
goto out_cleanup_context;
}
worker_params.field_mask = UCP_WORKER_PARAM_FIELD_THREAD_MODE;
worker_params.thread_mode = UCS_THREAD_MODE_MULTI;
status = ucp_worker_create(context, &worker_params, &worker);
if (status != UCS_OK) {
printf("<Failed to create UCP worker>\n");
goto out_cleanup_context;
}
if (print_opts & PRINT_UCP_WORKER) {
ucp_worker_print_info(worker, stdout);
}
if (print_opts & PRINT_UCP_EP) {
status = ucp_worker_get_address(worker, &address, &address_length);
if (status != UCS_OK) {
printf("<Failed to get UCP worker address>\n");
goto out_destroy_worker;
}
ep_params.field_mask = UCP_EP_PARAM_FIELD_REMOTE_ADDRESS;
ep_params.address = address;
status = ucp_ep_create(worker, &ep_params, &ep);
ucp_worker_release_address(worker, address);
if (status != UCS_OK) {
printf("<Failed to get create UCP endpoint>\n");
goto out_destroy_worker;
}
ucp_ep_print_info(ep, stdout);
ucp_ep_destroy(ep);
}
out_destroy_worker:
ucp_worker_destroy(worker);
out_cleanup_context:
ucp_cleanup(context);
out_release_config:
ucp_config_release(config);
}
int main(int argc, char **argv)
{
/* UCP temporary vars */
ucp_params_t ucp_params;
ucp_config_t *config;
ucs_status_t status;
/* UCP handler objects */
ucp_context_h ucp_context;
ucp_worker_h ucp_worker;
/* OOB connection vars */
uint64_t addr_len = 0;
char *server = NULL;
int oob_sock = -1;
int ret = -1;
/* Parse the command line */
if (parse_cmd(argc, argv, &server) != UCS_OK) {
goto err;
}
/* UCP initialization */
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
goto err;
}
ucp_params.features = UCP_FEATURE_TAG;
if (ucp_test_mode == TEST_MODE_WAIT || ucp_test_mode == TEST_MODE_EVENTFD) {
ucp_params.features |= UCP_FEATURE_WAKEUP;
}
ucp_params.request_size = sizeof(struct ucx_context);
ucp_params.request_init = request_init;
ucp_params.request_cleanup = NULL;
status = ucp_init(&ucp_params, config, &ucp_context);
ucp_config_print(config, stdout, NULL, UCS_CONFIG_PRINT_CONFIG);
ucp_config_release(config);
if (status != UCS_OK) {
goto err;
}
status = ucp_worker_create(ucp_context, UCS_THREAD_MODE_SINGLE, &ucp_worker);
if (status != UCS_OK) {
goto err_cleanup;
}
status = ucp_worker_get_address(ucp_worker, &local_addr, &local_addr_len);
if (status != UCS_OK) {
goto err_worker;
}
printf("[0x%x] local address length: %zu\n",
(unsigned int)pthread_self(), local_addr_len);
/* OOB connection establishment */
if (server) {
peer_addr_len = local_addr_len;
oob_sock = run_client(server);
if (oob_sock < 0) {
goto err_addr;
}
ret = recv(oob_sock, &addr_len, sizeof(addr_len), 0);
if (ret < 0) {
fprintf(stderr, "failed to receive address length\n");
goto err_addr;
}
peer_addr_len = addr_len;
peer_addr = malloc(peer_addr_len);
if (!peer_addr) {
fprintf(stderr, "unable to allocate memory\n");
goto err_addr;
}
ret = recv(oob_sock, peer_addr, peer_addr_len, 0);
if (ret < 0) {
fprintf(stderr, "failed to receive address\n");
goto err_peer_addr;
}
} else {
oob_sock = run_server();
if (oob_sock < 0) {
goto err_peer_addr;
}
addr_len = local_addr_len;
ret = send(oob_sock, &addr_len, sizeof(addr_len), 0);
if (ret < 0 || ret != sizeof(addr_len)) {
fprintf(stderr, "failed to send address length\n");
goto err_peer_addr;
}
ret = send(oob_sock, local_addr, local_addr_len, 0);
if (ret < 0 || ret != local_addr_len) {
fprintf(stderr, "failed to send address\n");
goto err_peer_addr;
}
}
close(oob_sock);
ret = run_test(server, ucp_worker);
err_peer_addr:
free(peer_addr);
err_addr:
ucp_worker_release_address(ucp_worker, local_addr);
err_worker:
ucp_worker_destroy(ucp_worker);
err_cleanup:
ucp_cleanup(ucp_context);
err:
return ret;
}
void print_ucp_info(int print_opts, ucs_config_print_flags_t print_flags,
uint64_t ctx_features, const ucp_ep_params_t *base_ep_params,
size_t estimated_num_eps, unsigned dev_type_bitmap,
const char *mem_size)
{
ucp_config_t *config;
ucs_status_t status;
ucs_status_ptr_t status_ptr;
ucp_context_h context;
ucp_worker_h worker;
ucp_params_t params;
ucp_worker_params_t worker_params;
ucp_ep_params_t ep_params;
ucp_address_t *address;
size_t address_length;
resource_usage_t usage;
ucp_ep_h ep;
status = ucp_config_read(NULL, NULL, &config);
if (status != UCS_OK) {
return;
}
memset(¶ms, 0, sizeof(params));
params.field_mask = UCP_PARAM_FIELD_FEATURES |
UCP_PARAM_FIELD_ESTIMATED_NUM_EPS;
params.features = ctx_features;
params.estimated_num_eps = estimated_num_eps;
get_resource_usage(&usage);
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_SELF))) {
ucp_config_modify(config, "SELF_DEVICES", "");
}
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_SHM))) {
ucp_config_modify(config, "SHM_DEVICES", "");
}
if (!(dev_type_bitmap & UCS_BIT(UCT_DEVICE_TYPE_NET))) {
ucp_config_modify(config, "NET_DEVICES", "");
}
status = ucp_init(¶ms, config, &context);
if (status != UCS_OK) {
printf("<Failed to create UCP context>\n");
goto out_release_config;
}
if ((print_opts & PRINT_MEM_MAP) && (mem_size != NULL)) {
ucp_mem_print_info(mem_size, context, stdout);
}
if (print_opts & PRINT_UCP_CONTEXT) {
ucp_context_print_info(context, stdout);
print_resource_usage(&usage, "UCP context");
}
if (!(print_opts & (PRINT_UCP_WORKER|PRINT_UCP_EP))) {
goto out_cleanup_context;
}
worker_params.field_mask = UCP_WORKER_PARAM_FIELD_THREAD_MODE;
worker_params.thread_mode = UCS_THREAD_MODE_MULTI;
get_resource_usage(&usage);
status = ucp_worker_create(context, &worker_params, &worker);
if (status != UCS_OK) {
printf("<Failed to create UCP worker>\n");
goto out_cleanup_context;
}
if (print_opts & PRINT_UCP_WORKER) {
ucp_worker_print_info(worker, stdout);
print_resource_usage(&usage, "UCP worker");
}
if (print_opts & PRINT_UCP_EP) {
status = ucp_worker_get_address(worker, &address, &address_length);
if (status != UCS_OK) {
printf("<Failed to get UCP worker address>\n");
goto out_destroy_worker;
}
ep_params = *base_ep_params;
ep_params.field_mask |= UCP_EP_PARAM_FIELD_REMOTE_ADDRESS;
ep_params.address = address;
status = ucp_ep_create(worker, &ep_params, &ep);
ucp_worker_release_address(worker, address);
if (status != UCS_OK) {
printf("<Failed to create UCP endpoint>\n");
goto out_destroy_worker;
}
ucp_ep_print_info(ep, stdout);
status_ptr = ucp_disconnect_nb(ep);
if (UCS_PTR_IS_PTR(status_ptr)) {
do {
ucp_worker_progress(worker);
status = ucp_request_test(status_ptr, NULL);
} while (status == UCS_INPROGRESS);
ucp_request_release(status_ptr);
}
}
out_destroy_worker:
ucp_worker_destroy(worker);
out_cleanup_context:
ucp_cleanup(context);
out_release_config:
ucp_config_release(config);
}
static int component_init(bool enable_progress_threads, bool enable_mpi_threads) {
ucp_config_t *config = NULL;
ucp_params_t context_params;
bool progress_registered = false, requests_created = false;
int ret = OMPI_SUCCESS;
ucs_status_t status;
mca_osc_ucx_component.ucp_context = NULL;
mca_osc_ucx_component.ucp_worker = NULL;
mca_osc_ucx_component.enable_mpi_threads = enable_mpi_threads;
status = ucp_config_read("MPI", NULL, &config);
if (UCS_OK != status) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_config_read failed: %d\n",
__FILE__, __LINE__, status);
return OMPI_ERROR;
}
OBJ_CONSTRUCT(&mca_osc_ucx_component.requests, opal_free_list_t);
requests_created = true;
ret = opal_free_list_init (&mca_osc_ucx_component.requests,
sizeof(ompi_osc_ucx_request_t),
opal_cache_line_size,
OBJ_CLASS(ompi_osc_ucx_request_t),
0, 0, 8, 0, 8, NULL, 0, NULL, NULL, NULL);
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: opal_free_list_init failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
mca_osc_ucx_component.num_incomplete_req_ops = 0;
ret = opal_progress_register(progress_callback);
progress_registered = true;
if (OMPI_SUCCESS != ret) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: opal_progress_register failed: %d\n",
__FILE__, __LINE__, ret);
goto error;
}
/* initialize UCP context */
memset(&context_params, 0, sizeof(ucp_context_h));
context_params.field_mask = UCP_PARAM_FIELD_FEATURES |
UCP_PARAM_FIELD_MT_WORKERS_SHARED |
UCP_PARAM_FIELD_ESTIMATED_NUM_EPS |
UCP_PARAM_FIELD_REQUEST_INIT |
UCP_PARAM_FIELD_REQUEST_SIZE;
context_params.features = UCP_FEATURE_RMA | UCP_FEATURE_AMO32 | UCP_FEATURE_AMO64;
context_params.mt_workers_shared = 0;
context_params.estimated_num_eps = ompi_proc_world_size();
context_params.request_init = internal_req_init;
context_params.request_size = sizeof(ompi_osc_ucx_internal_request_t);
status = ucp_init(&context_params, config, &mca_osc_ucx_component.ucp_context);
ucp_config_release(config);
if (UCS_OK != status) {
opal_output_verbose(1, ompi_osc_base_framework.framework_output,
"%s:%d: ucp_init failed: %d\n",
__FILE__, __LINE__, status);
ret = OMPI_ERROR;
goto error;
}
return ret;
error:
if (progress_registered) opal_progress_unregister(progress_callback);
if (requests_created) OBJ_DESTRUCT(&mca_osc_ucx_component.requests);
if (mca_osc_ucx_component.ucp_context) ucp_cleanup(mca_osc_ucx_component.ucp_context);
return ret;
}
